import os import numpy as np import math import time from time import sleep from random import random from random import uniform from stable_baselines3 import PPO from stable_baselines3.common.vec_env import SubprocVecEnv 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.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 # 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.reset_joint_noise_rad = 0.015 self.reset_perturb_steps = 3 self.reset_recover_steps = 8 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.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 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 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.Player.server.receive() self.Player.world.update() 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) 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 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.walk_cycle_step = 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.Player.server.receive() self.Player.world.update() 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) # Build target in the robot's current forward direction instead of fixed global +x. 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.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]) orientation_quat_inv = R.from_quat(self.Player.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(self.Player.robot.gyroscope) ang_vel_mag = float(np.linalg.norm(ang_vel)) is_fallen = height < 0.3 if is_fallen: # remain = max(0, 800 - self.step_counter) # return -8.0 - 0.01 * remain return -1.0 # # 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 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 # + exploration_bonus # + height_down_penalty ) if time.time() - self.start_time >= 1200: 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"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 self.previous_action = action 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=1.0 ) self.previous_action = action 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] < 0.3 # 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 # Reduced from 8 to decrease CPU/network pressure during init if n_envs < 1: raise ValueError("GYM_CPU_N_ENVS must be >= 1") 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 = 30000000 learning_rate = 1e-4 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): def thunk(): return WalkEnv(self.ip, self.server_p + i_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) # include 1 extra server for testing # Wait for servers to start print(f"Starting {n_envs + 1} rcssservermj servers...") print("Servers started, creating environments...") env = SubprocVecEnv([init_env(i) for i in range(n_envs)]) eval_env = SubprocVecEnv([init_env(n_envs)]) 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=0.03, # Entropy coefficient for exploration clip_range=0.13, # PPO clipping parameter gae_lambda=0.95, # GAE lambda gamma=0.95 , # Discount factor target_kl=0.03, n_epochs=5 ) model_path = self.learn_model(model, total_steps, model_path, eval_env=eval_env, eval_freq=n_steps_per_env * 10, save_freq=n_steps_per_env * 10, 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) server = Train_Server(self.server_p - 1, self.monitor_p, 1) 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) trainer.train({"model_file": "scripts/gyms/logs/Walk_R0_004/best_model.zip"}) # trainer.test({"model_file": "scripts/gyms/logs/Walk_R0_004/best_model.zip", # "folder_dir": "scripts/gyms/logs/Walk_R0_004/",})