update Walk.py and update models for Walk gym

.gitignore

@@ -20,3 +20,4 @@ best_model.zip
 *.yaml
 *.iml
 *.TXT
+events.out.tfevents.*

command.md

@@ -1,14 +0,0 @@
-训练（默认）
-bash train.sh
-
-测试（实时+显示画面)
-GYM_CPU_MODE=test GYM_CPU_TEST_MODEL=scripts/gyms/logs/Walk_R0_005/best_model.zip GYM_CPU_TEST_FOLDER=scripts/gyms/logs/Walk_R0_005/ GYM_CPU_TEST_NO_RENDER=0 GYM_CPU_TEST_NO_REALTIME=0 bash train.sh
-
-测试（无画面、非实时）
-GYM_CPU_MODE=test GYM_CPU_TEST_NO_RENDER=1 GYM_CPU_TEST_NO_REALTIME=1 bash train.sh
-
-retrain（继续训练）
-GYM_CPU_MODE=train GYM_CPU_TRAIN_MODEL=scripts/gyms/logs/Walk_R0_005/best_model.zip bash train.sh
-
-retrain+改训练超参
-GYM_CPU_MODE=train GYM_CPU_TRAIN_MODEL=scripts/gyms/logs/Walk_R0_004/best_model.zip GYM_CPU_TRAIN_LR=2e-4 GYM_CPU_TRAIN_CLIP_RANGE=0.13 GYM_CPU_TRAIN_BATCH_SIZE=256 YM_CPU_TRAIN_GAMMA=0.95 GYM_CPU_TRAIN_ENT_COEF=0.05 GYM_CPU_TRAIN_EPOCHS=8 bash train.sh

communication/server.py

@@ -1,4 +1,5 @@
 import logging
+import os
 import socket
 import time
 from select import select
@@ -15,6 +16,11 @@ class Server:
         self.__socket: socket.socket = self._create_socket()
         self.__send_buff = []
         self.__rcv_buffer_size = 1024
+        self.__rcv_buffer_default_size = 1024
+        self.__max_msg_size = 1048576
+        self.__shrink_threshold = 8192
+        self.__shrink_after_msgs = 200
+        self.__small_msg_streak = 0
         self.__rcv_buffer = bytearray(self.__rcv_buffer_size)
 
     def _create_socket(self) -> socket.socket:
@@ -105,6 +111,10 @@ class Server:
 
             msg_size = int.from_bytes(self.__rcv_buffer[:4], byteorder="big", signed=False)
 
+            # Guard against corrupted frame lengths that would trigger huge allocations.
+            if msg_size <= 0 or msg_size > self.__max_msg_size:
+                raise ConnectionResetError
+
             if msg_size > self.__rcv_buffer_size:
                 self.__rcv_buffer_size = msg_size
                 self.__rcv_buffer = bytearray(self.__rcv_buffer_size)
@@ -120,6 +130,15 @@ class Server:
                 message=self.__rcv_buffer[:msg_size].decode()
             )
 
+            if msg_size <= self.__shrink_threshold and self.__rcv_buffer_size > self.__rcv_buffer_default_size:
+                self.__small_msg_streak += 1
+                if self.__small_msg_streak >= self.__shrink_after_msgs:
+                    self.__rcv_buffer_size = self.__rcv_buffer_default_size
+                    self.__rcv_buffer = bytearray(self.__rcv_buffer_size)
+                    self.__small_msg_streak = 0
+            else:
+                self.__small_msg_streak = 0
+
             # 如果socket没有更多数据就退出
             if len(select([self.__socket], [], [], 0.0)[0]) == 0:
                 break

scripts/commons/Server.py

@@ -1,9 +1,14 @@
 import subprocess
 import os
 import time
+import threading
 
 
 class Server():
+    WATCHDOG_ENABLED = True
+    WATCHDOG_INTERVAL_SEC = 30.0
+    WATCHDOG_RSS_MB_LIMIT = 800
+
     def __init__(self, first_server_p, first_monitor_p, n_servers, no_render=True, no_realtime=True) -> None:
         try:
             import psutil
@@ -14,6 +19,10 @@ class Server():
         self.first_server_p = first_server_p
         self.n_servers = n_servers
         self.rcss_processes = []
+        self._server_specs = []
+        self._watchdog_stop = threading.Event()
+        self._watchdog_lock = threading.Lock()
+        self._watchdog_thread = None
         first_monitor_p = first_monitor_p + 100
 
         # makes it easier to kill test servers without affecting train servers
@@ -23,7 +32,15 @@ class Server():
         for i in range(n_servers):
             port = first_server_p + i
             mport = first_monitor_p + i
+            self._server_specs.append((port, mport, cmd, render_arg, realtime_arg))
+            proc = self._spawn_server(port, mport, cmd, render_arg, realtime_arg)
+            self.rcss_processes.append(proc)
 
+        if self.WATCHDOG_ENABLED:
+            self._watchdog_thread = threading.Thread(target=self._watchdog_loop, daemon=True)
+            self._watchdog_thread.start()
+
+    def _spawn_server(self, port, mport, cmd, render_arg, realtime_arg):
         server_cmd = f"{cmd} -c {port} -m {mport} {render_arg} {realtime_arg}".strip()
 
         proc = subprocess.Popen(
@@ -42,19 +59,79 @@ class Server():
                 f"rcssservermj exited early (code={rc}) on server port {port}, monitor port {mport}"
             )
 
-            self.rcss_processes.append(proc)
+        return proc
+
+    @staticmethod
+    def _pid_rss_mb(pid):
+        try:
+            with open(f"/proc/{pid}/status", "r", encoding="utf-8") as f:
+                for line in f:
+                    if line.startswith("VmRSS:"):
+                        parts = line.split()
+                        if len(parts) >= 2:
+                            # VmRSS is kB
+                            return float(parts[1]) / 1024.0
+        except (FileNotFoundError, ProcessLookupError, PermissionError, OSError):
+            return 0.0
+        return 0.0
+
+    def _restart_server_at_index(self, idx, reason):
+        port, mport, cmd, render_arg, realtime_arg = self._server_specs[idx]
+        old_proc = self.rcss_processes[idx]
+        try:
+            old_proc.terminate()
+            old_proc.wait(timeout=1.0)
+        except Exception:
+            try:
+                old_proc.kill()
+            except Exception:
+                pass
+
+        new_proc = self._spawn_server(port, mport, cmd, render_arg, realtime_arg)
+        self.rcss_processes[idx] = new_proc
+        print(
+            f"[ServerWatchdog] Restarted server idx={idx} port={port} monitor={mport} reason={reason}"
+        )
+
+    def _watchdog_loop(self):
+        while not self._watchdog_stop.wait(self.WATCHDOG_INTERVAL_SEC):
+            with self._watchdog_lock:
+                for i, proc in enumerate(self.rcss_processes):
+                    rc = proc.poll()
+                    if rc is not None:
+                        self._restart_server_at_index(i, f"exited:{rc}")
+                        continue
+
+                    rss_mb = self._pid_rss_mb(proc.pid)
+                    if rss_mb > self.WATCHDOG_RSS_MB_LIMIT:
+                        self._restart_server_at_index(i, f"rss_mb:{rss_mb:.1f}")
 
     def check_running_servers(self, psutil, first_server_p, first_monitor_p, n_servers):
         ''' Check if any server is running on chosen ports '''
         found = False
-        p_list = [p for p in psutil.process_iter() if p.cmdline() and "rcssservermj" in " ".join(p.cmdline())]
         range1 = (first_server_p, first_server_p + n_servers)
         range2 = (first_monitor_p, first_monitor_p + n_servers)
         bad_processes = []
 
+        def safe_cmdline(proc):
+            try:
+                return proc.cmdline()
+            except (psutil.ZombieProcess, psutil.NoSuchProcess, psutil.AccessDenied, OSError):
+                return []
+
+        p_list = []
+        for p in psutil.process_iter():
+            cmdline = safe_cmdline(p)
+            if cmdline and "rcssservermj" in " ".join(cmdline):
+                p_list.append(p)
+
         for p in p_list:
             # currently ignoring remaining default port when only one of the ports is specified (uncommon scenario)
-            ports = [int(arg) for arg in p.cmdline()[1:] if arg.isdigit()]
+            cmdline = safe_cmdline(p)
+            if not cmdline:
+                continue
+
+            ports = [int(arg) for arg in cmdline[1:] if arg.isdigit()]
             if len(ports) == 0:
                 ports = [60000, 60100]  # default server ports (changing this is unlikely)
 
@@ -66,7 +143,7 @@ class Server():
                     print("\nThere are already servers running on the same port(s)!")
                     found = True
                 bad_processes.append(p)
-                print(f"Port(s) {','.join(conflicts)} already in use by \"{' '.join(p.cmdline())}\" (PID:{p.pid})")
+                print(f"Port(s) {','.join(conflicts)} already in use by \"{' '.join(cmdline)}\" (PID:{p.pid})")
 
         if found:
             print()
@@ -78,6 +155,9 @@ class Server():
                     return
 
     def kill(self):
+        self._watchdog_stop.set()
+        if self._watchdog_thread is not None:
+            self._watchdog_thread.join(timeout=1.0)
         for p in self.rcss_processes:
             p.kill()
         print(f"Killed {self.n_servers} rcssservermj processes starting at {self.first_server_p}")

scripts/commons/Train_Base.py

@@ -6,7 +6,7 @@ from scripts.commons.UI import UI
 from shutil import copy
 from stable_baselines3 import PPO
 from stable_baselines3.common.base_class import BaseAlgorithm
-from stable_baselines3.common.callbacks import EvalCallback, CheckpointCallback, CallbackList, BaseCallback
+from stable_baselines3.common.callbacks import EvalCallback, CheckpointCallback, CallbackList, BaseCallback, StopTrainingOnNoModelImprovement
 from typing import Callable
 # from world.world import World
 from xml.dom import minidom
@@ -266,11 +266,28 @@ class Train_Base():
 
         evaluate = bool(eval_env is not None and eval_freq is not None)
 
+        # Optional early stop: stop training when eval reward does not improve for N eval rounds.
+        no_improve_evals = int(os.environ.get("GYM_CPU_EARLY_STOP_NO_IMPROVE_EVALS", "0"))
+        min_evals_before_stop = int(os.environ.get("GYM_CPU_EARLY_STOP_MIN_EVALS", "6"))
+        stop_on_no_improve = None
+        if evaluate and no_improve_evals > 0:
+            stop_on_no_improve = StopTrainingOnNoModelImprovement(
+                max_no_improvement_evals=no_improve_evals,
+                min_evals=min_evals_before_stop,
+                verbose=1,
+            )
+
         # Create evaluation callback
-        eval_callback = None if not evaluate else EvalCallback(eval_env, n_eval_episodes=eval_eps, eval_freq=eval_freq,
+        eval_callback = None if not evaluate else EvalCallback(
+            eval_env,
+            n_eval_episodes=eval_eps,
+            eval_freq=eval_freq,
             log_path=path,
-                                                               best_model_save_path=path, deterministic=True,
-                                                               render=False)
+            best_model_save_path=path,
+            deterministic=True,
+            render=False,
+            callback_after_eval=stop_on_no_improve,
+        )
 
         # Create custom callback to display evaluations
         custom_callback = None if not evaluate else Cyclic_Callback(eval_freq,

scripts/gyms/Walk.py

@@ -7,7 +7,7 @@ from random import random
 from random import uniform
 from itertools import count
 
-from stable_baselines3 import PPO
+from stable_baselines3 import PPO, TD3, DDPG, SAC, A2C
 from stable_baselines3.common.monitor import Monitor
 from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
 
@@ -53,6 +53,10 @@ class WalkEnv(gym.Env):
         self.route_completed = False
         self.debug_every_n_steps = 5
         self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = 600.0
+        self.reward_debug_burst_steps = 10
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
         self.calibrate_nominal_from_neutral = True
         self.auto_calibrate_train_sim_flip = True
         self.nominal_calibrated_once = False
@@ -60,6 +64,11 @@ class WalkEnv(gym.Env):
         self._target_hz = 0.0
         self._target_dt = 0.0
         self._last_sync_time = None
+        self._speed_estimate = 0.0
+        self._speed_from_acc = 0.0
+        self._prev_accelerometer = np.zeros(3, dtype=np.float32)
+        self._speed_smoothing = 0.85
+        self._fallback_dt = 0.02
         target_hz_env = 0
         if target_hz_env:
             try:
@@ -92,32 +101,32 @@ class WalkEnv(gym.Env):
         # 中立姿态
         self.joint_nominal_position = np.array(
             [
-                0.0,
-                0.0,
-                0.0,
-                1.4,
-                0.0,
-                -0.4,
-                0.0,
-                -1.4,
-                0.0,
-                0.4,
-                0.0,
-                -0.4,
-                0.0,
-                0.0,
-                0.8,
-                -0.4,
-                0.0,
-                0.4,
-                0.0,
-                0.0,
-                -0.8,
-                0.4,
-                0.0,
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
             ]
         )
-        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        # self.joint_nominal_position = np.zeros(self.no_of_actions)
         self.train_sim_flip = np.array(
             [
                 1.0,  # 0: Head_yaw (he1)
@@ -146,22 +155,108 @@ class WalkEnv(gym.Env):
             ]
         )
 
-        self.scaling_factor = 0.3
+        self.scaling_factor = 0.5
         # self.scaling_factor = 1
 
         # Encourage a minimum lateral stance so the policy avoids feet overlap.
         self.min_stance_rad = 0.10
 
         # Small reset perturbations for robustness training.
-        self.enable_reset_perturb = True
-        self.reset_beam_yaw_range_deg = 180  # randomize target direction fully to encourage learning a real walk instead of a fixed gait
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = 180.0
+        self.reset_target_bearing_range_deg = 0.0
+        self.reset_target_distance_min = 5
+        self.reset_target_distance_max = 10
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
         self.reset_joint_noise_rad = 0.025
         self.reset_perturb_steps = 4
         self.reset_recover_steps = 8
 
+        self.reward_smoothness_scale = 0.03
+        self.reward_smoothness_cap = 0.45
+        self.reward_forward_stability_gate = 0.35
+        self.reward_forward_tilt_hard_threshold = 0.50
+        self.reward_forward_tilt_hard_scale = 0.20
+        self.reward_head_toward_bonus = 1.0
+        self.turn_stationary_radius = 0.2
+        self.turn_stationary_penalty_scale = 3.0
+        self.stationary_start_steps = 20
+        self.stationary_step_eps = 0.015
+        self.stationary_penalty_scale = 1.2
+        self.train_stage = "walk"
+        self.in_place_radius = 0.18
+        self.in_place_center_reward_scale = 0.60
+        self.in_place_drift_penalty_scale = 1.20
+        self.waypoint_reach_distance = 0.3
+        self.num_waypoints = 1
+        self.exploration_start_steps = 40
+        self.exploration_scale = 0.012
+        self.exploration_cap = 0.2
+        self.exploration_target_novelty = 1.0
+        self.exploration_sigma = 0.7
+        self.reward_stride_swing_scale = 0.20
+        self.reward_stride_phase_scale = 0.18
+        self.reward_knee_drive_scale = 0.10
+        self.reward_knee_lift_scale = 0.12
+        self.reward_knee_lift_target = 0.15
+        self.reward_knee_lift_shortfall_scale = 0.05
+        self.reward_knee_overbend_threshold = 0.60
+        self.reward_knee_overbend_scale = 0.35
+        self.reward_hip_lift_scale = 0.12
+        self.reward_hip_lift_target = 0.80
+        self.reward_knee_alternate_scale = 0.10
+        self.reward_knee_bilateral_scale = 0.16
+        self.reward_single_leg_penalty_scale = 0.22
+        self.reward_knee_phase_switch_scale = 0.14
+        self.knee_phase_deadband = 0.10
+        self.knee_phase_min_interval = 18
+        self.knee_phase_target_interval = 22
+        self.knee_phase_fast_switch_penalty_scale = 0.10
+        self.knee_phase_max_hold_frames = 28
+        self.knee_phase_hold_penalty_scale = 0.18
+        self.reward_stride_cap = 0.80
+        self.reward_knee_explore_scale = 0.03
+        self.reward_knee_explore_delta_scale = 0.03
+        self.reward_knee_explore_cap = 0.10
+        self.reward_hip_pitch_explore_scale = 0.07
+        self.reward_hip_pitch_explore_delta_scale = 0.07
+        self.reward_hip_pitch_explore_cap = 0.10
+        self.reward_progress_scale = 18
+        self.reward_survival_scale = 0.5
+        self.reward_idle_penalty_scale = 0.6
+        self.reward_accel_penalty_scale = 0.08
+        self.reward_accel_penalty_cap = 0.40
+        self.reward_accel_abs_limit = 13.5
+        self.reward_accel_abs_penalty_scale = 0.05
+        self.reward_accel_abs_penalty_cap = 0.40
+        self.reward_heading_align_scale = 0.28
+        self.reward_heading_error_scale = 0.05
+        self.reward_progress_tilt_gate_start = 0.28
+        self.reward_progress_knee_gate_min = 0.16
+        self.reward_progress_hip_gate_over = 0.18
+        self.reward_progress_gate_floor = 0.25
+        self.reward_knee_straight_threshold = 0.18
+        self.reward_knee_straight_penalty_scale = 0.45
+        self.reward_hip_overextend_threshold = 0.9
+        self.reward_hip_overextend_penalty_scale = 1
+        self.reward_leg_stretch_penalty_scale = 0.35
+        self.reward_stretch_lean_combo_scale = 0.55
+
         self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
         self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.action_history_len = 50
+        self.prev_action_history = np.zeros((self.action_history_len, self.no_of_actions), dtype=np.float32)
+        self.history_idx = 0
         self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.prev_knee_balance = 0.0
+        self.prev_knee_phase_sign = 0
+        self.knee_phase_frames_since_switch = 0
+        self.knee_phase_hold_frames = 0
         self.Player.server.connect()
         # sleep(2.0)  # Longer wait for connection to establish completely
         self.Player.server.send_immediate(
@@ -204,6 +299,10 @@ class WalkEnv(gym.Env):
         except OSError:
             pass
 
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
     def observe(self, init=False):
 
         """获取当前观测值"""
@@ -312,19 +411,30 @@ class WalkEnv(gym.Env):
         if seed is not None:
             np.random.seed(seed)
 
-        length1 = 2  # randomize target distance
-        length2 = np.random.uniform(0.6, 1)  # randomize target distance
-        length3 = np.random.uniform(0.6, 1)  # randomize target distance
-        angle2 = np.random.uniform(-30, 30)  # randomize initial orientation
-        angle3 = np.random.uniform(-30, 30)  # randomize target direction
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
 
         self.step_counter = 0
         self.waypoint_index = 0
         self.route_completed = False
         self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
         self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.prev_action_history.fill(0.0)
+        self.history_idx = 0
         self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.prev_knee_balance = 0.0
+        self.prev_knee_phase_sign = 0
+        self.knee_phase_frames_since_switch = 0
+        self.knee_phase_hold_frames = 0
         self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+        self._speed_estimate = 0.0
+        self._speed_from_acc = 0.0
+        self._prev_accelerometer = np.array(
+            getattr(self.Player.robot, "accelerometer", np.zeros(3)),
+            dtype=np.float32,
+        )
 
         # 随机 beam 目标位置和朝向，增加训练多样性
         beam_x = (random() - 0.5) * 10
@@ -379,14 +489,29 @@ class WalkEnv(gym.Env):
         self.initial_position = np.array(self.Player.world.global_position[:2])
         self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
         self.act = np.zeros(self.no_of_actions, np.float32)
-        # Build target in the robot's current forward direction instead of fixed global +x.
+        # Generate multiple waypoints along a path
         heading_deg = float(r.global_orientation_euler[2])
-        forward_offset = MathOps.rotate_2d_vec(np.array([length1, 0.0]), heading_deg, is_rad=False)
-        point1 = self.initial_position + forward_offset
-        point2 = point1 + MathOps.rotate_2d_vec(np.array([length2, 0]), angle2, is_rad=False)
-        point3 = point2 + MathOps.rotate_2d_vec(np.array([length3, 0]), angle3, is_rad=False)
-        self.point_list = [point1]
+        self.point_list = []
+        current_point = self.initial_position.copy()
+        
+        for i in range(self.num_waypoints):
+            # Each waypoint is placed further along the path
+            target_distance_wp = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+            self.target_distance_wp = target_distance_wp
+            target_bearing_deg_wp = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+            
+            target_offset = MathOps.rotate_2d_vec(
+                np.array([target_distance_wp, 0.0]),
+                heading_deg + target_bearing_deg_wp,
+                is_rad=False,
+            )
+            next_point = current_point + target_offset
+            self.point_list.append(next_point)
+            current_point = next_point.copy()
+        
         self.target_position = self.point_list[self.waypoint_index]
+        if self.train_stage == "in_place":
+            self.target_position = self.initial_position.copy()
         self.initial_height = self.Player.world.global_position[2]
 
         return self.observe(True), {}
@@ -394,119 +519,251 @@ class WalkEnv(gym.Env):
     def render(self, mode='human', close=False):
         return
 
+    
     def compute_reward(self, previous_pos, current_pos, action):
         height = float(self.Player.world.global_position[2])
         robot = self.Player.robot
 
-        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
-        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
-        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
-        ang_vel = np.deg2rad(robot.gyroscope)
-        ang_vel_mag = float(np.linalg.norm(ang_vel))
+        prev_dist_to_target = float(np.linalg.norm(self.target_position - previous_pos))
+        curr_dist_to_target = float(np.linalg.norm(self.target_position - current_pos))
+        dist_delta = prev_dist_to_target - curr_dist_to_target
 
-        is_fallen = height < 0.55
+        is_fallen = height < 0.45
         if is_fallen:
-            # remain = max(0, 800 - self.step_counter)
-            # return -8.0 - 0.01 * remain
-            return -1.0
+            return -2.0
         
-
-
-        # # 目标方向
-        # to_target = self.target_position - current_pos
-        # dist_to_target = float(np.linalg.norm(to_target))
-        # if dist_to_target < 0.5:
-        #     return 15.0
-
-        # forward_dir = to_target / dist_to_target if dist_to_target > 0.1 else np.array([1.0, 0.0])
-        # delta_pos = current_pos - previous_pos
-        # forward_step = float(np.dot(delta_pos, forward_dir))
-        # lateral_step = float(np.linalg.norm(delta_pos - forward_dir * forward_step))
-
-        # 奖励项
-        # progress_reward = 2 * forward_step
-        # lateral_penalty = -0.1 * lateral_step
-        alive_bonus = 2.0
-
-        # action_penalty = -0.01 * float(np.linalg.norm(action))
-        smoothness_penalty = -0.01 * float(np.linalg.norm(action - self.last_action_for_reward))
-
-        posture_penalty = -0.3 * (tilt_mag)
-        ang_vel_penalty = -0.02 * ang_vel_mag
-
-        # Use simulator joint readings in training frame to shape lateral stance.
         joint_pos = np.deg2rad(
             [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
         ) * self.train_sim_flip
-        left_hip_roll = float(joint_pos[12])
+        left_hip_roll = -float(joint_pos[12])
         right_hip_roll = float(joint_pos[18])
-        left_ankle_roll = float(joint_pos[16])
+
+        left_ankle_roll = -float(joint_pos[16])
         right_ankle_roll = float(joint_pos[22])
+        left_knee_flex = abs(float(joint_pos[14]))
+        right_knee_flex = abs(float(joint_pos[20]))
+        avg_knee_flex = 0.5 * (left_knee_flex + right_knee_flex)
 
-        hip_spread = left_hip_roll - right_hip_roll
-        ankle_spread = left_ankle_roll - right_ankle_roll
-        stance_metric = 0.6 * abs(hip_spread) + 0.4 * abs(ankle_spread)
+        max_leg_roll = 0.5  # 防止劈叉姿势
+        split_penalty = -0.12 * max(0.0, (left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = -float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
 
-        # Penalize narrow stance (feet too close) and scissoring (cross-leg pattern).
-        stance_collapse_penalty = -4.0 * max(0.0, self.min_stance_rad - stance_metric)
-        cross_leg_penalty = -1.2 * max(0.0, -(hip_spread * ankle_spread))
+        min_leg_separation = 0.04  # 最小腿间距（防止贴得太近）
+        inward_penalty = 0.3 * min(0.0, (left_hip_roll-min_leg_separation)) + 0.3 * min(0.0, (right_hip_roll-min_leg_separation))  # 惩罚左右腿过度内扣
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.12 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.12 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.2  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.6 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.6 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+
+        target_vec = self.target_position - current_pos
+        target_dist = float(np.linalg.norm(target_vec))
+        if target_dist > 1e-6:
+            target_heading = math.atan2(float(target_vec[1]), float(target_vec[0]))
+            robot_heading = math.radians(float(robot.global_orientation_euler[2]))
+            heading_error = self._wrap_to_pi(target_heading - robot_heading)
+            heading_align_reward = self.reward_heading_align_scale * math.cos(heading_error)
+            heading_error_penalty = -self.reward_heading_error_scale * abs(heading_error)
+        else:
+            heading_align_reward = 0.0
+            heading_error_penalty = 0.0
+
+        # Forward-progress reward (distance delta) with anti-stuck shaping.
+        progress_reward_raw = self.reward_progress_scale * dist_delta
+        survival_reward = self.reward_survival_scale
+        smoothness_penalty = -self.reward_smoothness_scale * float(np.linalg.norm(action - self.last_action_for_reward))
+        step_displacement = float(np.linalg.norm(current_pos - previous_pos))
+        accel_signal = 0.0
+        accel_source = "imu_delta"
+        accel_now = np.array(getattr(robot, "accelerometer", np.zeros(3)), dtype=np.float32)
+        if accel_now.shape[0] >= 3:
+            # Use IMU acceleration delta to reduce gravity bias and punish abrupt bursts.
+            accel_signal = float(np.linalg.norm(accel_now[:3] - self._prev_accelerometer[:3]))
+            self._prev_accelerometer = accel_now
+        accel_penalty = -min(
+            self.reward_accel_penalty_cap,
+            self.reward_accel_penalty_scale * accel_signal,
+        )
+        accel_abs = float(np.linalg.norm(accel_now[:3])) if accel_now.shape[0] >= 3 else 0.0
+        accel_abs_over = max(0.0, accel_abs - self.reward_accel_abs_limit)
+        accel_abs_penalty = -min(
+            self.reward_accel_abs_penalty_cap,
+            self.reward_accel_abs_penalty_scale * accel_abs_over,
+        )
+        if self.step_counter > 30 and step_displacement < 0.015 and self.target_distance_wp > 0.3:
+            idle_penalty = -self.reward_idle_penalty_scale
+        else:
+            idle_penalty = 0.0
+
+        if self.step_counter > self.exploration_start_steps:
+            displacement_novelty = step_displacement / max(1e-6, self.stationary_step_eps)
+            exploration_bonus = min(
+                self.exploration_cap,
+                self.exploration_scale * max(0.0, displacement_novelty - self.exploration_target_novelty),
+            )
+        else:
+            exploration_bonus = 0.0
+
+        # Encourage active/varied knee motions early in training without dominating progress reward.
+        left_knee_act = float(action[14])
+        right_knee_act = float(action[20])
+        left_knee_delta = abs(left_knee_act - float(self.last_action_for_reward[14])) 
+        right_knee_delta = abs(right_knee_act - float(self.last_action_for_reward[20])) 
+        knee_action_mag = 0.5 * (abs(left_knee_act) + abs(right_knee_act))
+        knee_action_delta = 0.5 * (left_knee_delta + right_knee_delta)
+        if self.step_counter > 10:
+            knee_explore_reward = min(
+                self.reward_knee_explore_cap,
+                self.reward_knee_explore_scale * knee_action_mag
+                + self.reward_knee_explore_delta_scale * knee_action_delta,
+            )
+        else:
+            knee_explore_reward = 0.0
+
+        # Directly encourage observable knee flexion instead of only action exploration.
+        knee_lift_shortfall_penalty = -self.reward_knee_lift_shortfall_scale * max(
+            0.0, self.reward_knee_lift_target - avg_knee_flex
+        )
+
+        # Encourage hip-pitch exploration to improve forward stride generation.
+        left_hip_pitch_act = float(action[11])
+        right_hip_pitch_act = float(action[17])
+        left_hip_pitch_delta = abs(left_hip_pitch_act - float(self.last_action_for_reward[11]))
+        right_hip_pitch_delta = abs(right_hip_pitch_act - float(self.last_action_for_reward[17]))
+        hip_pitch_action_mag = 0.5 * (abs(left_hip_pitch_act) + abs(right_hip_pitch_act))
+        hip_pitch_action_delta = 0.5 * (left_hip_pitch_delta + right_hip_pitch_delta)
+        if self.step_counter > 10:
+            hip_pitch_explore_reward = min(
+                self.reward_hip_pitch_explore_cap,
+                self.reward_hip_pitch_explore_scale * hip_pitch_action_mag
+                + self.reward_hip_pitch_explore_delta_scale * hip_pitch_action_delta,
+            )
+        else:
+            hip_pitch_explore_reward = 0.0
+
+        if curr_dist_to_target < 0.3:
+            arrival_bonus = self.target_distance_wp * 8 ## 奖励到达目标点
+        else:
+            arrival_bonus = 0.0
             
         target_height = self.initial_height
         height_error =  height - target_height
-        height_penalty = -0.5 * abs(height_error)  # 惩罚高度偏离，系数可调
+        height_error = height - target_height
 
-        # # 在 compute_reward 开头附近，添加高度变化率计算
-        # if not hasattr(self, 'last_height'):
-        #     self.last_height = height
-        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
-        # height_rate = height - self.last_height  # 正为上升，负为下降
-        # self.last_height = height
+        height_penalty = -0.5 * (math.exp(15*abs(height_error))-1)
 
-        # 惩罚高度下降（负变化率）
-        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
 
-        # # 在 compute_reward 中
-        # if self.step_counter > 50:
-        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
-        #     novelty = float(np.linalg.norm(action - avg_prev_action))
-        #     exploration_bonus = 0.05 * novelty
-        # else:
-        #     exploration_bonus = 0
+        # Gate progress reward when posture quality is poor.
+        # Important: include leg overextension so upright torso cannot exploit progress reward.
+        tilt_excess = max(0.0, tilt_mag - self.reward_progress_tilt_gate_start)
+        knee_gate_excess = max(0.0, self.reward_progress_knee_gate_min - avg_knee_flex)
+        left_hip_pitch = float(joint_pos[11])
+        right_hip_pitch = float(joint_pos[17])
+        left_hip_over = max(0.0, abs(left_hip_pitch) - self.reward_hip_overextend_threshold)
+        right_hip_over = max(0.0, abs(right_hip_pitch) - self.reward_hip_overextend_threshold)
+        hip_over_mean = 0.5 * (left_hip_over + right_hip_over)
 
-        # self.prev_action_history[self.history_idx] = action
-        # self.history_idx = (self.history_idx + 1) % 50
+        hip_gate_excess = max(0.0, hip_over_mean - self.reward_progress_hip_gate_over)
+        posture_gate = 1.0 - 1.2 * tilt_excess - 2.0 * knee_gate_excess - 1.8 * hip_gate_excess
+        posture_gate = float(np.clip(posture_gate, self.reward_progress_gate_floor, 1.0))
+        progress_reward = progress_reward_raw * posture_gate
 
+        knee_straight_penalty = -self.reward_knee_straight_penalty_scale * max(
+            0.0, self.reward_knee_straight_threshold - avg_knee_flex
+        )
 
+        hip_overextend_penalty = -self.reward_hip_overextend_penalty_scale * (left_hip_over + right_hip_over)
+
+        # Penalize over-stretched legs even if torso stays upright.
+        stretch_amount = left_hip_over + right_hip_over
+        straight_amount = max(0.0, self.reward_knee_straight_threshold - avg_knee_flex)
+        leg_stretch_penalty = -self.reward_leg_stretch_penalty_scale * stretch_amount * (1.0 + 2.5 * straight_amount)
+
+        # Keep extra combo penalty, but no longer vanish when torso is upright.
+        stretch_lean_combo_penalty = -self.reward_stretch_lean_combo_scale * (0.5 + tilt_mag) * stretch_amount * (1.0 + 3.0 * straight_amount)
+        posture_penalty = -0.6 * (tilt_mag)
         total = (
-            # progress_reward  + 
-                 alive_bonus + 
-                 # lateral_penalty +
-                 # action_penalty +
-                smoothness_penalty +
-                 posture_penalty
-                 + ang_vel_penalty
-                 + height_penalty
-                 + stance_collapse_penalty
-                 + cross_leg_penalty
+            progress_reward
+            + survival_reward
+            + smoothness_penalty
+            + accel_penalty
+            + accel_abs_penalty
+            + idle_penalty
+            + split_penalty
+            + inward_penalty
+            + ankle_roll_penalty
+            + ankle_roll_cross_penalty
+            + left_hip_yaw_penalty
+            + right_hip_yaw_penalty
+            + heading_align_reward
+            + heading_error_penalty
+            # + knee_straight_penalty
+            + hip_overextend_penalty
+            + leg_stretch_penalty
+            + stretch_lean_combo_penalty
             # + exploration_bonus
-                # + height_down_penalty
-                 )
-        if time.time() - self.start_time >= 600:
-            self.start_time = time.time()
-            print(
-                #   f"progress_reward:{progress_reward:.4f}",
-                #   f"lateral_penalty:{lateral_penalty:.4f}",
-                #   f"action_penalty:{action_penalty:.4f}"s, 
-                  f"height_penalty:{height_penalty:.4f}",
-                  f"smoothness_penalty:{smoothness_penalty:.4f},",
-                  f"posture_penalty:{posture_penalty:.4f}",
-                  f"stance_collapse_penalty:{stance_collapse_penalty:.4f}",
-                  f"cross_leg_penalty:{cross_leg_penalty:.4f}",
-                #   f"ang_vel_penalty:{ang_vel_penalty:.4f}",
-                #   f"height_down_penalty:{height_down_penalty:.4f}",
-                #   f"exploration_bonus:{exploration_bonus:.4f}"
+            # + knee_explore_reward
+            # + knee_lift_shortfall_penalty
+            # + hip_pitch_explore_reward
+            + arrival_bonus
+            + height_penalty
+            + posture_penalty
         )
 
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"progress_reward:{progress_reward:.4f},"
+                f"survival_reward:{survival_reward:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"accel_penalty:{accel_penalty:.4f},"
+                f"accel_source:{accel_source},"
+                f"accel_signal:{accel_signal:.4f},"
+                f"accel_abs:{accel_abs:.4f},"
+                f"accel_abs_penalty:{accel_abs_penalty:.4f},"
+                f"idle_penalty:{idle_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"heading_align_reward:{heading_align_reward:.4f},"
+                f"heading_error_penalty:{heading_error_penalty:.4f},"
+                f"knee_straight_penalty:{knee_straight_penalty:.4f},"
+                f"hip_overextend_penalty:{hip_overextend_penalty:.4f},"
+                f"leg_stretch_penalty:{leg_stretch_penalty:.4f},"
+                f"stretch_lean_combo_penalty:{stretch_lean_combo_penalty:.4f},"
+                f"posture_gate:{posture_gate:.4f},"
+                f"progress_reward_raw:{progress_reward_raw:.4f},"
+                # f"exploration_bonus:{exploration_bonus:.4f},"
+                f"height_penalty:{height_penalty:.4f},"
+                # f"knee_explore_reward:{knee_explore_reward:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                # f"knee_lift_shortfall_penalty:{knee_lift_shortfall_penalty:.4f},"
+                # f"hip_pitch_explore_reward:{hip_pitch_explore_reward:.4f},"
+                f"arrival_bonus:{arrival_bonus:.4f},"
+                f"total:{total:.4f}"
+            )
         return total
 
 
@@ -514,7 +771,32 @@ class WalkEnv(gym.Env):
     def step(self, action):
 
         r = self.Player.robot
-        self.previous_action = action
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        # Loosen upper-body constraints: keep motion bounded but no longer hard-lock head/arms/waist.
+        action[0:2] = 0
+        action[3] = np.clip(action[3], 3, 5)
+        action[7] = np.clip(action[7], -5, -3)
+        action[2] = np.clip(action[2], -6, 6)
+        action[6] = np.clip(action[6], -6, 6)
+        action[4] = 0
+        action[5] = np.clip(action[5], -8, -2)
+        action[8] = 0
+        action[9] = np.clip(action[9], 8, 2)
+        action[10] = np.clip(action[10], -0.6, 0.6)
+        # Boost knee command range so policy can produce visible knee flexion earlier.
+        action[14] = np.clip(action[14], 0, 10.0)
+        action[20] = np.clip(action[20], -10.0, 0)
+        # action[14] = 1 # the correct left knee sign
+        # action[20] = -1 # the correct right knee sign
+        # action[11] = 2
+        # action[17] = -2
+        # action[12] = -1
+        # action[18] = 1
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
 
         self.target_joint_positions = (
                 # self.joint_nominal_position +
@@ -524,10 +806,10 @@ class WalkEnv(gym.Env):
 
         for idx, target in enumerate(self.target_joint_positions):
             r.set_motor_target_position(
-                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=40, kd=1.0
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=60, kd=1.2
             )
 
-        self.previous_action = action
+        self.previous_action = action.copy()
 
         self.sync()  # run simulation step
         self.step_counter += 1
@@ -539,13 +821,28 @@ class WalkEnv(gym.Env):
 
         # Compute reward based on movement from previous step
         reward = self.compute_reward(self.previous_pos, current_pos, action)
-
-        # Update previous position
         self.previous_pos = current_pos.copy()
+
+        self.prev_action_history[self.history_idx] = action.copy()
+        self.history_idx = (self.history_idx + 1) % self.action_history_len
+
         self.last_action_for_reward = action.copy()
         
+        # Check if current waypoint is reached
+        if self.train_stage != "in_place":
+            dist_to_waypoint = float(np.linalg.norm(current_pos - self.target_position))
+            if dist_to_waypoint < self.waypoint_reach_distance:
+                # Move to next waypoint
+                self.waypoint_index += 1
+                if self.waypoint_index >= len(self.point_list):
+                    # All waypoints completed
+                    self.route_completed = True
+                else:
+                    # Update target to next waypoint
+                    self.target_position = self.point_list[self.waypoint_index]
+
         # Fall detection and penalty
-        is_fallen = self.Player.world.global_position[2] < 0.55
+        is_fallen = self.Player.world.global_position[2] < 0.45
 
         # terminal state: the robot is falling or timeout
         terminated = is_fallen or self.step_counter > 800 or self.route_completed
@@ -561,14 +858,21 @@ class Train(Train_Base):
     def train(self, args):
 
         # --------------------------------------- Learning parameters
-        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
-        if n_envs < 1:
-            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
-        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
-        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "256"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
-        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
-        total_steps = 30000000
-        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        n_envs = 20
+        server_warmup_sec = 3.0
+        n_steps_per_env = 256  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = 512  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 90000000
+        learning_rate = 2e-4
+        ent_coef = 0.035
+        clip_range = 0.2
+        gamma = 0.97
+        n_epochs = 3
+        enable_eval = True
+        monitor_train_env = False
+        eval_freq_mult = 60
+        save_freq_mult = 60
+        eval_eps = 7
         folder_name = f'Walk_R{self.robot_type}'
         model_path = f'./scripts/gyms/logs/{folder_name}/'
 
@@ -585,6 +889,10 @@ class Train(Train_Base):
 
             return thunk
 
+        env = None
+        eval_env = None
+        servers = None
+        try:
             server_log_dir = os.path.join(model_path, "server_logs")
             os.makedirs(server_log_dir, exist_ok=True)
             servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
@@ -596,11 +904,11 @@ class Train(Train_Base):
                 sleep(server_warmup_sec)
             print("Servers started, creating environments...")
 
-        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)])
+            env = SubprocVecEnv([init_env(i, monitor=monitor_train_env) for i in range(n_envs)], start_method="spawn")
             # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+            if enable_eval:
                 eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
 
-        try:
             # Custom policy network architecture
             policy_kwargs = dict(
                 net_arch=dict(
@@ -623,26 +931,30 @@ class Train(Train_Base):
                     learning_rate=learning_rate,
                     device="cpu",
                     policy_kwargs=policy_kwargs,
-                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
-                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    ent_coef=ent_coef,  # Entropy coefficient for exploration
+                    clip_range=clip_range,  # PPO clipping parameter
                     gae_lambda=0.95,  # GAE lambda
-                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    gamma=gamma, # Discount factor
                     # target_kl=0.03,
-                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    n_epochs=n_epochs,
                     tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
                 )
 
             model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
-                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=100,
+                                          eval_freq=n_steps_per_env * max(1, eval_freq_mult),
+                                          save_freq=n_steps_per_env * max(1, save_freq_mult),
+                                          eval_eps=max(1, eval_eps),
                                           backup_env_file=__file__)
         except KeyboardInterrupt:
             sleep(1)  # wait for child processes
             print("\nctrl+c pressed, aborting...\n")
-            servers.kill()
             return
-
+        finally:
+            if env is not None:
                 env.close()
+            if eval_env is not None:
                 eval_env.close()
+            if servers is not None:
                 servers.kill()
 
     def test(self, args):
@@ -650,8 +962,8 @@ class Train(Train_Base):
         # Uses different server and monitor ports
         server_log_dir = os.path.join(args["folder_dir"], "server_logs")
         os.makedirs(server_log_dir, exist_ok=True)
-        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
-        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+        test_no_render = False
+        test_no_realtime = False
 
         server = Train_Server(
             self.server_p - 1,
@@ -694,8 +1006,8 @@ if __name__ == "__main__":
     run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
 
     if run_mode == "test":
-        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Walk_R0_004/best_model.zip")
-        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Walk_R0_004/")
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
         trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
     else:
         retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()

scripts/gyms/logs/Turn_R0_000/Walk.py

@@ -0,0 +1,832 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,
+                0.0,
+                0.0,
+                1.4,
+                0.0,
+                -0.4,
+                0.0,
+                -1.4,
+                0.0,
+                0.4,
+                0.0,
+                -0.4,
+                0.0,
+                0.0,
+                0.8,
+                -0.4,
+                0.0,
+                0.4,
+                0.0,
+                0.0,
+                -0.8,
+                0.4,
+                0.0,
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "0.7"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    def compute_reward(self, previous_pos, current_pos, action):
+        print(time.time(), self.step_counter)
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        # is_fallen = height < 0.55
+        # if is_fallen:
+        #     remain = max(0, 800 - self.step_counter)
+        #     # Strong terminal penalty discourages risky turn-and-fall behaviors.
+        #     return -1
+        
+
+
+        # # 目标方向
+        # to_target = self.target_position - current_pos
+        # dist_to_target = float(np.linalg.norm(to_target))
+        # if dist_to_target < 0.5:
+        #     return 15.0
+
+        # forward_dir = to_target / dist_to_target if dist_to_target > 0.1 else np.array([1.0, 0.0])
+        # delta_pos = current_pos - previous_pos
+        # forward_step = float(np.dot(delta_pos, forward_dir))
+        # lateral_step = float(np.linalg.norm(delta_pos - forward_dir * forward_step))
+
+        # Keep reward simple: turn correctly, stay stable, avoid jerky actions.
+        
+        delta_action_norm = float(np.linalg.norm(action - self.last_action_for_reward))
+        # Cap smoothness penalty so it regularizes behavior without dominating total reward.
+        smoothness_penalty = -min(self.reward_smoothness_cap, self.reward_smoothness_scale * delta_action_norm)
+
+        posture_penalty = -0.45 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.04 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        hip_spread = left_hip_roll - right_hip_roll
+        ankle_spread = left_ankle_roll - right_ankle_roll
+        stance_metric = 0.6 * abs(hip_spread) + 0.4 * abs(ankle_spread)
+
+        # Penalize narrow stance (feet too close) and scissoring (cross-leg pattern).
+        stance_collapse_penalty = -4 * max(0.0, self.min_stance_rad - stance_metric)
+        cross_leg_penalty = -2.5 * max(0.0, -(hip_spread * ankle_spread))
+
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = self._wrap_to_pi(target_yaw - robot_yaw)
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+
+        # Reward reducing heading error between consecutive steps.
+        # Use a deadzone and smaller gain to avoid high-frequency jitter near alignment.
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward = 0.70 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.70, 0.70))
+        self.last_yaw_error = yaw_error
+
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+        turn_dir = float(np.sign(yaw_error))
+        # Continuous turn shaping prevents reward discontinuity near small heading error.
+        turn_gate = min(1.0, abs_yaw_error / math.radians(45.0))
+        turn_rate_reward = 0.70 * turn_gate * math.tanh(2.0 * turn_dir * yaw_rate)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(8.0) else 0.0
+        # After roughly aligning with target, prioritize standing stability over continued aggressive turning.
+        aligned_gate = max(0.0, 1.0 - abs_yaw_error / math.radians(18.0))
+        post_turn_ang_vel_penalty = -0.10 * aligned_gate * min(rp_ang_vel_mag, math.radians(60.0))
+        lower_body_speed_mag = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        post_turn_pose_bonus = 0.30 * aligned_gate * math.exp(-tilt_mag / 0.20) * math.exp(-lower_body_speed_mag / 1.10)
+        # Keep feet separation when aligned so robot does not collapse stance after turning.
+        aligned_stance_bonus = 0.20 * aligned_gate * min(1.0, stance_metric / max(self.min_stance_rad, 1e-4))
+        # Once roughly aligned, damp yaw oscillation and reward keeping a stable stance.
+        anti_oscillation_penalty = -0.08 * min(yaw_rate_abs, math.radians(35.0)) if abs_yaw_error < math.radians(7.0) else 0.0
+        stabilize_bonus = 0.45 if (
+            abs_yaw_error < math.radians(8.0)
+            and yaw_rate_abs < math.radians(10.0)
+            and tilt_mag < 0.28
+        ) else 0.0
+
+        # 改进（线性分段，sigmoid 过渡）
+        if abs_yaw_error < math.radians(15.0):
+            alive_bonus = 2 * (1.0 - abs_yaw_error / math.radians(15.0)) ** 0.5  # 平方根让小角度更敏感
+        else:
+            alive_bonus = max(0.1, 2 * (1.0 - (abs_yaw_error - math.radians(15.0)) / math.radians(75.0)))
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        # 改进（分段，偏离越多惩罚越重）
+        height_error = height - target_height
+        if abs(height_error) < 0.04:
+            height_penalty = -2.5 * abs(height_error)  # 小偏离，保持线性
+        else:
+            height_penalty = -2.5 * 0.04 - 4.0 * (abs(height_error) - 0.04)  # 大偏离，惩罚加速
+
+        total = (
+            alive_bonus
+            + smoothness_penalty
+            + posture_penalty
+            + ang_vel_penalty
+            + linkage_reward
+            + waist_only_turn_penalty
+            + yaw_link_reward
+            + head_toward_bonus
+            + heading_progress_reward
+            + anti_oscillation_penalty
+            + stabilize_bonus
+            + height_penalty
+            # + post_turn_ang_vel_penalty
+            # + post_turn_pose_bonus
+            # + aligned_stance_bonus
+            # + heading_align_reward
+            + turn_rate_reward
+            # + stance_collapse_penalty
+            # + cross_leg_penalty
+        )
+
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            # print(
+            #     f"reward_debug: step={self.step_counter}, "
+            #     f"alive_bonus:{alive_bonus:.4f}, "
+            #     # f"heading_align_reward:{heading_align_reward:.4f}, "
+            #     # f"heading_progress_reward:{heading_progress_reward:.4f}, "
+            #     f"head_towards_bonus:{head_toward_bonus},"
+            #     f"posture_penalty:{posture_penalty:.4f}, "
+            #     f"ang_vel_penalty:{ang_vel_penalty:.4f}, "
+            #     f"smoothness_penalty:{smoothness_penalty:.4f}, "
+            #     f"linkage_reward:{linkage_reward:.4f}, "
+            #     f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f}, "
+            #     f"yaw_link_reward:{yaw_link_reward:.4f}, "
+            #     f"anti_oscillation_penalty:{anti_oscillation_penalty:.4f}, "
+            #     f"stabilize_bonus:{stabilize_bonus:.4f}, "
+            #     f"turn_rate_reward:{turn_rate_reward:.4f}, "
+            #     f"total:{total:.4f}"
+            # )
+            
+            self.debug_log(
+                f"reward_debug: step={self.step_counter}, "
+                f"alive_bonus:{alive_bonus:.4f}, "
+                # f"heading_align_reward:{heading_align_reward:.4f}, "
+                f"heading_progress_reward:{heading_progress_reward:.4f}, "
+                f"head_towards_bonus:{head_toward_bonus},"
+                f"posture_penalty:{posture_penalty:.4f}, "
+                f"ang_vel_penalty:{ang_vel_penalty:.4f}, "
+                f"smoothness_penalty:{smoothness_penalty:.4f}, "
+                f"heading_progress_reward:{heading_progress_reward:.4f}, "
+                f"linkage_reward:{linkage_reward:.4f}, "
+                f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f}, "
+                f"yaw_link_reward:{yaw_link_reward:.4f}, "
+                f"anti_oscillation_penalty:{anti_oscillation_penalty:.4f}, "
+                f"stabilize_bonus:{stabilize_bonus:.4f}, "
+                f"height_penalty:{height_penalty:.4f}, "
+                # f"post_turn_ang_vel_penalty:{post_turn_ang_vel_penalty:.4f}, "
+                # f"post_turn_pose_bonus:{post_turn_pose_bonus:.4f}, "
+                f"aligned_stance_bonus:{aligned_stance_bonus:.4f}, "
+                # f"turn_rate_reward:{turn_rate_reward:.4f}, "
+                f"stance_collapse_penalty:{stance_collapse_penalty:.4f}, "
+                f"cross_leg_penalty:{cross_leg_penalty:.4f}, "
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.1# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=40, kd=1.2
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        # Update previous position
+        self.previous_pos = current_pos.copy()
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "256"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/",
+                    max_grad_norm=float(os.environ.get("GYM_CPU_TRAIN_MAX_GRAD_NORM", "0.5"))
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=5,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_001/Walk.py

@@ -0,0 +1,831 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,
+                0.0,
+                0.0,
+                1.4,
+                0.0,
+                -0.4,
+                0.0,
+                -1.4,
+                0.0,
+                0.4,
+                0.0,
+                -0.4,
+                0.0,
+                0.0,
+                0.8,
+                -0.4,
+                0.0,
+                0.4,
+                0.0,
+                0.0,
+                -0.8,
+                0.4,
+                0.0,
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "0.7"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        # is_fallen = height < 0.55
+        # if is_fallen:
+        #     remain = max(0, 800 - self.step_counter)
+        #     # Strong terminal penalty discourages risky turn-and-fall behaviors.
+        #     return -1
+        
+
+
+        # # 目标方向
+        # to_target = self.target_position - current_pos
+        # dist_to_target = float(np.linalg.norm(to_target))
+        # if dist_to_target < 0.5:
+        #     return 15.0
+
+        # forward_dir = to_target / dist_to_target if dist_to_target > 0.1 else np.array([1.0, 0.0])
+        # delta_pos = current_pos - previous_pos
+        # forward_step = float(np.dot(delta_pos, forward_dir))
+        # lateral_step = float(np.linalg.norm(delta_pos - forward_dir * forward_step))
+
+        # Keep reward simple: turn correctly, stay stable, avoid jerky actions.
+        
+        delta_action_norm = float(np.linalg.norm(action - self.last_action_for_reward))
+        # Cap smoothness penalty so it regularizes behavior without dominating total reward.
+        smoothness_penalty = -min(self.reward_smoothness_cap, self.reward_smoothness_scale * delta_action_norm)
+
+        posture_penalty = -0.45 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.04 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        hip_spread = left_hip_roll - right_hip_roll
+        ankle_spread = left_ankle_roll - right_ankle_roll
+        stance_metric = 0.6 * abs(hip_spread) + 0.4 * abs(ankle_spread)
+
+        # Penalize narrow stance (feet too close) and scissoring (cross-leg pattern).
+        stance_collapse_penalty = -4 * max(0.0, self.min_stance_rad - stance_metric)
+        cross_leg_penalty = -2.5 * max(0.0, -(hip_spread * ankle_spread))
+
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = self._wrap_to_pi(target_yaw - robot_yaw)
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+
+        # Reward reducing heading error between consecutive steps.
+        # Use a deadzone and smaller gain to avoid high-frequency jitter near alignment.
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward = 0.70 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.70, 0.70))
+        self.last_yaw_error = yaw_error
+
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+        turn_dir = float(np.sign(yaw_error))
+        # Continuous turn shaping prevents reward discontinuity near small heading error.
+        turn_gate = min(1.0, abs_yaw_error / math.radians(45.0))
+        turn_rate_reward = 0.70 * turn_gate * math.tanh(2.0 * turn_dir * yaw_rate)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(8.0) else 0.0
+        # After roughly aligning with target, prioritize standing stability over continued aggressive turning.
+        aligned_gate = max(0.0, 1.0 - abs_yaw_error / math.radians(18.0))
+        post_turn_ang_vel_penalty = -0.10 * aligned_gate * min(rp_ang_vel_mag, math.radians(60.0))
+        lower_body_speed_mag = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        post_turn_pose_bonus = 0.30 * aligned_gate * math.exp(-tilt_mag / 0.20) * math.exp(-lower_body_speed_mag / 1.10)
+        # Keep feet separation when aligned so robot does not collapse stance after turning.
+        aligned_stance_bonus = 0.20 * aligned_gate * min(1.0, stance_metric / max(self.min_stance_rad, 1e-4))
+        # Once roughly aligned, damp yaw oscillation and reward keeping a stable stance.
+        anti_oscillation_penalty = -0.08 * min(yaw_rate_abs, math.radians(35.0)) if abs_yaw_error < math.radians(7.0) else 0.0
+        stabilize_bonus = 0.45 if (
+            abs_yaw_error < math.radians(8.0)
+            and yaw_rate_abs < math.radians(10.0)
+            and tilt_mag < 0.28
+        ) else 0.0
+
+        # 改进（线性分段，sigmoid 过渡）
+        if abs_yaw_error < math.radians(15.0):
+            alive_bonus = 2 * (1.0 - abs_yaw_error / math.radians(15.0)) ** 0.5  # 平方根让小角度更敏感
+        else:
+            alive_bonus = max(0.1, 2 * (1.0 - (abs_yaw_error - math.radians(15.0)) / math.radians(75.0)))
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        # 改进（分段，偏离越多惩罚越重）
+        height_error = height - target_height
+        if abs(height_error) < 0.04:
+            height_penalty = -2.5 * abs(height_error)  # 小偏离，保持线性
+        else:
+            height_penalty = -2.5 * 0.04 - 4.0 * (abs(height_error) - 0.04)  # 大偏离，惩罚加速
+
+        total = (
+            alive_bonus
+            + smoothness_penalty
+            + posture_penalty
+            + ang_vel_penalty
+            + linkage_reward
+            + waist_only_turn_penalty
+            + yaw_link_reward
+            + head_toward_bonus
+            + heading_progress_reward
+            + anti_oscillation_penalty
+            + stabilize_bonus
+            + height_penalty
+            # + post_turn_ang_vel_penalty
+            # + post_turn_pose_bonus
+            # + aligned_stance_bonus
+            # + heading_align_reward
+            + turn_rate_reward
+            # + stance_collapse_penalty
+            # + cross_leg_penalty
+        )
+
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            # print(
+            #     f"reward_debug: step={self.step_counter}, "
+            #     f"alive_bonus:{alive_bonus:.4f}, "
+            #     # f"heading_align_reward:{heading_align_reward:.4f}, "
+            #     # f"heading_progress_reward:{heading_progress_reward:.4f}, "
+            #     f"head_towards_bonus:{head_toward_bonus},"
+            #     f"posture_penalty:{posture_penalty:.4f}, "
+            #     f"ang_vel_penalty:{ang_vel_penalty:.4f}, "
+            #     f"smoothness_penalty:{smoothness_penalty:.4f}, "
+            #     f"linkage_reward:{linkage_reward:.4f}, "
+            #     f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f}, "
+            #     f"yaw_link_reward:{yaw_link_reward:.4f}, "
+            #     f"anti_oscillation_penalty:{anti_oscillation_penalty:.4f}, "
+            #     f"stabilize_bonus:{stabilize_bonus:.4f}, "
+            #     f"turn_rate_reward:{turn_rate_reward:.4f}, "
+            #     f"total:{total:.4f}"
+            # )
+            
+            self.debug_log(
+                f"reward_debug: step={self.step_counter}, "
+                f"alive_bonus:{alive_bonus:.4f}, "
+                # f"heading_align_reward:{heading_align_reward:.4f}, "
+                f"heading_progress_reward:{heading_progress_reward:.4f}, "
+                f"head_towards_bonus:{head_toward_bonus},"
+                f"posture_penalty:{posture_penalty:.4f}, "
+                f"ang_vel_penalty:{ang_vel_penalty:.4f}, "
+                f"smoothness_penalty:{smoothness_penalty:.4f}, "
+                f"heading_progress_reward:{heading_progress_reward:.4f}, "
+                f"linkage_reward:{linkage_reward:.4f}, "
+                f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f}, "
+                f"yaw_link_reward:{yaw_link_reward:.4f}, "
+                f"anti_oscillation_penalty:{anti_oscillation_penalty:.4f}, "
+                f"stabilize_bonus:{stabilize_bonus:.4f}, "
+                f"height_penalty:{height_penalty:.4f}, "
+                # f"post_turn_ang_vel_penalty:{post_turn_ang_vel_penalty:.4f}, "
+                # f"post_turn_pose_bonus:{post_turn_pose_bonus:.4f}, "
+                f"aligned_stance_bonus:{aligned_stance_bonus:.4f}, "
+                # f"turn_rate_reward:{turn_rate_reward:.4f}, "
+                f"stance_collapse_penalty:{stance_collapse_penalty:.4f}, "
+                f"cross_leg_penalty:{cross_leg_penalty:.4f}, "
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.1# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=40, kd=1.2
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        # Update previous position
+        self.previous_pos = current_pos.copy()
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "256"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/",
+                    max_grad_norm=float(os.environ.get("GYM_CPU_TRAIN_MAX_GRAD_NORM", "0.5"))
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=5,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_002/Walk.py

@@ -0,0 +1,831 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,
+                0.0,
+                0.0,
+                1.4,
+                0.0,
+                -0.4,
+                0.0,
+                -1.4,
+                0.0,
+                0.4,
+                0.0,
+                -0.4,
+                0.0,
+                0.0,
+                0.8,
+                -0.4,
+                0.0,
+                0.4,
+                0.0,
+                0.0,
+                -0.8,
+                0.4,
+                0.0,
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        # is_fallen = height < 0.55
+        # if is_fallen:
+        #     remain = max(0, 800 - self.step_counter)
+        #     # Strong terminal penalty discourages risky turn-and-fall behaviors.
+        #     return -1
+        
+
+
+        # # 目标方向
+        # to_target = self.target_position - current_pos
+        # dist_to_target = float(np.linalg.norm(to_target))
+        # if dist_to_target < 0.5:
+        #     return 15.0
+
+        # forward_dir = to_target / dist_to_target if dist_to_target > 0.1 else np.array([1.0, 0.0])
+        # delta_pos = current_pos - previous_pos
+        # forward_step = float(np.dot(delta_pos, forward_dir))
+        # lateral_step = float(np.linalg.norm(delta_pos - forward_dir * forward_step))
+
+        # Keep reward simple: turn correctly, stay stable, avoid jerky actions.
+        
+        delta_action_norm = float(np.linalg.norm(action - self.last_action_for_reward))
+        # Cap smoothness penalty so it regularizes behavior without dominating total reward.
+        smoothness_penalty = -min(self.reward_smoothness_cap, self.reward_smoothness_scale * delta_action_norm)
+
+        posture_penalty = -0.45 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.04 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        hip_spread = left_hip_roll - right_hip_roll
+        ankle_spread = left_ankle_roll - right_ankle_roll
+        stance_metric = 0.5 * abs(hip_spread) + 0.5 * abs(ankle_spread)
+
+        # Penalize narrow stance (feet too close) and scissoring (cross-leg pattern).
+        stance_collapse_penalty = -3 * max(0.0, self.min_stance_rad - stance_metric)
+        cross_leg_penalty = -2.5 * max(0.0, -(hip_spread * ankle_spread))
+
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = self._wrap_to_pi(target_yaw - robot_yaw)
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+
+        # Reward reducing heading error between consecutive steps.
+        # Use a deadzone and smaller gain to avoid high-frequency jitter near alignment.
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward = progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, 1, 1))
+        self.last_yaw_error = yaw_error
+
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+        turn_dir = float(np.sign(yaw_error))
+        # Continuous turn shaping prevents reward discontinuity near small heading error.
+        turn_gate = min(1.0, abs_yaw_error / math.radians(45.0))
+        turn_rate_reward = 0.70 * turn_gate * math.tanh(2.0 * turn_dir * yaw_rate)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(8.0) else 0.0
+        # After roughly aligning with target, prioritize standing stability over continued aggressive turning.
+        aligned_gate = max(0.0, 1.0 - abs_yaw_error / math.radians(18.0))
+        post_turn_ang_vel_penalty = -0.10 * aligned_gate * min(rp_ang_vel_mag, math.radians(60.0))
+        lower_body_speed_mag = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        post_turn_pose_bonus = 0.30 * aligned_gate * math.exp(-tilt_mag / 0.20) * math.exp(-lower_body_speed_mag / 1.10)
+        # Keep feet separation when aligned so robot does not collapse stance after turning.
+        aligned_stance_bonus = 0.20 * aligned_gate * min(1.0, stance_metric / max(self.min_stance_rad, 1e-4))
+        # Once roughly aligned, damp yaw oscillation and reward keeping a stable stance.
+        anti_oscillation_penalty = -0.08 * min(yaw_rate_abs, math.radians(35.0)) if abs_yaw_error < math.radians(7.0) else 0.0
+        stabilize_bonus = 0.6 if (
+            abs_yaw_error < math.radians(8.0)
+            and yaw_rate_abs < math.radians(10.0)
+            and tilt_mag < 0.28
+        ) else 0.0
+
+        # 改进（线性分段，sigmoid 过渡）
+        if abs_yaw_error < math.radians(15.0):
+            alive_bonus = 2 * (1.0 - abs_yaw_error / math.radians(15.0)) ** 0.5  # 平方根让小角度更敏感
+        else:
+            alive_bonus = max(0.1, 2 * (1.0 - (abs_yaw_error - math.radians(15.0)) / math.radians(75.0)))
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        # 改进（分段，偏离越多惩罚越重）
+        height_error = height - target_height
+        if abs(height_error) < 0.04:
+            height_penalty = -2.5 * abs(height_error)  # 小偏离，保持线性
+        else:
+            height_penalty = -2.5 * 0.04 - 4.0 * (abs(height_error) - 0.04)  # 大偏离，惩罚加速
+
+        total = (
+            alive_bonus
+            + smoothness_penalty
+            + posture_penalty
+            + ang_vel_penalty
+            + linkage_reward
+            + waist_only_turn_penalty
+            + yaw_link_reward
+            + head_toward_bonus
+            + heading_progress_reward
+            + anti_oscillation_penalty
+            + stabilize_bonus
+            + height_penalty
+            # + post_turn_ang_vel_penalty
+            # + post_turn_pose_bonus
+            # + aligned_stance_bonus
+            # + heading_align_reward
+            + turn_rate_reward
+            + stance_collapse_penalty
+            + cross_leg_penalty
+        )
+
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            # print(
+            #     f"reward_debug: step={self.step_counter}, "
+            #     f"alive_bonus:{alive_bonus:.4f}, "
+            #     # f"heading_align_reward:{heading_align_reward:.4f}, "
+            #     # f"heading_progress_reward:{heading_progress_reward:.4f}, "
+            #     f"head_towards_bonus:{head_toward_bonus},"
+            #     f"posture_penalty:{posture_penalty:.4f}, "
+            #     f"ang_vel_penalty:{ang_vel_penalty:.4f}, "
+            #     f"smoothness_penalty:{smoothness_penalty:.4f}, "
+            #     f"linkage_reward:{linkage_reward:.4f}, "
+            #     f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f}, "
+            #     f"yaw_link_reward:{yaw_link_reward:.4f}, "
+            #     f"anti_oscillation_penalty:{anti_oscillation_penalty:.4f}, "
+            #     f"stabilize_bonus:{stabilize_bonus:.4f}, "
+            #     f"turn_rate_reward:{turn_rate_reward:.4f}, "
+            #     f"total:{total:.4f}"
+            # )
+            
+            self.debug_log(
+                f"reward_debug: step={self.step_counter}, "
+                f"alive_bonus:{alive_bonus:.4f}, "
+                # f"heading_align_reward:{heading_align_reward:.4f}, "
+                f"heading_progress_reward:{heading_progress_reward:.4f}, "
+                f"head_towards_bonus:{head_toward_bonus},"
+                f"posture_penalty:{posture_penalty:.4f}, "
+                f"ang_vel_penalty:{ang_vel_penalty:.4f}, "
+                f"smoothness_penalty:{smoothness_penalty:.4f}, "
+                f"heading_progress_reward:{heading_progress_reward:.4f}, "
+                f"linkage_reward:{linkage_reward:.4f}, "
+                f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f}, "
+                f"yaw_link_reward:{yaw_link_reward:.4f}, "
+                f"anti_oscillation_penalty:{anti_oscillation_penalty:.4f}, "
+                f"stabilize_bonus:{stabilize_bonus:.4f}, "
+                f"height_penalty:{height_penalty:.4f}, "
+                # f"post_turn_ang_vel_penalty:{post_turn_ang_vel_penalty:.4f}, "
+                # f"post_turn_pose_bonus:{post_turn_pose_bonus:.4f}, "
+                f"aligned_stance_bonus:{aligned_stance_bonus:.4f}, "
+                # f"turn_rate_reward:{turn_rate_reward:.4f}, "
+                f"stance_collapse_penalty:{stance_collapse_penalty:.4f}, "
+                f"cross_leg_penalty:{cross_leg_penalty:.4f}, "
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.1# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=40, kd=1.2
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        # Update previous position
+        self.previous_pos = current_pos.copy()
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "256"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/",
+                    max_grad_norm=float(os.environ.get("GYM_CPU_TRAIN_MAX_GRAD_NORM", "0.5"))
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_003/Walk.py

@@ -0,0 +1,755 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,
+                0.0,
+                0.0,
+                1.4,
+                0.0,
+                -0.4,
+                0.0,
+                -1.4,
+                0.0,
+                0.4,
+                0.0,
+                -0.4,
+                0.0,
+                0.0,
+                0.8,
+                -0.4,
+                0.0,
+                0.4,
+                0.0,
+                0.0,
+                -0.8,
+                0.4,
+                0.0,
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -1.0
+        
+
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.01 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.45 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.04 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        hip_spread = left_hip_roll - right_hip_roll
+        ankle_spread = left_ankle_roll - right_ankle_roll
+        stance_metric = 0.6 * abs(hip_spread) + 0.4 * abs(ankle_spread)
+
+        # Penalize narrow stance (feet too close) and scissoring (cross-leg pattern).
+        stance_collapse_penalty = -4.0 * max(0.0, self.min_stance_rad - stance_metric)
+        cross_leg_penalty = -1.2 * max(0.0, -(hip_spread * ankle_spread))
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_penalty = -0.5 * abs(height_error)  # 惩罚高度偏离，系数可调
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                 alive_bonus + 
+                 # lateral_penalty +
+                 # action_penalty +
+                smoothness_penalty +
+                 posture_penalty
+                 + ang_vel_penalty
+                 + height_penalty
+                 + stance_collapse_penalty
+                 + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f}",
+                f"smoothness_penalty:{smoothness_penalty:.4f},",
+                f"posture_penalty:{posture_penalty:.4f}",
+                f"stance_collapse_penalty:{stance_collapse_penalty:.4f}",
+                f"cross_leg_penalty:{cross_leg_penalty:.4f}",
+                f"ang_vel_penalty:{ang_vel_penalty:.4f}",
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+        if time.time() - self.start_time >= 600:
+            self.start_time = time.time()
+            self.debug_log(
+                #   f"progress_reward:{progress_reward:.4f}",
+                #   f"lateral_penalty:{lateral_penalty:.4f}",
+                #   f"action_penalty:{action_penalty:.4f}"s, 
+                  f"height_penalty:{height_penalty:.4f}",
+                  f"smoothness_penalty:{smoothness_penalty:.4f},",
+                  f"posture_penalty:{posture_penalty:.4f}",
+                  f"stance_collapse_penalty:{stance_collapse_penalty:.4f}",
+                  f"cross_leg_penalty:{cross_leg_penalty:.4f}",
+                #   f"ang_vel_penalty:{ang_vel_penalty:.4f}",
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+            )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.1# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=40, kd=1.2
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        # Update previous position
+        self.previous_pos = current_pos.copy()
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/",
+                    max_grad_norm=float(os.environ.get("GYM_CPU_TRAIN_MAX_GRAD_NORM", "0.5"))
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_004/Walk.py

@@ -0,0 +1,754 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,
+                0.0,
+                0.0,
+                1.4,
+                0.0,
+                -0.4,
+                0.0,
+                -1.4,
+                0.0,
+                0.4,
+                0.0,
+                -0.4,
+                0.0,
+                0.0,
+                0.8,
+                -0.4,
+                0.0,
+                0.4,
+                0.0,
+                0.0,
+                -0.8,
+                0.4,
+                0.0,
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -1.0
+        
+
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.7, 0.7))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.02 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        hip_spread = left_hip_roll - right_hip_roll if right_hip_roll > 0.03 and left_hip_roll > 0.03 else 0.0
+        ankle_spread = left_ankle_roll - right_ankle_roll if right_ankle_roll > 0.03 and left_ankle_roll > 0.03 else 0.0
+        stance_metric = 0.6 * abs(hip_spread) + 0.4 * abs(ankle_spread)
+
+        # Penalize narrow stance (feet too close) and scissoring (cross-leg pattern).
+        stance_collapse_penalty = -4.0 * max(0.0, self.min_stance_rad - stance_metric) 
+        cross_leg_penalty = -1.2 * max(0.0, -(hip_spread * ankle_spread))
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -math.exp(15*abs(height_error))
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.1# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=40, kd=1.2
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        # Update previous position
+        self.previous_pos = current_pos.copy()
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/",
+                    max_grad_norm=float(os.environ.get("GYM_CPU_TRAIN_MAX_GRAD_NORM", "0.5"))
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_005/Walk.py

@@ -0,0 +1,755 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,
+                0.0,
+                0.0,
+                1.4,
+                0.0,
+                -0.4,
+                0.0,
+                -1.4,
+                0.0,
+                0.4,
+                0.0,
+                -0.4,
+                0.0,
+                0.0,
+                0.8,
+                -0.4,
+                0.0,
+                0.4,
+                0.0,
+                0.0,
+                -0.8,
+                0.4,
+                0.0,
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -1.0
+        
+
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.7, 0.7))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.02 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        hip_spread = left_hip_roll - right_hip_roll if right_hip_roll > 0.03 and left_hip_roll > 0.03 else 0.0
+        ankle_spread = left_ankle_roll - right_ankle_roll if right_ankle_roll > 0.03 and left_ankle_roll > 0.03 else 0.0
+        stance_metric = 0.6 * abs(hip_spread) + 0.4 * abs(ankle_spread)
+
+        # Penalize narrow stance (feet too close) and scissoring (cross-leg pattern).
+        stance_collapse_penalty = -4.0 * max(0.0, self.min_stance_rad - stance_metric) 
+        cross_leg_penalty = -1.2 * max(0.0, -(hip_spread * ankle_spread))
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.1# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=40, kd=1.2
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        # Update previous position
+        self.previous_pos = current_pos.copy()
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/",
+                    max_grad_norm=float(os.environ.get("GYM_CPU_TRAIN_MAX_GRAD_NORM", "0.5"))
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_006/Walk.py

@@ -0,0 +1,821 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -1.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -0.1 * float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.7, 0.7))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.02 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.75  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.1  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.35  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 1  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                + left_hip_yaw_penalty
+                + right_hip_yaw_penalty
+                + hip_yaw_cross_penalty
+                + position_penalty
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=150, kd=40
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_007/Walk.py

@@ -0,0 +1,823 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -20.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(4.0) else 0.0
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  0.8 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.4, 0.4))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.02 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.75  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.1  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.35  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 1  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                head_toward_bonus +
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                + left_hip_yaw_penalty
+                + right_hip_yaw_penalty
+                + hip_yaw_cross_penalty
+                + position_penalty
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=150, kd=40
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_008/Walk.py

@@ -0,0 +1,849 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -20.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -3 * float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(4.0) else 0.0
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  0.8 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.4, 0.4))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.05 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.15  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.05  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.5  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                head_toward_bonus +
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                + left_hip_yaw_penalty
+                + right_hip_yaw_penalty
+                + hip_yaw_cross_penalty
+                + position_penalty
+                # + linkage_reward
+                # + waist_only_turn_penalty
+                # + yaw_link_reward
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"linkage_reward:{linkage_reward:.4f},"
+                # f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f},"
+                # f"yaw_link_reward:{yaw_link_reward:.4f}"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        action[10] = 0
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=80, kd=10
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_009/Walk.py

@@ -0,0 +1,853 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -20.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -3 * float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(4.0) else 0.0
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  0.8 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.4, 0.4))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.05 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_hip_pitch = float(joint_pos[11])
+        right_hip_pitch = float(joint_pos[17])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.15  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.05  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.3  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                head_toward_bonus +
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                + left_hip_yaw_penalty
+                + right_hip_yaw_penalty
+                + hip_yaw_cross_penalty
+                + position_penalty
+                # + linkage_reward
+                # + waist_only_turn_penalty
+                # + yaw_link_reward
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"linkage_reward:{linkage_reward:.4f},"
+                # f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f},"
+                # f"yaw_link_reward:{yaw_link_reward:.4f}"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        action[10] = 0
+        action[11] = np.clip(action[11], -0.3, 0.3)
+        action[17] = np.clip(action[17], -0.3, 0.3)
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=80, kd=10
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_010/Walk.py

@@ -0,0 +1,853 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -20.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -3 * float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(4.0) else 0.0
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  0.8 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.4, 0.4))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.05 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_hip_pitch = float(joint_pos[11])
+        right_hip_pitch = float(joint_pos[17])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.15  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.05  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.3  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                head_toward_bonus +
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                + left_hip_yaw_penalty
+                + right_hip_yaw_penalty
+                + hip_yaw_cross_penalty
+                + position_penalty
+                # + linkage_reward
+                # + waist_only_turn_penalty
+                # + yaw_link_reward
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"linkage_reward:{linkage_reward:.4f},"
+                # f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f},"
+                # f"yaw_link_reward:{yaw_link_reward:.4f}"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        action[10] = 0
+        action[11] = np.clip(action[11], -0.1, 0.1)
+        action[17] = np.clip(action[17], -0.1, 0.1)
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=110, kd=29.5
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_011/Walk.py

@@ -0,0 +1,853 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -20.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -3 * float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(4.0) else 0.0
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  0.8 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.4, 0.4))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.05 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_hip_pitch = float(joint_pos[11])
+        right_hip_pitch = float(joint_pos[17])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.15  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.05  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.3  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                head_toward_bonus +
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                + left_hip_yaw_penalty
+                + right_hip_yaw_penalty
+                + hip_yaw_cross_penalty
+                + position_penalty
+                # + linkage_reward
+                # + waist_only_turn_penalty
+                # + yaw_link_reward
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"linkage_reward:{linkage_reward:.4f},"
+                # f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f},"
+                # f"yaw_link_reward:{yaw_link_reward:.4f}"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        action[10] = 0
+        action[11] = np.clip(action[11], -0.1, 0.1)
+        action[17] = np.clip(action[17], -0.1, 0.1)
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=110, kd=6
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_012/Walk.py

@@ -0,0 +1,853 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "90"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -20.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -3 * float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(4.0) else 0.0
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  0.8 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.4, 0.4))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.05 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_hip_pitch = float(joint_pos[11])
+        right_hip_pitch = float(joint_pos[17])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.2  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.05  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.4  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                head_toward_bonus +
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                # + left_hip_yaw_penalty
+                # + right_hip_yaw_penalty
+                # + hip_yaw_cross_penalty
+                + position_penalty
+                # + linkage_reward
+                # + waist_only_turn_penalty
+                # + yaw_link_reward
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"linkage_reward:{linkage_reward:.4f},"
+                # f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f},"
+                # f"yaw_link_reward:{yaw_link_reward:.4f}"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+        # print(f"abs_yaw_error:{abs_yaw_error:.4f}")
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        action[10] = 0
+        action[11] = np.clip(action[11], -0.4, 0.4)
+        action[17] = np.clip(action[17], -0.4, 0.4)
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=110, kd=6
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_013/Walk.py

@@ -0,0 +1,853 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "90"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -20.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -3 * float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(4.0) else 0.0
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  0.8 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.4, 0.4))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.05 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_hip_pitch = float(joint_pos[11])
+        right_hip_pitch = float(joint_pos[17])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.2  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.05  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.4  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                head_toward_bonus +
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                # + left_hip_yaw_penalty
+                # + right_hip_yaw_penalty
+                # + hip_yaw_cross_penalty
+                + position_penalty
+                # + linkage_reward
+                # + waist_only_turn_penalty
+                # + yaw_link_reward
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"linkage_reward:{linkage_reward:.4f},"
+                # f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f},"
+                # f"yaw_link_reward:{yaw_link_reward:.4f}"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+        # print(f"abs_yaw_error:{abs_yaw_error:.4f}")
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        action[10] = 0
+        action[11] = np.clip(action[11], -0.4, 0.4)
+        action[17] = np.clip(action[17], -0.4, 0.4)
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=80, kd=4.67
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_R0_014/Walk.py

@@ -0,0 +1,853 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "120"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -20.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -3 * float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(4.0) else 0.0
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =   progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -1, 1))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.05 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_hip_pitch = float(joint_pos[11])
+        right_hip_pitch = float(joint_pos[17])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.2  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.05  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.5  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                head_toward_bonus +
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                + left_hip_yaw_penalty
+                + right_hip_yaw_penalty
+                + hip_yaw_cross_penalty
+                + position_penalty
+                # + linkage_reward
+                # + waist_only_turn_penalty
+                # + yaw_link_reward
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"linkage_reward:{linkage_reward:.4f},"
+                # f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f},"
+                # f"yaw_link_reward:{yaw_link_reward:.4f}"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+        # print(f"abs_yaw_error:{abs_yaw_error:.4f}")
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        action[10] = 0
+        action[11] = np.clip(action[11], -0.7, 0.7)
+        action[17] = np.clip(action[17], -0.7, 0.7)
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=80, kd=4.67
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_around_normal_60deg/Walk.py

@@ -0,0 +1,853 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "90"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -20.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -3 * float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(4.0) else 0.0
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  0.8 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.4, 0.4))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.05 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_hip_pitch = float(joint_pos[11])
+        right_hip_pitch = float(joint_pos[17])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.2  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.05  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.4  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                head_toward_bonus +
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                # + left_hip_yaw_penalty
+                # + right_hip_yaw_penalty
+                # + hip_yaw_cross_penalty
+                + position_penalty
+                # + linkage_reward
+                # + waist_only_turn_penalty
+                # + yaw_link_reward
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"linkage_reward:{linkage_reward:.4f},"
+                # f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f},"
+                # f"yaw_link_reward:{yaw_link_reward:.4f}"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+        # print(f"abs_yaw_error:{abs_yaw_error:.4f}")
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        action[10] = 0
+        action[11] = np.clip(action[11], -0.4, 0.4)
+        action[17] = np.clip(action[17], -0.4, 0.4)
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=80, kd=4.67
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Turn_around_unnormal/Walk.py

@@ -0,0 +1,853 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = float(os.environ.get("GYM_CPU_REWARD_DEBUG_INTERVAL_SEC", "600"))
+        self.reward_debug_burst_steps = int(os.environ.get("GYM_CPU_REWARD_DEBUG_BURST_STEPS", "10"))
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.3
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = float(os.environ.get("GYM_CPU_RESET_BEAM_YAW_RANGE_DEG", "180"))
+        self.reset_target_bearing_range_deg = float(os.environ.get("GYM_CPU_RESET_TARGET_BEARING_RANGE_DEG", "45"))
+        self.reset_target_distance_min = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MIN", "1.2"))
+        self.reset_target_distance_max = float(os.environ.get("GYM_CPU_RESET_TARGET_DISTANCE_MAX", "2.8"))
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_SCALE", "0.06"))
+        self.reward_smoothness_cap = float(os.environ.get("GYM_CPU_REWARD_SMOOTHNESS_CAP", "0.45"))
+        self.reward_head_toward_bonus = float(os.environ.get("GYM_CPU_REWARD_HEAD_TOWARD_BONUS", "1"))
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Randomize global target bearing so policy must learn to rotate toward it first.
+        heading_deg = float(r.global_orientation_euler[2])
+        target_offset = MathOps.rotate_2d_vec(
+            np.array([target_distance, 0.0]),
+            heading_deg + target_bearing_deg,
+            is_rad=False,
+        )
+        point1 = self.initial_position + target_offset
+        self.point_list = [point1]
+        self.target_position = self.point_list[self.waypoint_index]
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+
+        joint_pos_rad = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        joint_speed_rad = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        ang_vel = np.deg2rad(robot.gyroscope)
+        rp_ang_vel_mag = float(np.linalg.norm(ang_vel[:2]))
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            # remain = max(0, 800 - self.step_counter)
+            # return -8.0 - 0.01 * remain
+            return -20.0
+        
+
+        if np.linalg.norm(current_pos - previous_pos) > 0.005:
+            position_penalty = -3 * float(np.linalg.norm(current_pos - previous_pos))
+        else:
+            position_penalty = 0.0
+
+
+        # Turn-to-target shaping.
+        to_target = self.target_position - current_pos
+        dist_to_target = float(np.linalg.norm(to_target))
+        if dist_to_target > 1e-6:
+            target_yaw = math.atan2(float(to_target[1]), float(to_target[0]))
+        else:
+            target_yaw = 0.0
+
+        robot_yaw = math.radians(float(robot.global_orientation_euler[2]))
+        yaw_error = target_yaw - robot_yaw
+
+        # Main heading objective: face the target direction.
+        # heading_align_reward = 1.0 * math.cos(yaw_error)
+
+        abs_yaw_error = abs(yaw_error)
+        alive_bonus = 2.0 * max(0.0, 1.0 - abs_yaw_error / math.pi)
+        head_toward_bonus = self.reward_head_toward_bonus if abs_yaw_error < math.radians(4.0) else 0.0
+
+        if self.last_yaw_error is None:
+            heading_progress_reward = 0.0
+        else:
+            prev_abs_yaw_error = abs(self.last_yaw_error)
+            yaw_err_delta = prev_abs_yaw_error - abs_yaw_error
+            progress_gate = 1.0 if abs_yaw_error > math.radians(4.0) else 0.0
+            heading_progress_reward =  0.8 * progress_gate * yaw_err_delta
+            heading_progress_reward = float(np.clip(heading_progress_reward, -0.4, 0.4))
+        self.last_yaw_error = yaw_error
+
+        # action_penalty = -0.01 * float(np.linalg.norm(action))
+        smoothness_penalty = -0.05 * float(np.linalg.norm(action - self.last_action_for_reward))
+
+        posture_penalty = -0.6 * tilt_mag
+        # Penalize roll/pitch rotational shake but do not penalize yaw turning directly.
+        ang_vel_penalty = -0.06 * rp_ang_vel_mag
+
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+        left_hip_pitch = float(joint_pos[11])
+        right_hip_pitch = float(joint_pos[17])
+
+        left_ankle_roll = float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+
+        max_leg_roll = 0.15  # 防止劈叉姿势
+        split_penalty = -0.8 * max(0.0, (-left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.05  # 最小腿间距（防止贴得太近）
+        # 惩罚腿过分靠拢（内收）- 基于两腿间距
+        leg_separation = -left_hip_roll + right_hip_roll
+        inward_penalty = -0.25 * max(0.0, (min_leg_separation - leg_separation) / min_leg_separation)
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.5 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.3 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.3  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.4 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.4 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+        # 智能交叉腿惩罚：只在站立时惩罚，转身时允许交叉腿
+        yaw_rate = float(np.deg2rad(robot.gyroscope[2]))
+        yaw_rate_abs = abs(yaw_rate)
+
+        # 当转身速度较小时才惩罚交叉腿（站立状态）
+        cross_leg_gate = max(0.0, 1.0 - yaw_rate_abs / math.radians(8.0))
+        hip_yaw_cross_penalty = -1.0 * cross_leg_gate * max(0.0, -(left_hip_yaw * right_hip_yaw)) if left_hip_yaw > 0 and right_hip_yaw < 0 else 0.0
+
+        # Torso-lower-body linkage: reward coordinated turning, punish waist-only spinning.
+        waist_speed = abs(float(joint_speed_rad[10]))
+        lower_body_speed = float(np.mean(np.abs(joint_speed_rad[11:23])))
+        lower_body_follow_ratio = lower_body_speed / (waist_speed + 1e-4)
+        linkage_reward = 0.24 * min(1.0, lower_body_follow_ratio) * min(1.0, waist_speed / 1.2)
+        waist_only_turn_penalty = -0.20 * max(0.0, waist_speed - 1.35 * lower_body_speed)
+
+        # Extra posture linkage in yaw joints to avoid decoupled torso twist.
+        waist_yaw = abs(float(joint_pos_rad[10]))
+        hip_yaw_mean = 0.5 * (abs(float(joint_pos_rad[13])) + abs(float(joint_pos_rad[19])))
+        yaw_link_reward = 0.12 * math.exp(-abs(waist_yaw - hip_yaw_mean) / 0.22)
+
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+        
+        height_penalty = -(math.exp(12*abs(height_error))-1) if height_error > 0.04 else 0
+
+        # # 在 compute_reward 开头附近，添加高度变化率计算
+        # if not hasattr(self, 'last_height'):
+        #     self.last_height = height
+        #     self.last_height_time = self.step_counter  # 可选，用于时间间隔
+        # height_rate = height - self.last_height  # 正为上升，负为下降
+        # self.last_height = height
+
+        # 惩罚高度下降（负变化率）
+        # height_down_penalty = -5.0 * max(0, -height_rate)  # 系数可调，-height_rate 为正表示下降幅度
+
+        # # 在 compute_reward 中
+        # if self.step_counter > 50:
+        #     avg_prev_action = np.mean(self.prev_action_history, axis=0)
+        #     novelty = float(np.linalg.norm(action - avg_prev_action))
+        #     exploration_bonus = 0.05 * novelty
+        # else:
+        #     exploration_bonus = 0
+
+        # self.prev_action_history[self.history_idx] = action
+        # self.history_idx = (self.history_idx + 1) % 50
+
+
+        total = (
+            # progress_reward  + 
+                alive_bonus + 
+                head_toward_bonus +
+                heading_progress_reward +
+                # lateral_penalty +
+                # action_penalty +
+                smoothness_penalty +
+                posture_penalty
+                + ang_vel_penalty
+                + height_penalty
+                + ankle_roll_penalty
+                + ankle_roll_cross_penalty
+                + split_penalty
+                + inward_penalty
+                # + leg_proximity_penalty
+                + left_hip_yaw_penalty
+                + right_hip_yaw_penalty
+                + hip_yaw_cross_penalty
+                + position_penalty
+                # + linkage_reward
+                # + waist_only_turn_penalty
+                # + yaw_link_reward
+                # + stance_collapse_penalty
+                # + hip_yaw_yaw_cross_penalty
+                # + stance_collapse_penalty
+                #  + cross_leg_penalty
+                #  + exploration_bonus
+                # + height_down_penalty
+                 )
+        # print(height_error, height_penalty)
+        
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"height_penalty:{height_penalty:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                f"heading_progress_reward:{heading_progress_reward:.4f},"
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"cross_leg_penalty:{cross_leg_penalty:.4f},"
+                f"ang_vel_penalty:{ang_vel_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"hip_yaw_cross_penalty:{hip_yaw_cross_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"position_penalty:{position_penalty:.4f},"
+                # f"linkage_reward:{linkage_reward:.4f},"
+                # f"waist_only_turn_penalty:{waist_only_turn_penalty:.4f},"
+                # f"yaw_link_reward:{yaw_link_reward:.4f}"
+                # f"leg_proximity_penalty:{leg_proximity_penalty:.4f},"
+
+                # f"stance_collapse_penalty:{stance_collapse_penalty:.4f},"
+                # f"hip_yaw_yaw_cross_penalty:{hip_yaw_yaw_cross_penalty:.4f},"
+                #   f"height_down_penalty:{height_down_penalty:.4f}",
+                #   f"exploration_bonus:{exploration_bonus:.4f}"
+                f"alive_bonus:{alive_bonus:.4f},"
+                f"abs_yaw_error:{abs_yaw_error:.4f}"
+                f"total:{total:.4f}"
+                )
+
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        action[0:2] = 0
+        action[3] = 4
+        action[7] = -4
+        action[2] = 0
+        action[6] = 0
+        action[4] = 0
+        action[5] = -5
+        action[8] = 0
+        action[9] = 5
+        action[10] = 0
+        action[11] = np.clip(action[11], -0.1, 0.1)
+        action[17] = np.clip(action[17], -0.1, 0.1)
+        # action[12] = -1.0
+        # action[18] = 1.0
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=110, kd=29.5
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        if self.step_counter % 10 == 0:
+            self.previous_pos = current_pos.copy()
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+
+        self.last_action_for_reward = action.copy()
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = int(os.environ.get("GYM_CPU_N_ENVS", "20"))
+        if n_envs < 1:
+            raise ValueError("GYM_CPU_N_ENVS must be >= 1")
+        server_warmup_sec = float(os.environ.get("GYM_CPU_SERVER_WARMUP_SEC", "3.0"))
+        n_steps_per_env = int(os.environ.get("GYM_CPU_TRAIN_STEPS_PER_ENV", "512"))  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = int(os.environ.get("GYM_CPU_TRAIN_BATCH_SIZE", "512"))  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = float(os.environ.get("GYM_CPU_TRAIN_LR", "3e-4"))
+        folder_name = f'Turn_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        server_log_dir = os.path.join(model_path, "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+        # Wait for servers to start
+        print(f"Starting {n_envs + 1} rcssservermj servers...")
+        if server_warmup_sec > 0:
+            print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+            sleep(server_warmup_sec)
+        print("Servers started, creating environments...")
+
+        env = SubprocVecEnv([init_env(i, monitor=True) for i in range(n_envs)], start_method="spawn")
+        # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+        eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+        try:
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=float(os.environ.get("GYM_CPU_TRAIN_ENT_COEF", "0.05")),  # Entropy coefficient for exploration
+                    clip_range=float(os.environ.get("GYM_CPU_TRAIN_CLIP_RANGE", "0.2")),  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=float(os.environ.get("GYM_CPU_TRAIN_GAMMA", "0.95")), # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=int(os.environ.get("GYM_CPU_TRAIN_EPOCHS", "5")),
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * 20, save_freq=n_steps_per_env * 20, eval_eps=7,
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            servers.kill()
+            return
+
+        env.close()
+        eval_env.close()
+        servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = os.environ.get("GYM_CPU_TEST_NO_RENDER", "0") == "1"
+        test_no_realtime = os.environ.get("GYM_CPU_TEST_NO_REALTIME", "0") == "1"
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

scripts/gyms/logs/Walk_version_0.10/Walk.py

@@ -0,0 +1,968 @@
+import os
+import numpy as np
+import math
+import time
+from time import sleep
+from random import random
+from random import uniform
+from itertools import count
+
+from stable_baselines3 import PPO, TD3, DDPG, SAC, A2C
+from stable_baselines3.common.monitor import Monitor
+from stable_baselines3.common.vec_env import SubprocVecEnv, DummyVecEnv
+
+import gymnasium as gym
+from gymnasium import spaces
+
+from scripts.commons.Train_Base import Train_Base
+from scripts.commons.Server import Server as Train_Server
+
+from agent.base_agent import Base_Agent
+from utils.math_ops import MathOps
+
+from scipy.spatial.transform import Rotation as R
+
+'''
+Objective:
+Learn how to run forward using step primitive
+----------
+- class Basic_Run: implements an OpenAI custom gym
+- class Train:  implements algorithms to train a new model or test an existing model
+'''
+
+
+class WalkEnv(gym.Env):
+    def __init__(self, ip, server_p) -> None:
+
+        # Args: Server IP, Agent Port, Monitor Port, Uniform No., Robot Type, Team Name, Enable Log, Enable Draw
+        self.Player = player = Base_Agent(
+            team_name="Gym",
+            number=1,
+            host=ip,
+            port=server_p
+        )
+        self.robot_type = self.Player.robot
+        self.step_counter = 0  # to limit episode size
+        self.force_play_on = True
+
+        self.target_position = np.array([0.0, 0.0])  # target position in the x-y plane
+        self.initial_position = np.array([0.0, 0.0])  # initial position in the x-y plane
+        self.target_direction = 0.0  # target direction in the x-y plane (relative to the robot's orientation)
+        self.isfallen = False
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.debug_every_n_steps = 5
+        self.enable_debug_joint_status = False
+        self.reward_debug_interval_sec = 600.0
+        self.reward_debug_burst_steps = 10
+        self._reward_debug_last_time = time.time()
+        self._reward_debug_steps_left = 0
+        self.calibrate_nominal_from_neutral = True
+        self.auto_calibrate_train_sim_flip = True
+        self.nominal_calibrated_once = False
+        self.flip_calibrated_once = False
+        self._target_hz = 0.0
+        self._target_dt = 0.0
+        self._last_sync_time = None
+        self._speed_estimate = 0.0
+        self._speed_from_acc = 0.0
+        self._prev_accelerometer = np.zeros(3, dtype=np.float32)
+        self._speed_smoothing = 0.85
+        self._fallback_dt = 0.02
+        target_hz_env = 0
+        if target_hz_env:
+            try:
+                self._target_hz = float(target_hz_env)
+            except ValueError:
+                self._target_hz = 0.0
+        if self._target_hz > 0.0:
+            self._target_dt = 1.0 / self._target_hz
+
+        # State space
+        # 原始观测大小: 78
+        obs_size = 78
+        self.obs = np.zeros(obs_size, np.float32)
+        self.observation_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(obs_size,),
+            dtype=np.float32
+        )
+
+        action_dim = len(self.Player.robot.ROBOT_MOTORS)
+        self.no_of_actions = action_dim
+        self.action_space = spaces.Box(
+            low=-10.0,
+            high=10.0,
+            shape=(action_dim,),
+            dtype=np.float32
+        )
+
+        # 中立姿态
+        self.joint_nominal_position = np.array(
+            [
+                0.0,  # 0: Head_yaw (he1)
+                0.0,  # 1: Head_pitch (he2)
+                0.0,  # 2: Left_Shoulder_Pitch (lae1)
+                0.0,  # 3: Left_Shoulder_Roll (lae2)
+                0.0,  # 4: Left_Elbow_Pitch (lae3)
+                0.0,  # 5: Left_Elbow_Yaw (lae4)
+                0.0,  # 6: Right_Shoulder_Pitch (rae1)
+                0.0,  # 7: Right_Shoulder_Roll (rae2)
+                0.0,  # 8: Right_Elbow_Pitch (rae3)
+                0.0,  # 9: Right_Elbow_Yaw (rae4)
+                0.0,  # 10: Waist (te1)
+                0.0,  # 11: Left_Hip_Pitch (lle1)
+                0.0,  # 12: Left_Hip_Roll (lle2)
+                1.0,  # 13: Left_Hip_Yaw (lle3)
+                0.0,  # 14: Left_Knee_Pitch (lle4)
+                0.0,  # 15: Left_Ankle_Pitch (lle5)
+                0.0,  # 16: Left_Ankle_Roll (lle6)
+                0.0,  # 17: Right_Hip_Pitch (rle1)
+                0.0,  # 18: Right_Hip_Roll (rle2)
+                1.0,  # 19: Right_Hip_Yaw (rle3)
+                0.0,  # 20: Right_Knee_Pitch (rle4)
+                0.0,  # 21: Right_Ankle_Pitch (rle5)
+                0.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+        # self.joint_nominal_position = np.zeros(self.no_of_actions)
+        self.train_sim_flip = np.array(
+            [
+                1.0,  # 0: Head_yaw (he1)
+                -1.0,  # 1: Head_pitch (he2)
+                1.0,  # 2: Left_Shoulder_Pitch (lae1)
+                -1.0,  # 3: Left_Shoulder_Roll (lae2)
+                -1.0,  # 4: Left_Elbow_Pitch (lae3)
+                1.0,  # 5: Left_Elbow_Yaw (lae4)
+                -1.0,  # 6: Right_Shoulder_Pitch (rae1)
+                -1.0,  # 7: Right_Shoulder_Roll (rae2)
+                1.0,  # 8: Right_Elbow_Pitch (rae3)
+                1.0,  # 9: Right_Elbow_Yaw (rae4)
+                1.0,  # 10: Waist (te1)
+                1.0,  # 11: Left_Hip_Pitch (lle1)
+                -1.0,  # 12: Left_Hip_Roll (lle2)
+                -1.0,  # 13: Left_Hip_Yaw (lle3)
+                1.0,  # 14: Left_Knee_Pitch (lle4)
+                1.0,  # 15: Left_Ankle_Pitch (lle5)
+                -1.0,  # 16: Left_Ankle_Roll (lle6)
+                -1.0,  # 17: Right_Hip_Pitch (rle1)
+                -1.0,  # 18: Right_Hip_Roll (rle2)
+                -1.0,  # 19: Right_Hip_Yaw (rle3)
+                -1.0,  # 20: Right_Knee_Pitch (rle4)
+                -1.0,  # 21: Right_Ankle_Pitch (rle5)
+                -1.0,  # 22: Right_Ankle_Roll (rle6)
+            ]
+        )
+
+        self.scaling_factor = 0.5
+        # self.scaling_factor = 1
+
+        # Encourage a minimum lateral stance so the policy avoids feet overlap.
+        self.min_stance_rad = 0.10
+
+        # Small reset perturbations for robustness training.
+        self.enable_reset_perturb = False
+        self.reset_beam_yaw_range_deg = 180.0
+        self.reset_target_bearing_range_deg = 0.0
+        self.reset_target_distance_min = 5
+        self.reset_target_distance_max = 10
+        if self.reset_target_distance_min > self.reset_target_distance_max:
+            self.reset_target_distance_min, self.reset_target_distance_max = (
+                self.reset_target_distance_max,
+                self.reset_target_distance_min,
+            )
+        self.reset_joint_noise_rad = 0.025
+        self.reset_perturb_steps = 4
+        self.reset_recover_steps = 8
+
+        self.reward_smoothness_scale = 0.03
+        self.reward_smoothness_cap = 0.45
+        self.reward_forward_stability_gate = 0.35
+        self.reward_forward_tilt_hard_threshold = 0.50
+        self.reward_forward_tilt_hard_scale = 0.20
+        self.reward_head_toward_bonus = 1.0
+        self.turn_stationary_radius = 0.2
+        self.turn_stationary_penalty_scale = 3.0
+        self.stationary_start_steps = 20
+        self.stationary_step_eps = 0.015
+        self.stationary_penalty_scale = 1.2
+        self.train_stage = "walk"
+        self.in_place_radius = 0.18
+        self.in_place_center_reward_scale = 0.60
+        self.in_place_drift_penalty_scale = 1.20
+        self.waypoint_reach_distance = 0.3
+        self.num_waypoints = 1
+        self.exploration_start_steps = 40
+        self.exploration_scale = 0.012
+        self.exploration_cap = 0.2
+        self.exploration_target_novelty = 1.0
+        self.exploration_sigma = 0.7
+        self.reward_stride_swing_scale = 0.20
+        self.reward_stride_phase_scale = 0.18
+        self.reward_knee_drive_scale = 0.10
+        self.reward_knee_lift_scale = 0.12
+        self.reward_knee_lift_target = 0.15
+        self.reward_knee_lift_shortfall_scale = 0.05
+        self.reward_knee_overbend_threshold = 0.60
+        self.reward_knee_overbend_scale = 0.35
+        self.reward_hip_lift_scale = 0.12
+        self.reward_hip_lift_target = 0.80
+        self.reward_knee_alternate_scale = 0.10
+        self.reward_knee_bilateral_scale = 0.16
+        self.reward_single_leg_penalty_scale = 0.22
+        self.reward_knee_phase_switch_scale = 0.14
+        self.knee_phase_deadband = 0.10
+        self.knee_phase_min_interval = 18
+        self.knee_phase_target_interval = 22
+        self.knee_phase_fast_switch_penalty_scale = 0.10
+        self.knee_phase_max_hold_frames = 28
+        self.knee_phase_hold_penalty_scale = 0.18
+        self.reward_stride_cap = 0.80
+        self.reward_knee_explore_scale = 0.03
+        self.reward_knee_explore_delta_scale = 0.03
+        self.reward_knee_explore_cap = 0.10
+        self.reward_hip_pitch_explore_scale = 0.07
+        self.reward_hip_pitch_explore_delta_scale = 0.07
+        self.reward_hip_pitch_explore_cap = 0.10
+        self.reward_progress_scale = 18
+        self.reward_survival_scale = 0.5
+        self.reward_idle_penalty_scale = 0.6
+        self.reward_accel_penalty_scale = 0.08
+        self.reward_accel_penalty_cap = 0.40
+        self.reward_accel_abs_limit = 13.5
+        self.reward_accel_abs_penalty_scale = 0.05
+        self.reward_accel_abs_penalty_cap = 0.40
+        self.reward_heading_align_scale = 0.28
+        self.reward_heading_error_scale = 0.05
+
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.action_history_len = 50
+        self.prev_action_history = np.zeros((self.action_history_len, self.no_of_actions), dtype=np.float32)
+        self.history_idx = 0
+        self.previous_pos = np.array([0.0, 0.0])  # Track previous position
+        self.last_yaw_error = None
+        self.prev_knee_balance = 0.0
+        self.prev_knee_phase_sign = 0
+        self.knee_phase_frames_since_switch = 0
+        self.knee_phase_hold_frames = 0
+        self.Player.server.connect()
+        # sleep(2.0)  # Longer wait for connection to establish completely
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+        self.start_time = time.time()
+
+    def _reconnect_server(self):
+        try:
+            self.Player.server.shutdown()
+        except Exception:
+            pass
+
+        self.Player.server.connect()
+        self.Player.server.send_immediate(
+            f"(init {self.Player.robot.name} {self.Player.world.team_name} {self.Player.world.number})"
+        )
+
+    def _safe_receive_world_update(self, retries=1):
+        last_exc = None
+        for attempt in range(retries + 1):
+            try:
+                self.Player.server.receive()
+                self.Player.world.update()
+                return
+            except (ConnectionResetError, OSError) as exc:
+                last_exc = exc
+                if attempt >= retries:
+                    raise
+                self._reconnect_server()
+        if last_exc is not None:
+            raise last_exc
+
+    def debug_log(self, message):
+        print(message)
+        try:
+            log_path = os.path.join(os.path.dirname(os.path.dirname(__file__)), "comm_debug.log")
+            with open(log_path, "a", encoding="utf-8") as f:
+                f.write(message + "\n")
+        except OSError:
+            pass
+
+    @staticmethod
+    def _wrap_to_pi(angle_rad: float) -> float:
+        return (angle_rad + math.pi) % (2.0 * math.pi) - math.pi
+
+    def observe(self, init=False):
+
+        """获取当前观测值"""
+        robot = self.Player.robot
+        world = self.Player.world
+
+        # Safety check: ensure data is available
+
+        # 计算目标速度
+        raw_target = self.target_position - world.global_position[:2]
+        velocity = MathOps.rotate_2d_vec(
+            raw_target,
+            -robot.global_orientation_euler[2],
+            is_rad=False
+        )
+
+        # 计算相对方向
+        rel_orientation = MathOps.vector_angle(velocity) * 0.3
+        rel_orientation = np.clip(rel_orientation, -0.25, 0.25)
+
+        velocity = np.concatenate([velocity, np.array([rel_orientation])])
+        velocity[0] = np.clip(velocity[0], -0.5, 0.5)
+        velocity[1] = np.clip(velocity[1], -0.25, 0.25)
+
+        # 关节状态
+        radian_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        radian_joint_speeds = np.deg2rad(
+            [robot.motor_speeds[motor] for motor in robot.ROBOT_MOTORS]
+        )
+
+        qpos_qvel_previous_action = np.concatenate([
+            (radian_joint_positions * self.train_sim_flip - self.joint_nominal_position) / 4.6,
+            radian_joint_speeds / 110.0 * self.train_sim_flip,
+            self.previous_action / 10.0,
+        ])
+
+        # 角速度
+        ang_vel = np.clip(np.deg2rad(robot.gyroscope) / 50.0, -1.0, 1.0)
+
+        # 投影的重力方向
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+
+        # 组合观测
+        observation = np.concatenate([
+            qpos_qvel_previous_action,
+            ang_vel,
+            velocity,
+            projected_gravity,
+        ])
+
+        observation = np.clip(observation, -10.0, 10.0)
+        return observation.astype(np.float32)
+
+    def sync(self):
+        ''' Run a single simulation step '''
+        self._safe_receive_world_update(retries=1)
+        self.Player.robot.commit_motor_targets_pd()
+        self.Player.server.send()
+        if self._target_dt > 0.0:
+            now = time.time()
+            if self._last_sync_time is None:
+                self._last_sync_time = now
+                return
+            elapsed = now - self._last_sync_time
+            remaining = self._target_dt - elapsed
+            if remaining > 0.0:
+                time.sleep(remaining)
+                now = time.time()
+            self._last_sync_time = now
+
+    def debug_joint_status(self):
+        robot = self.Player.robot
+        actual_joint_positions = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        )
+        target_joint_positions = getattr(
+            self,
+            'target_joint_positions',
+            np.zeros(len(robot.ROBOT_MOTORS), dtype=np.float32)
+        )
+        joint_error = actual_joint_positions - target_joint_positions
+        leg_slice = slice(11, None)
+
+        self.debug_log(
+            "[WalkDebug] "
+            f"step={self.step_counter} "
+            f"pos={np.round(self.Player.world.global_position, 3).tolist()} "
+            f"target_xy={np.round(self.target_position, 3).tolist()} "
+            f"target_leg={np.round(target_joint_positions[leg_slice], 3).tolist()} "
+            f"actual_leg={np.round(actual_joint_positions[leg_slice], 3).tolist()} "
+            f"err_norm={float(np.linalg.norm(joint_error)):.4f} "
+            f"fallen={self.Player.world.global_position[2] < 0.3}"
+        )
+        print(f"waist target={target_joint_positions[10]:.3f}, actual={actual_joint_positions[10]:.3f}")
+
+    def reset(self, seed=None, options=None):
+        '''
+        Reset and stabilize the robot
+        Note: for some behaviors it would be better to reduce stabilization or add noise
+        '''
+        r = self.Player.robot
+        super().reset(seed=seed)
+        if seed is not None:
+            np.random.seed(seed)
+
+        target_distance = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+        target_bearing_deg = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+
+        self.step_counter = 0
+        self.waypoint_index = 0
+        self.route_completed = False
+        self.previous_action = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.last_action_for_reward = np.zeros(len(self.Player.robot.ROBOT_MOTORS))
+        self.prev_action_history.fill(0.0)
+        self.history_idx = 0
+        self.previous_pos = np.array([0.0, 0.0])  # Initialize for first step
+        self.last_yaw_error = None
+        self.prev_knee_balance = 0.0
+        self.prev_knee_phase_sign = 0
+        self.knee_phase_frames_since_switch = 0
+        self.knee_phase_hold_frames = 0
+        self.walk_cycle_step = 0
+        self._reward_debug_steps_left = 0
+        self._speed_estimate = 0.0
+        self._speed_from_acc = 0.0
+        self._prev_accelerometer = np.array(
+            getattr(self.Player.robot, "accelerometer", np.zeros(3)),
+            dtype=np.float32,
+        )
+
+        # 随机 beam 目标位置和朝向，增加训练多样性
+        beam_x = (random() - 0.5) * 10
+        beam_y = (random() - 0.5) * 10
+        beam_yaw = uniform(-self.reset_beam_yaw_range_deg, self.reset_beam_yaw_range_deg)
+
+        for _ in range(5):
+            self._safe_receive_world_update(retries=2)
+            self.Player.robot.commit_motor_targets_pd()
+            self.Player.server.commit_beam(pos2d=(beam_x, beam_y), rotation=beam_yaw)
+            self.Player.server.send()
+
+        # 执行 Neutral 技能直到完成，给机器人足够时间在 beam 位置稳定站立
+        finished_count = 0
+        for _ in range(50):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 20:  # 假设需要连续20次完成才算成功
+                    break
+
+        if self.enable_reset_perturb and self.reset_joint_noise_rad > 0.0:
+            perturb_action = np.zeros(self.no_of_actions, dtype=np.float32)
+            # Perturb waist + lower body only (10:), keep head/arms stable.
+            perturb_action[10:] = np.random.uniform(
+                -self.reset_joint_noise_rad,
+                self.reset_joint_noise_rad,
+                size=(self.no_of_actions - 10,)
+            )
+
+            for _ in range(self.reset_perturb_steps):
+                target_joint_positions = (self.joint_nominal_position + perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+            for i in range(self.reset_recover_steps):
+                # Linearly fade perturbation to help policy start from near-neutral.
+                alpha = 1.0 - float(i + 1) / float(self.reset_recover_steps)
+                target_joint_positions = (self.joint_nominal_position + alpha * perturb_action) * self.train_sim_flip
+                for idx, target in enumerate(target_joint_positions):
+                    r.set_motor_target_position(
+                        r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=25, kd=0.6
+                    )
+                self.sync()
+
+        # memory variables
+        self.sync()  
+        self.initial_position = np.array(self.Player.world.global_position[:2])
+        self.previous_pos = self.initial_position.copy()  # Critical: set to actual position
+        self.act = np.zeros(self.no_of_actions, np.float32)
+        # Generate multiple waypoints along a path
+        heading_deg = float(r.global_orientation_euler[2])
+        self.point_list = []
+        current_point = self.initial_position.copy()
+        
+        for i in range(self.num_waypoints):
+            # Each waypoint is placed further along the path
+            target_distance_wp = np.random.uniform(self.reset_target_distance_min, self.reset_target_distance_max)
+            self.target_distance_wp = target_distance_wp
+            target_bearing_deg_wp = np.random.uniform(-self.reset_target_bearing_range_deg, self.reset_target_bearing_range_deg)
+            
+            target_offset = MathOps.rotate_2d_vec(
+                np.array([target_distance_wp, 0.0]),
+                heading_deg + target_bearing_deg_wp,
+                is_rad=False,
+            )
+            next_point = current_point + target_offset
+            self.point_list.append(next_point)
+            current_point = next_point.copy()
+        
+        self.target_position = self.point_list[self.waypoint_index]
+        if self.train_stage == "in_place":
+            self.target_position = self.initial_position.copy()
+        self.initial_height = self.Player.world.global_position[2]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    
+    def compute_reward(self, previous_pos, current_pos, action):
+        height = float(self.Player.world.global_position[2])
+        robot = self.Player.robot
+
+        prev_dist_to_target = float(np.linalg.norm(self.target_position - previous_pos))
+        curr_dist_to_target = float(np.linalg.norm(self.target_position - current_pos))
+        dist_delta = prev_dist_to_target - curr_dist_to_target
+
+        is_fallen = height < 0.55
+        if is_fallen:
+            return -2.0
+        
+        joint_pos = np.deg2rad(
+            [robot.motor_positions[motor] for motor in robot.ROBOT_MOTORS]
+        ) * self.train_sim_flip
+        left_hip_roll = -float(joint_pos[12])
+        right_hip_roll = float(joint_pos[18])
+
+        left_ankle_roll = -float(joint_pos[16])
+        right_ankle_roll = float(joint_pos[22])
+        left_knee_flex = abs(float(joint_pos[14]))
+        right_knee_flex = abs(float(joint_pos[20]))
+        avg_knee_flex = 0.5 * (left_knee_flex + right_knee_flex)
+
+        max_leg_roll = 0.5  # 防止劈叉姿势
+        split_penalty = -0.12 * max(0.0, (left_hip_roll + right_hip_roll - 2 * max_leg_roll) / max_leg_roll)
+        left_hip_yaw = -float(joint_pos[13])
+        right_hip_yaw = float(joint_pos[19])
+
+        min_leg_separation = 0.04  # 最小腿间距（防止贴得太近）
+        inward_penalty = 0.3 * min(0.0, (left_hip_roll-min_leg_separation)) + 0.3 * min(0.0, (right_hip_roll-min_leg_separation))  # 惩罚左右腿过度内扣
+        
+        
+        # 脚踝roll角度检测：防止过度外翻或内翻
+        max_ankle_roll = 0.15  # 最大允许的脚踝roll角度
+        
+        # 惩罚脚踝过度外翻/内翻（绝对值过大）
+        ankle_roll_penalty = -0.12 * max(0.0, (abs(left_ankle_roll) + abs(right_ankle_roll) - 2 * max_ankle_roll) / max_ankle_roll)
+        
+        # 惩罚两脚踝roll方向相反（不稳定姿势）
+        ankle_roll_cross_penalty = -0.12 * max(0.0, -(left_ankle_roll * right_ankle_roll))
+
+       # 分别惩罚左右大腿过度转动
+        max_hip_yaw = 0.2  # 最大允许的yaw角度
+        left_hip_yaw_penalty = -0.6 * max(0.0, abs(left_hip_yaw) - max_hip_yaw)
+        right_hip_yaw_penalty = -0.6 * max(0.0, abs(right_hip_yaw) - max_hip_yaw)
+
+        target_vec = self.target_position - current_pos
+        target_dist = float(np.linalg.norm(target_vec))
+        if target_dist > 1e-6:
+            target_heading = math.atan2(float(target_vec[1]), float(target_vec[0]))
+            robot_heading = math.radians(float(robot.global_orientation_euler[2]))
+            heading_error = self._wrap_to_pi(target_heading - robot_heading)
+            heading_align_reward = self.reward_heading_align_scale * math.cos(heading_error)
+            heading_error_penalty = -self.reward_heading_error_scale * abs(heading_error)
+        else:
+            heading_align_reward = 0.0
+            heading_error_penalty = 0.0
+
+        # Forward-progress reward (distance delta) with anti-stuck shaping.
+        progress_reward = self.reward_progress_scale * dist_delta
+        survival_reward = self.reward_survival_scale
+        smoothness_penalty = -self.reward_smoothness_scale * float(np.linalg.norm(action - self.last_action_for_reward))
+        step_displacement = float(np.linalg.norm(current_pos - previous_pos))
+        accel_signal = 0.0
+        accel_source = "imu_delta"
+        accel_now = np.array(getattr(robot, "accelerometer", np.zeros(3)), dtype=np.float32)
+        if accel_now.shape[0] >= 3:
+            # Use IMU acceleration delta to reduce gravity bias and punish abrupt bursts.
+            accel_signal = float(np.linalg.norm(accel_now[:3] - self._prev_accelerometer[:3]))
+            self._prev_accelerometer = accel_now
+        accel_penalty = -min(
+            self.reward_accel_penalty_cap,
+            self.reward_accel_penalty_scale * accel_signal,
+        )
+        accel_abs = float(np.linalg.norm(accel_now[:3])) if accel_now.shape[0] >= 3 else 0.0
+        accel_abs_over = max(0.0, accel_abs - self.reward_accel_abs_limit)
+        accel_abs_penalty = -min(
+            self.reward_accel_abs_penalty_cap,
+            self.reward_accel_abs_penalty_scale * accel_abs_over,
+        )
+        if self.step_counter > 30 and step_displacement < 0.015 and self.target_distance_wp > 0.3:
+            idle_penalty = -self.reward_idle_penalty_scale
+        else:
+            idle_penalty = 0.0
+
+        if self.step_counter > self.exploration_start_steps:
+            displacement_novelty = step_displacement / max(1e-6, self.stationary_step_eps)
+            exploration_bonus = min(
+                self.exploration_cap,
+                self.exploration_scale * max(0.0, displacement_novelty - self.exploration_target_novelty),
+            )
+        else:
+            exploration_bonus = 0.0
+
+        # Encourage active/varied knee motions early in training without dominating progress reward.
+        left_knee_act = float(action[14])
+        right_knee_act = float(action[20])
+        left_knee_delta = abs(left_knee_act - float(self.last_action_for_reward[14])) 
+        right_knee_delta = abs(right_knee_act - float(self.last_action_for_reward[20])) 
+        knee_action_mag = 0.5 * (abs(left_knee_act) + abs(right_knee_act))
+        knee_action_delta = 0.5 * (left_knee_delta + right_knee_delta)
+        if self.step_counter > 10:
+            knee_explore_reward = min(
+                self.reward_knee_explore_cap,
+                self.reward_knee_explore_scale * knee_action_mag
+                + self.reward_knee_explore_delta_scale * knee_action_delta,
+            )
+        else:
+            knee_explore_reward = 0.0
+
+        # Directly encourage observable knee flexion instead of only action exploration.
+        knee_lift_shortfall_penalty = -self.reward_knee_lift_shortfall_scale * max(
+            0.0, self.reward_knee_lift_target - avg_knee_flex
+        )
+
+        # Encourage hip-pitch exploration to improve forward stride generation.
+        left_hip_pitch_act = float(action[11])
+        right_hip_pitch_act = float(action[17])
+        left_hip_pitch_delta = abs(left_hip_pitch_act - float(self.last_action_for_reward[11]))
+        right_hip_pitch_delta = abs(right_hip_pitch_act - float(self.last_action_for_reward[17]))
+        hip_pitch_action_mag = 0.5 * (abs(left_hip_pitch_act) + abs(right_hip_pitch_act))
+        hip_pitch_action_delta = 0.5 * (left_hip_pitch_delta + right_hip_pitch_delta)
+        if self.step_counter > 10:
+            hip_pitch_explore_reward = min(
+                self.reward_hip_pitch_explore_cap,
+                self.reward_hip_pitch_explore_scale * hip_pitch_action_mag
+                + self.reward_hip_pitch_explore_delta_scale * hip_pitch_action_delta,
+            )
+        else:
+            hip_pitch_explore_reward = 0.0
+
+        if curr_dist_to_target < 0.3:
+            arrival_bonus = self.target_distance_wp * 8 ## 奖励到达目标点
+        else:
+            arrival_bonus = 0.0
+            
+        target_height = self.initial_height
+        height_error =  height - target_height
+        height_error = height - target_height
+
+        height_penalty = -0.5 * (math.exp(15*abs(height_error))-1)
+
+        orientation_quat_inv = R.from_quat(robot._global_cheat_orientation).inv()
+        projected_gravity = orientation_quat_inv.apply(np.array([0.0, 0.0, -1.0]))
+        tilt_mag = float(np.linalg.norm(projected_gravity[:2]))
+        posture_penalty = -0.6 * (tilt_mag)
+        total = (
+            progress_reward
+            + survival_reward
+            + smoothness_penalty
+            + accel_penalty
+            + accel_abs_penalty
+            + idle_penalty
+            + split_penalty
+            + inward_penalty
+            + ankle_roll_penalty
+            + ankle_roll_cross_penalty
+            + left_hip_yaw_penalty
+            + right_hip_yaw_penalty
+            + heading_align_reward
+            + heading_error_penalty
+            # + exploration_bonus
+            # + knee_explore_reward
+            # + knee_lift_shortfall_penalty
+            # + hip_pitch_explore_reward
+            + arrival_bonus
+            + height_penalty
+            + posture_penalty
+        )
+
+        now = time.time()
+        if self.reward_debug_interval_sec > 0 and now - self._reward_debug_last_time >= self.reward_debug_interval_sec:
+            self._reward_debug_last_time = now
+            self._reward_debug_steps_left = max(1, self.reward_debug_burst_steps)
+
+        if self._reward_debug_steps_left > 0:
+            self._reward_debug_steps_left -= 1
+            self.debug_log(
+                f"progress_reward:{progress_reward:.4f},"
+                f"survival_reward:{survival_reward:.4f},"
+                f"smoothness_penalty:{smoothness_penalty:.4f},"
+                f"accel_penalty:{accel_penalty:.4f},"
+                f"accel_source:{accel_source},"
+                f"accel_signal:{accel_signal:.4f},"
+                f"accel_abs:{accel_abs:.4f},"
+                f"accel_abs_penalty:{accel_abs_penalty:.4f},"
+                f"idle_penalty:{idle_penalty:.4f},"
+                f"split_penalty:{split_penalty:.4f},"
+                f"inward_penalty:{inward_penalty:.4f},"
+                f"ankle_roll_penalty:{ankle_roll_penalty:.4f},"
+                f"ankle_roll_cross_penalty:{ankle_roll_cross_penalty:.4f},"
+                f"left_hip_yaw_penalty:{left_hip_yaw_penalty:.4f},"
+                f"right_hip_yaw_penalty:{right_hip_yaw_penalty:.4f},"
+                f"heading_align_reward:{heading_align_reward:.4f},"
+                f"heading_error_penalty:{heading_error_penalty:.4f},"
+                # f"exploration_bonus:{exploration_bonus:.4f},"
+                f"height_penalty:{height_penalty:.4f},"
+                # f"knee_explore_reward:{knee_explore_reward:.4f},"
+                f"posture_penalty:{posture_penalty:.4f},"
+                # f"knee_lift_shortfall_penalty:{knee_lift_shortfall_penalty:.4f},"
+                # f"hip_pitch_explore_reward:{hip_pitch_explore_reward:.4f},"
+                f"arrival_bonus:{arrival_bonus:.4f},"
+                f"total:{total:.4f}"
+            )
+        return total
+
+
+
+    def step(self, action):
+
+        r = self.Player.robot
+        max_action_delta =  0.5# Limit how much the action can change from the previous step to encourage smoother motions.
+        if self.previous_action is not None:
+            action = np.clip(action, self.previous_action - max_action_delta, self.previous_action + max_action_delta)
+        # Loosen upper-body constraints: keep motion bounded but no longer hard-lock head/arms/waist.
+        action[0:2] = 0
+        action[3] = np.clip(action[3], 3, 5)
+        action[7] = np.clip(action[7], -5, -3)
+        action[2] = np.clip(action[2], -6, 6)
+        action[6] = np.clip(action[6], -6, 6)
+        action[4] = 0
+        action[5] = np.clip(action[5], -8, -2)
+        action[8] = 0
+        action[9] = np.clip(action[9], 8, 2)
+        action[10] = np.clip(action[10], -0.6, 0.6)
+        # Boost knee command range so policy can produce visible knee flexion earlier.
+        action[14] = np.clip(action[14], 0, 10.0)
+        action[20] = np.clip(action[20], -10.0, 0)
+        # action[14] = 1 # the correct left knee sign
+        # action[20] = -1 # the correct right knee sign
+        # action[11] = 1
+        # action[17] = 1
+        # action[12] = -1
+        # action[18] = 1
+        # action[13] = -1.0
+        # action[19] = 1.0
+        self.previous_action = action.copy()
+
+        self.target_joint_positions = (
+                # self.joint_nominal_position +
+                 self.scaling_factor * action
+        )
+        self.target_joint_positions *= self.train_sim_flip
+
+        for idx, target in enumerate(self.target_joint_positions):
+            r.set_motor_target_position(
+                r.ROBOT_MOTORS[idx], target * 180 / math.pi, kp=60, kd=1.2
+            )
+
+        self.previous_action = action.copy()
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        if self.enable_debug_joint_status and self.step_counter % self.debug_every_n_steps == 0:
+            self.debug_joint_status()
+
+        current_pos = np.array(self.Player.world.global_position[:2], dtype=np.float32)
+
+        # Compute reward based on movement from previous step
+        reward = self.compute_reward(self.previous_pos, current_pos, action)
+        self.previous_pos = current_pos.copy()
+
+        self.prev_action_history[self.history_idx] = action.copy()
+        self.history_idx = (self.history_idx + 1) % self.action_history_len
+
+        self.last_action_for_reward = action.copy()
+        
+        # Check if current waypoint is reached
+        if self.train_stage != "in_place":
+            dist_to_waypoint = float(np.linalg.norm(current_pos - self.target_position))
+            if dist_to_waypoint < self.waypoint_reach_distance:
+                # Move to next waypoint
+                self.waypoint_index += 1
+                if self.waypoint_index >= len(self.point_list):
+                    # All waypoints completed
+                    self.route_completed = True
+                else:
+                    # Update target to next waypoint
+                    self.target_position = self.point_list[self.waypoint_index]
+
+        # Fall detection and penalty
+        is_fallen = self.Player.world.global_position[2] < 0.55
+
+        # terminal state: the robot is falling or timeout
+        terminated = is_fallen or self.step_counter > 800 or self.route_completed
+        truncated = False
+
+        return self.observe(), reward, terminated, truncated, {}
+
+
+class Train(Train_Base):
+    def __init__(self, script) -> None:
+        super().__init__(script)
+
+    def train(self, args):
+
+        # --------------------------------------- Learning parameters
+        n_envs = 20
+        server_warmup_sec = 3.0
+        n_steps_per_env = 256  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = 512  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 90000000
+        learning_rate = 2e-4
+        ent_coef = 0.035
+        clip_range = 0.2
+        gamma = 0.97
+        n_epochs = 3
+        enable_eval = True
+        monitor_train_env = False
+        eval_freq_mult = 60
+        save_freq_mult = 60
+        eval_eps = 7
+        folder_name = f'Walk_R{self.robot_type}'
+        model_path = f'./scripts/gyms/logs/{folder_name}/'
+
+        print(f"Model path: {model_path}")
+        print(f"Using {n_envs} parallel environments")
+
+        # --------------------------------------- Run algorithm
+        def init_env(i_env, monitor=False):
+            def thunk():
+                env = WalkEnv(self.ip, self.server_p + i_env)
+                if monitor:
+                    env = Monitor(env)
+                return env
+
+            return thunk
+
+        env = None
+        eval_env = None
+        servers = None
+        try:
+            server_log_dir = os.path.join(model_path, "server_logs")
+            os.makedirs(server_log_dir, exist_ok=True)
+            servers = Train_Server(self.server_p, self.monitor_p_1000, n_envs + 1, no_render=True, no_realtime=True)  # include 1 extra server for testing
+
+            # Wait for servers to start
+            print(f"Starting {n_envs + 1} rcssservermj servers...")
+            if server_warmup_sec > 0:
+                print(f"Waiting {server_warmup_sec:.1f}s for server warmup...")
+                sleep(server_warmup_sec)
+            print("Servers started, creating environments...")
+
+            env = SubprocVecEnv([init_env(i, monitor=monitor_train_env) for i in range(n_envs)], start_method="spawn")
+            # Use single-process eval env to avoid extra subprocess fragility during callback evaluation.
+            if enable_eval:
+                eval_env = DummyVecEnv([init_env(n_envs, monitor=True)])
+
+            # Custom policy network architecture
+            policy_kwargs = dict(
+                net_arch=dict(
+                    pi=[512, 256, 128],  # Policy network: 3 layers
+                    vf=[512, 256, 128]  # Value network: 3 layers
+                ),
+                activation_fn=__import__('torch.nn', fromlist=['ELU']).ELU,
+            )
+
+            if "model_file" in args:  # retrain
+                model = PPO.load(args["model_file"], env=env, device="cpu", n_envs=n_envs, n_steps=n_steps_per_env,
+                                 batch_size=minibatch_size, learning_rate=learning_rate)
+            else:  # train new model
+                model = PPO(
+                    "MlpPolicy",
+                    env=env,
+                    verbose=1,
+                    n_steps=n_steps_per_env,
+                    batch_size=minibatch_size,
+                    learning_rate=learning_rate,
+                    device="cpu",
+                    policy_kwargs=policy_kwargs,
+                    ent_coef=ent_coef,  # Entropy coefficient for exploration
+                    clip_range=clip_range,  # PPO clipping parameter
+                    gae_lambda=0.95,  # GAE lambda
+                    gamma=gamma, # Discount factor
+                    # target_kl=0.03,
+                    n_epochs=n_epochs,
+                    tensorboard_log=f"./scripts/gyms/logs/{folder_name}/tensorboard/"
+                )
+
+            model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env,
+                                          eval_freq=n_steps_per_env * max(1, eval_freq_mult),
+                                          save_freq=n_steps_per_env * max(1, save_freq_mult),
+                                          eval_eps=max(1, eval_eps),
+                                          backup_env_file=__file__)
+        except KeyboardInterrupt:
+            sleep(1)  # wait for child processes
+            print("\nctrl+c pressed, aborting...\n")
+            return
+        finally:
+            if env is not None:
+                env.close()
+            if eval_env is not None:
+                eval_env.close()
+            if servers is not None:
+                servers.kill()
+
+    def test(self, args):
+
+        # Uses different server and monitor ports
+        server_log_dir = os.path.join(args["folder_dir"], "server_logs")
+        os.makedirs(server_log_dir, exist_ok=True)
+        test_no_render = False
+        test_no_realtime = False
+
+        server = Train_Server(
+            self.server_p - 1,
+            self.monitor_p,
+            1,
+            no_render=test_no_render,
+            no_realtime=test_no_realtime,
+        )
+        env = WalkEnv(self.ip, self.server_p - 1)
+        model = PPO.load(args["model_file"], env=env)
+
+        try:
+            self.export_model(args["model_file"], args["model_file"] + ".pkl",
+                              False)  # Export to pkl to create custom behavior
+            self.test_model(model, env, log_path=args["folder_dir"], model_path=args["folder_dir"])
+        except KeyboardInterrupt:
+            print()
+
+        env.close()
+        server.kill()
+
+
+if __name__ == "__main__":
+    from types import SimpleNamespace
+
+    # 创建默认参数
+    script_args = SimpleNamespace(
+        args=SimpleNamespace(
+            i='127.0.0.1',  # Server IP
+            p=3100,  # Server port
+            m=3200,  # Monitor port
+            r=0,  # Robot type
+            t='Gym',  # Team name
+            u=1  # Uniform number
+        )
+    )
+
+    trainer = Train(script_args)
+
+    run_mode = os.environ.get("GYM_CPU_MODE", "train").strip().lower()
+
+    if run_mode == "test":
+        test_model_file = os.environ.get("GYM_CPU_TEST_MODEL", "scripts/gyms/logs/Turn_R0_004/best_model.zip")
+        test_folder = os.environ.get("GYM_CPU_TEST_FOLDER", "scripts/gyms/logs/Turn_R0_004/")
+        trainer.test({"model_file": test_model_file, "folder_dir": test_folder})
+    else:
+        retrain_model = os.environ.get("GYM_CPU_TRAIN_MODEL", "").strip()
+        if retrain_model:
+            trainer.train({"model_file": retrain_model})
+        else:
+            trainer.train({})

train.sh

@@ -24,36 +24,14 @@ CPU_QUOTA="$((CORES * UTIL_PERCENT))%"
 MEMORY_MAX="${MEMORY_MAX:-0}"
 
 # ------------------------------
-# 训练运行参数（由 scripts/gyms/Walk.py 读取）
+# 精简运行参数（由 scripts/gyms/Walk.py 读取）
 # ------------------------------
-# 运行模式：train 或 test
+# 仅保留最常用开关，避免超长环境变量命令。
 GYM_CPU_MODE="${GYM_CPU_MODE:-train}"
-
-# 并行环境数量：越大通常吞吐越高，但也更容易触发 OOM 或连接不稳定。
-# 默认使用更稳妥的 12，确认稳定后再升到 16/20。
-GYM_CPU_N_ENVS="${GYM_CPU_N_ENVS:-20}"
-# 服务器预热时间（秒）：
-# 在批量拉起 rcssserver 后等待一段时间，再创建 SubprocVecEnv，
-# 可降低 ConnectionReset/EOFError 概率。
-GYM_CPU_SERVER_WARMUP_SEC="${GYM_CPU_SERVER_WARMUP_SEC:-10}"
-
-# 训练专用参数
-GYM_CPU_TRAIN_STEPS_PER_ENV="${GYM_CPU_TRAIN_STEPS_PER_ENV:-256}"
-GYM_CPU_TRAIN_BATCH_SIZE="${GYM_CPU_TRAIN_BATCH_SIZE:-512}"
-GYM_CPU_TRAIN_LR="${GYM_CPU_TRAIN_LR:-1e-4}"
-GYM_CPU_TRAIN_ENT_COEF="${GYM_CPU_TRAIN_ENT_COEF:-0.03}"
-GYM_CPU_TRAIN_CLIP_RANGE="${GYM_CPU_TRAIN_CLIP_RANGE:-0.13}"
-GYM_CPU_TRAIN_GAMMA="${GYM_CPU_TRAIN_GAMMA:-0.95}"
-GYM_CPU_TRAIN_EPOCHS="${GYM_CPU_TRAIN_EPOCHS:-5}"
+GYM_CPU_TRAIN_STAGE="${GYM_CPU_TRAIN_STAGE:-walk}"
 GYM_CPU_TRAIN_MODEL="${GYM_CPU_TRAIN_MODEL:-}"
-
-# 测试专用参数
 GYM_CPU_TEST_MODEL="${GYM_CPU_TEST_MODEL:-scripts/gyms/logs/Walk_R0_004/best_model.zip}"
 GYM_CPU_TEST_FOLDER="${GYM_CPU_TEST_FOLDER:-scripts/gyms/logs/Walk_R0_004/}"
-# 测试默认实时且显示画面：默认均为 0
-# 设为 1 表示关闭对应能力
-GYM_CPU_TEST_NO_RENDER="${GYM_CPU_TEST_NO_RENDER:-0}"
-GYM_CPU_TEST_NO_REALTIME="${GYM_CPU_TEST_NO_REALTIME:-0}"
 
 # Python 解释器选择策略：
 # 1) 优先使用你手动传入的 PYTHON_BIN
@@ -93,7 +71,7 @@ SCRIPT_DIR="$(cd "$(dirname "$0")" && pwd)"
 # 打印当前生效配置，方便排障和复现实验。
 echo "Starting training with limits: CPU=${CPU_QUOTA}, Memory=${MEMORY_MAX}"
 echo "Mode: ${GYM_CPU_MODE}"
-echo "Runtime knobs: GYM_CPU_N_ENVS=${GYM_CPU_N_ENVS}, GYM_CPU_SERVER_WARMUP_SEC=${GYM_CPU_SERVER_WARMUP_SEC}"
+echo "Run knobs: GYM_CPU_MODE=${GYM_CPU_MODE}, GYM_CPU_TRAIN_STAGE=${GYM_CPU_TRAIN_STAGE}"
 echo "Using Python: ${PYTHON_EXEC}"
 if [[ -n "${CONDA_DEFAULT_ENV:-}" ]]; then
 	echo "Detected conda env: ${CONDA_DEFAULT_ENV}"
@@ -118,19 +96,9 @@ systemd-run --user --scope \
 	"${SYSTEMD_PROPS[@]}" \
 	env \
 	GYM_CPU_MODE="${GYM_CPU_MODE}" \
-	GYM_CPU_N_ENVS="${GYM_CPU_N_ENVS}" \
-	GYM_CPU_SERVER_WARMUP_SEC="${GYM_CPU_SERVER_WARMUP_SEC}" \
-	GYM_CPU_TRAIN_STEPS_PER_ENV="${GYM_CPU_TRAIN_STEPS_PER_ENV}" \
-	GYM_CPU_TRAIN_BATCH_SIZE="${GYM_CPU_TRAIN_BATCH_SIZE}" \
-	GYM_CPU_TRAIN_LR="${GYM_CPU_TRAIN_LR}" \
-	GYM_CPU_TRAIN_ENT_COEF="${GYM_CPU_TRAIN_ENT_COEF}" \
-	GYM_CPU_TRAIN_CLIP_RANGE="${GYM_CPU_TRAIN_CLIP_RANGE}" \
-	GYM_CPU_TRAIN_GAMMA="${GYM_CPU_TRAIN_GAMMA}" \
-	GYM_CPU_TRAIN_EPOCHS="${GYM_CPU_TRAIN_EPOCHS}" \
+	GYM_CPU_TRAIN_STAGE="${GYM_CPU_TRAIN_STAGE}" \
 	GYM_CPU_TRAIN_MODEL="${GYM_CPU_TRAIN_MODEL}" \
 	GYM_CPU_TEST_MODEL="${GYM_CPU_TEST_MODEL}" \
 	GYM_CPU_TEST_FOLDER="${GYM_CPU_TEST_FOLDER}" \
-	GYM_CPU_TEST_NO_RENDER="${GYM_CPU_TEST_NO_RENDER}" \
-	GYM_CPU_TEST_NO_REALTIME="${GYM_CPU_TEST_NO_REALTIME}" \
 	"${PYTHON_EXEC}" "-m" "scripts.gyms.Walk"
 

world/world.py

@@ -47,7 +47,7 @@ class World:
         self.their_team_players: list[OtherRobot] = [OtherRobot(is_teammate=False) for _ in
                                                      range(self.MAX_PLAYERS_PER_TEAM)]
         self.field: Field = self.__initialize_field(field_name=field_name)
-        self.WORLD_STEPTIME: float = 0.005  # Time step of the world in seconds
+        self.WORLD_STEPTIME: float = 0.02  # Time step of the world in seconds
 
     def update(self) -> None:
         """