change model and policy

This reverts commit 648cf32e9c.

behaviors/custom/reinforcement/walk/walk.py

@@ -98,8 +98,12 @@ class Walk(Behavior):
         velocity[0] = np.clip(velocity[0], -0.5, 0.5)
         velocity[1] = np.clip(velocity[1], -0.25, 0.25)
 
-        radian_joint_positions = np.deg2rad(list(robot.motor_positions.values()))
-        radian_joint_speeds = np.deg2rad(list(robot.motor_speeds.values()))
+        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.vstack(
             (

communication/server.py

@@ -72,37 +72,37 @@ class Server:
         self.commit(msg)
         self.send()
 
-    def receive(self) -> None:
-        """
-        Receive the next message from the TCP/IP socket and updates world
-        """
+    def receive(self):
 
-        # Receive message length information
-        if (
-            self.__socket.recv_into(
-                self.__rcv_buffer, nbytes=4, flags=socket.MSG_WAITALL
+        while True:
+
+            if (
+                    self.__socket.recv_into(
+                        self.__rcv_buffer, nbytes=4, flags=socket.MSG_WAITALL
+                    ) != 4
+            ):
+                raise ConnectionResetError
+
+            msg_size = int.from_bytes(self.__rcv_buffer[:4], byteorder="big", signed=False)
+
+            if msg_size > self.__rcv_buffer_size:
+                self.__rcv_buffer_size = msg_size
+                self.__rcv_buffer = bytearray(self.__rcv_buffer_size)
+
+            if (
+                    self.__socket.recv_into(
+                        self.__rcv_buffer, nbytes=msg_size, flags=socket.MSG_WAITALL
+                    ) != msg_size
+            ):
+                raise ConnectionResetError
+
+            self.world_parser.parse(
+                message=self.__rcv_buffer[:msg_size].decode()
             )
-            != 4
-        ):
-            raise ConnectionResetError
 
-        msg_size = int.from_bytes(self.__rcv_buffer[:4], byteorder="big", signed=False)
-
-        # Ensure receive buffer is large enough to hold the message
-        if msg_size > self.__rcv_buffer_size:
-            self.__rcv_buffer_size = msg_size
-            self.__rcv_buffer = bytearray(self.__rcv_buffer_size)
-
-        # Receive message with the specified length
-        if (
-            self.__socket.recv_into(
-                self.__rcv_buffer, nbytes=msg_size, flags=socket.MSG_WAITALL
-            )
-            != msg_size
-        ):
-            raise ConnectionResetError
-
-        self.world_parser.parse(message=self.__rcv_buffer[:msg_size].decode())
+            # 如果socket没有更多数据就退出
+            if len(select([self.__socket], [], [], 0.0)[0]) == 0:
+                break
 
     def commit_beam(self, pos2d: list, rotation: float) -> None:
         assert len(pos2d) == 2

readme.md

@@ -74,21 +74,6 @@ poetry run ./build_binary.sh <team-name>
 
 Once binary generation is finished, the result will be inside the build folder, as ```<team-name>.tar.gz```
 
-### GYM
-
-To use the gym, you need to install the following dependencies:
-```bash
-pip install gymnasium
-pip install psutil
-pip install stable-baselines3
-```
-
-Then, you can run gym examples under the ```GYM_CPU``` folder:
-```bash
-python3 -m scripts.gyms.Walk # Run the Walk gym example
-                             # of course, you can run other gym examples
-```
-
 ### Authors and acknowledgment
 This project was developed and contributed by:
 - **Chenxi Liu**

scripts/commons/Server.py

@@ -0,0 +1,71 @@
+import subprocess
+import os
+
+
+class Server():
+    def __init__(self, first_server_p, first_monitor_p, n_servers) -> None:
+        try:
+            import psutil
+            self.check_running_servers(psutil, first_server_p, first_monitor_p, n_servers)
+        except ModuleNotFoundError:
+            print("Info: Cannot check if the server is already running, because the psutil module was not found")
+
+        self.first_server_p = first_server_p
+        self.n_servers = n_servers
+        self.rcss_processes = []
+        first_monitor_p = first_monitor_p + 100
+
+        # makes it easier to kill test servers without affecting train servers
+        cmd = "rcssservermj"
+        for i in range(n_servers):
+            port = first_server_p + i
+            mport = first_monitor_p + i
+
+            server_cmd = f"{cmd} --aport {port} --mport {mport} "
+
+            self.rcss_processes.append(
+                subprocess.Popen(
+                    server_cmd.split(),
+                    stdout=subprocess.DEVNULL,
+                    stderr=subprocess.STDOUT,
+                    start_new_session=True
+                )
+            )
+
+    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 = []
+
+        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()]
+            if len(ports) == 0:
+                ports = [60000, 60100]  # default server ports (changing this is unlikely)
+
+            conflicts = [str(port) for port in ports if (
+                    (range1[0] <= port < range1[1]) or (range2[0] <= port < range2[1]))]
+
+            if len(conflicts) > 0:
+                if not found:
+                    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})")
+
+        if found:
+            print()
+            while True:
+                inp = input("Enter 'kill' to kill these processes or ctrl+c to abort. ")
+                if inp == "kill":
+                    for p in bad_processes:
+                        p.kill()
+                    return
+
+    def kill(self):
+        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

@@ -0,0 +1,601 @@
+from datetime import datetime, timedelta
+from itertools import count
+from os import listdir
+from os.path import isdir, join, isfile
+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 typing import Callable
+# from world.world import World
+from xml.dom import minidom
+import numpy as np
+import os, time, math, csv, select, sys
+import pickle
+import xml.etree.ElementTree as ET
+import shutil
+
+
+class Train_Base():
+    def __init__(self, script) -> None:
+        '''
+        When training with multiple environments (multiprocessing):
+            The server port is incremented as follows:
+                self.server_p, self.server_p+1, self.server_p+2, ...
+            We add +1000 to the initial monitor port, so than we can have more than 100 environments:
+                self.monitor_p+1000, self.monitor_p+1001, self.monitor_p+1002, ...
+        When testing we use self.server_p and self.monitor_p
+        '''
+
+        args = script.args
+        self.script = script
+        self.ip = args.i
+        self.server_p = args.p  # (initial) server port
+        self.monitor_p = args.m + 100  # monitor port when testing
+        self.monitor_p_1000 = args.m + 1100  # initial monitor port when training
+        self.robot_type = args.r
+        self.team = args.t
+        self.uniform = args.u
+        self.cf_last_time = 0
+        self.cf_delay = 0
+        # self.cf_target_period = World.STEPTIME # target simulation speed while testing (default: real-time)
+
+    @staticmethod
+    def prompt_user_for_model(self):
+
+        gyms_logs_path = "./scripts/gyms/logs/"
+        folders = [f for f in listdir(gyms_logs_path) if isdir(join(gyms_logs_path, f))]
+        folders.sort(key=lambda f: os.path.getmtime(join(gyms_logs_path, f)), reverse=True)  # sort by modification date
+
+        while True:
+            try:
+                folder_name = UI.print_list(folders, prompt="Choose folder (ctrl+c to return): ")[1]
+            except KeyboardInterrupt:
+                print()
+                return None  # ctrl+c
+
+            folder_dir = os.path.join(gyms_logs_path, folder_name)
+            models = [m[:-4] for m in listdir(folder_dir) if isfile(join(folder_dir, m)) and m.endswith(".zip")]
+
+            if not models:
+                print("The chosen folder does not contain any .zip file!")
+                continue
+
+            models.sort(key=lambda m: os.path.getmtime(join(folder_dir, m + ".zip")),
+                        reverse=True)  # sort by modification date
+
+            try:
+                model_name = UI.print_list(models, prompt="Choose model (ctrl+c to return): ")[1]
+                break
+            except KeyboardInterrupt:
+                print()
+
+        return {"folder_dir": folder_dir, "folder_name": folder_name,
+                "model_file": os.path.join(folder_dir, model_name + ".zip")}
+
+    # def control_fps(self, read_input = False):
+    #     ''' Add delay to control simulation speed '''
+
+    #     if read_input:
+    #         speed = input()
+    #         if speed == '':
+    #             self.cf_target_period = 0
+    #             print(f"Changed simulation speed to MAX")
+    #         else:
+    #             if speed == '0':
+    #                 inp = input("Paused. Set new speed or '' to use previous speed:")
+    #                 if inp != '':
+    #                     speed = inp
+
+    #             try:
+    #                 speed = int(speed)
+    #                 assert speed >= 0
+    #                 self.cf_target_period = World.STEPTIME * 100 / speed
+    #                 print(f"Changed simulation speed to {speed}%")
+    #             except:
+    #                 print("""Train_Base.py:
+    # Error: To control the simulation speed, enter a non-negative integer.
+    # To disable this control module, use test_model(..., enable_FPS_control=False) in your gyms environment.""")
+
+    #     now = time.time()
+    #     period = now - self.cf_last_time
+    #     self.cf_last_time = now
+    #     self.cf_delay += (self.cf_target_period - period)*0.9
+    #     if self.cf_delay > 0:
+    #         time.sleep(self.cf_delay)
+    #     else:
+    #         self.cf_delay = 0
+
+    def test_model(self, model: BaseAlgorithm, env, log_path: str = None, model_path: str = None, max_episodes=0,
+                   enable_FPS_control=True, verbose=1):
+        '''
+        Test model and log results
+
+        Parameters
+        ----------
+        model : BaseAlgorithm
+            Trained model
+        env : Env
+            Gym-like environment
+        log_path : str
+            Folder where statistics file is saved, default is `None` (no file is saved)
+        model_path : str
+            Folder where it reads evaluations.npz to plot it and create evaluations.csv, default is `None` (no plot, no csv)
+        max_episodes : int
+            Run tests for this number of episodes
+            Default is 0 (run until user aborts)
+        verbose : int
+            0 - no output (except if enable_FPS_control=True)
+            1 - print episode statistics
+        '''
+
+        if model_path is not None:
+            assert os.path.isdir(model_path), f"{model_path} is not a valid path"
+            self.display_evaluations(model_path)
+
+        if log_path is not None:
+            assert os.path.isdir(log_path), f"{log_path} is not a valid path"
+
+            # If file already exists, don't overwrite
+            if os.path.isfile(log_path + "/test.csv"):
+                for i in range(1000):
+                    p = f"{log_path}/test_{i:03}.csv"
+                    if not os.path.isfile(p):
+                        log_path = p
+                        break
+            else:
+                log_path += "/test.csv"
+
+            with open(log_path, 'w') as f:
+                f.write("reward,ep. length,rew. cumulative avg., ep. len. cumulative avg.\n")
+            print("Train statistics are saved to:", log_path)
+
+        if enable_FPS_control:  # control simulation speed (using non blocking user input)
+            print("\nThe simulation speed can be changed by sending a non-negative integer\n"
+                  "(e.g. '50' sets speed to 50%, '0' pauses the simulation, '' sets speed to MAX)\n")
+
+        ep_reward = 0
+        ep_length = 0
+        rewards_sum = 0
+        reward_min = math.inf
+        reward_max = -math.inf
+        ep_lengths_sum = 0
+        ep_no = 0
+
+        obs, _ = env.reset()
+        while True:
+            action, _states = model.predict(obs, deterministic=True)
+            obs, reward, terminated, truncated, info = env.step(action)
+            done = terminated or truncated
+            ep_reward += reward
+            ep_length += 1
+
+            # if enable_FPS_control:  # control simulation speed (using non blocking user input)
+            #     self.control_fps(select.select([sys.stdin], [], [], 0)[0])
+
+            if done:
+                obs, _ = env.reset()
+                rewards_sum += ep_reward
+                ep_lengths_sum += ep_length
+                reward_max = max(ep_reward, reward_max)
+                reward_min = min(ep_reward, reward_min)
+                ep_no += 1
+                avg_ep_lengths = ep_lengths_sum / ep_no
+                avg_rewards = rewards_sum / ep_no
+
+                if verbose > 0:
+                    print(
+                        f"\rEpisode: {ep_no:<3}  Ep.Length: {ep_length:<4.0f}  Reward: {ep_reward:<6.2f}                                                             \n",
+                        end=f"--AVERAGE--   Ep.Length: {avg_ep_lengths:<4.0f}  Reward: {avg_rewards:<6.2f}  (Min: {reward_min:<6.2f}  Max: {reward_max:<6.2f})",
+                        flush=True)
+
+                if log_path is not None:
+                    with open(log_path, 'a') as f:
+                        writer = csv.writer(f)
+                        writer.writerow([ep_reward, ep_length, avg_rewards, avg_ep_lengths])
+
+                if ep_no == max_episodes:
+                    return
+
+                ep_reward = 0
+                ep_length = 0
+
+    def learn_model(self, model: BaseAlgorithm, total_steps: int, path: str, eval_env=None, eval_freq=None, eval_eps=5,
+                    save_freq=None, backup_env_file=None, export_name=None):
+        '''
+        Learn Model for a specific number of time steps
+
+        Parameters
+        ----------
+        model : BaseAlgorithm
+            Model to train
+        total_steps : int
+            The total number of samples (env steps) to train on
+        path : str
+            Path where the trained model is saved
+            If the path already exists, an incrementing number suffix is added
+        eval_env : Env
+            Environment to periodically test the model
+            Default is None (no periodical evaluation)
+        eval_freq : int
+            Evaluate the agent every X steps
+            Default is None (no periodical evaluation)
+        eval_eps : int
+            Evaluate the agent for X episodes (both eval_env and eval_freq must be defined)
+            Default is 5
+        save_freq : int
+            Saves model at every X steps
+            Default is None (no periodical checkpoint)
+        backup_gym_file : str
+            Generates backup of environment file in model's folder
+            Default is None (no backup)
+        export_name : str
+            If export_name and save_freq are defined, a model is exported every X steps
+            Default is None (no export)
+
+        Returns
+        -------
+        model_path : str
+            Directory where model was actually saved (considering incremental suffix)
+
+        Notes
+        -----
+        If `eval_env` and `eval_freq` were specified:
+            - The policy will be evaluated in `eval_env` every `eval_freq` steps
+            - Evaluation results will be saved in `path` and shown at the end of training
+            - Every time the results improve, the model is saved
+        '''
+
+        start = time.time()
+        start_date = datetime.now().strftime("%d/%m/%Y %H:%M:%S")
+
+        # If path already exists, add suffix to avoid overwriting
+        if os.path.isdir(path):
+            for i in count():
+                p = path.rstrip("/") + f'_{i:03}/'
+                if not os.path.isdir(p):
+                    path = p
+                    break
+        os.makedirs(path)
+
+        # Backup environment file
+        if backup_env_file is not None:
+            backup_file = os.path.join(path, os.path.basename(backup_env_file))
+            copy(backup_env_file, backup_file)
+
+        evaluate = bool(eval_env is not None and eval_freq is not None)
+
+        # Create evaluation callback
+        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)
+
+        # Create custom callback to display evaluations
+        custom_callback = None if not evaluate else Cyclic_Callback(eval_freq,
+                                                                    lambda: self.display_evaluations(path, True))
+
+        # Create checkpoint callback
+        checkpoint_callback = None if save_freq is None else CheckpointCallback(save_freq=save_freq, save_path=path,
+                                                                                name_prefix="model", verbose=1)
+
+        # Create custom callback to export checkpoint models
+        export_callback = None if save_freq is None or export_name is None else Export_Callback(save_freq, path,
+                                                                                                export_name)
+
+        callbacks = CallbackList(
+            [c for c in [eval_callback, custom_callback, checkpoint_callback, export_callback] if c is not None])
+
+        model.learn(total_timesteps=total_steps, callback=callbacks)
+        model.save(os.path.join(path, "last_model"))
+
+        # Display evaluations if they exist
+        if evaluate:
+            self.display_evaluations(path)
+
+        # Display timestamps + Model path
+        end_date = datetime.now().strftime('%d/%m/%Y %H:%M:%S')
+        duration = timedelta(seconds=int(time.time() - start))
+        print(f"Train start:     {start_date}")
+        print(f"Train end:       {end_date}")
+        print(f"Train duration:  {duration}")
+        print(f"Model path:      {path}")
+
+        # Append timestamps to backup environment file
+        if backup_env_file is not None:
+            with open(backup_file, 'a') as f:
+                f.write(f"\n# Train start:    {start_date}\n")
+                f.write(f"# Train end:      {end_date}\n")
+                f.write(f"# Train duration: {duration}")
+
+        return path
+
+    def display_evaluations(self, path, save_csv=False):
+
+        eval_npz = os.path.join(path, "evaluations.npz")
+
+        if not os.path.isfile(eval_npz):
+            return
+
+        console_width = 80
+        console_height = 18
+        symb_x = "\u2022"
+        symb_o = "\u007c"
+        symb_xo = "\u237f"
+
+        with np.load(eval_npz) as data:
+            time_steps = data["timesteps"]
+            results_raw = np.mean(data["results"], axis=1)
+            ep_lengths_raw = np.mean(data["ep_lengths"], axis=1)
+        sample_no = len(results_raw)
+
+        xvals = np.linspace(0, sample_no - 1, 80)
+        results = np.interp(xvals, range(sample_no), results_raw)
+        ep_lengths = np.interp(xvals, range(sample_no), ep_lengths_raw)
+
+        results_limits = np.min(results), np.max(results)
+        ep_lengths_limits = np.min(ep_lengths), np.max(ep_lengths)
+
+        results_discrete = np.digitize(results, np.linspace(results_limits[0] - 1e-5, results_limits[1] + 1e-5,
+                                                            console_height + 1)) - 1
+        ep_lengths_discrete = np.digitize(ep_lengths,
+                                          np.linspace(0, ep_lengths_limits[1] + 1e-5, console_height + 1)) - 1
+
+        matrix = np.zeros((console_height, console_width, 2), int)
+        matrix[results_discrete[0]][0][0] = 1  # draw 1st column
+        matrix[ep_lengths_discrete[0]][0][1] = 1  # draw 1st column
+        rng = [[results_discrete[0], results_discrete[0]], [ep_lengths_discrete[0], ep_lengths_discrete[0]]]
+
+        # Create continuous line for both plots
+        for k in range(2):
+            for i in range(1, console_width):
+                x = [results_discrete, ep_lengths_discrete][k][i]
+                if x > rng[k][1]:
+                    rng[k] = [rng[k][1] + 1, x]
+                elif x < rng[k][0]:
+                    rng[k] = [x, rng[k][0] - 1]
+                else:
+                    rng[k] = [x, x]
+                for j in range(rng[k][0], rng[k][1] + 1):
+                    matrix[j][i][k] = 1
+
+        print(f'{"-" * console_width}')
+        for l in reversed(range(console_height)):
+            for c in range(console_width):
+                if np.all(matrix[l][c] == 0):
+                    print(end=" ")
+                elif np.all(matrix[l][c] == 1):
+                    print(end=symb_xo)
+                elif matrix[l][c][0] == 1:
+                    print(end=symb_x)
+                else:
+                    print(end=symb_o)
+            print()
+        print(f'{"-" * console_width}')
+        print(f"({symb_x})-reward          min:{results_limits[0]:11.2f}    max:{results_limits[1]:11.2f}")
+        print(
+            f"({symb_o})-ep. length      min:{ep_lengths_limits[0]:11.0f}    max:{ep_lengths_limits[1]:11.0f}    {time_steps[-1] / 1000:15.0f}k steps")
+        print(f'{"-" * console_width}')
+
+        # save CSV
+        if save_csv:
+            eval_csv = os.path.join(path, "evaluations.csv")
+            with open(eval_csv, 'a+') as f:
+                writer = csv.writer(f)
+                if sample_no == 1:
+                    writer.writerow(["time_steps", "reward ep.", "length"])
+                writer.writerow([time_steps[-1], results_raw[-1], ep_lengths_raw[-1]])
+
+    # def generate_slot_behavior(self, path, slots, auto_head:bool, XML_name):
+    #     '''
+    #     Function that generates the XML file for the optimized slot behavior, overwriting previous files
+    #     '''
+
+    #     file = os.path.join( path, XML_name )
+
+    #     # create the file structure
+    #     auto_head = '1' if auto_head else '0'
+    #     EL_behavior = ET.Element('behavior',{'description':'Add description to XML file', "auto_head":auto_head})
+
+    #     for i,s in enumerate(slots):
+    #         EL_slot = ET.SubElement(EL_behavior, 'slot', {'name':str(i), 'delta':str(s[0]/1000)})
+    #         for j in s[1]: # go through all joint indices
+    #             ET.SubElement(EL_slot, 'move', {'id':str(j), 'angle':str(s[2][j])})
+
+    #     # create XML file
+    #     xml_rough = ET.tostring( EL_behavior, 'utf-8' )
+    #     xml_pretty = minidom.parseString(xml_rough).toprettyxml(indent="  ")
+    #     with open(file, "w") as x:
+    #         x.write(xml_pretty)
+
+    #     print(file, "was created!")
+
+    # @staticmethod
+    # def linear_schedule(initial_value: float) -> Callable[[float], float]:
+    #     '''
+    #     Linear learning rate schedule
+
+    #     Parameters
+    #     ----------
+    #     initial_value : float
+    #         Initial learning rate
+
+    #     Returns
+    #     -------
+    #     schedule : Callable[[float], float]
+    #         schedule that computes current learning rate depending on remaining progress
+    #     '''
+    #     def func(progress_remaining: float) -> float:
+    #         '''
+    #         Compute learning rate according to current progress
+
+    #         Parameters
+    #         ----------
+    #         progress_remaining : float
+    #             Progress will decrease from 1 (beginning) to 0
+
+    #         Returns
+    #         -------
+    #         learning_rate : float
+    #             Learning rate according to current progress
+    #         '''
+    #         return progress_remaining * initial_value
+
+    #     return func
+
+    @staticmethod
+    def export_model(input_file, output_file, add_sufix=True):
+        '''
+        Export model weights to binary file
+
+        Parameters
+        ----------
+        input_file : str
+            Input file, compatible with algorithm
+        output_file : str
+            Output file, including directory
+        add_sufix : bool
+            If true, a suffix is appended to the file name: output_file + "_{index}.pkl"
+        '''
+
+        # If file already exists, don't overwrite
+        if add_sufix:
+            for i in count():
+                f = f"{output_file}_{i:03}.pkl"
+                if not os.path.isfile(f):
+                    output_file = f
+                    break
+
+        model = PPO.load(input_file)
+        weights = model.policy.state_dict()  # dictionary containing network layers
+
+        w = lambda name: weights[name].detach().cpu().numpy()  # extract weights from policy
+
+        var_list = []
+        for i in count(0, 2):  # add hidden layers (step=2 because that's how SB3 works)
+            if f"mlp_extractor.policy_net.{i}.bias" not in weights:
+                break
+            var_list.append(
+                [w(f"mlp_extractor.policy_net.{i}.bias"), w(f"mlp_extractor.policy_net.{i}.weight"), "tanh"])
+
+        var_list.append([w("action_net.bias"), w("action_net.weight"), "none"])  # add final layer
+
+        with open(output_file, "wb") as f:
+            pickle.dump(var_list, f, protocol=4)  # protocol 4 is backward compatible with Python 3.4
+
+    def print_list(data, numbering=True, prompt=None, divider=" | ", alignment="<", min_per_col=6):
+        '''
+        Print list - prints list, using as many columns as possible
+
+        Parameters
+        ----------
+        data : `list`
+            list of items
+        numbering : `bool`
+            assigns number to each option
+        prompt : `str`
+            the prompt string, if given, is printed after the table before reading input
+        divider : `str`
+            string that divides columns
+        alignment : `str`
+            f-string style alignment ( '<', '>', '^' )
+        min_per_col : int
+            avoid splitting columns with fewer items
+
+        Returns
+        -------
+        item : `int`, item
+            returns tuple with global index of selected item and the item object,
+            or `None` (if `numbering` or `prompt` are `None`)
+
+        '''
+
+        WIDTH = shutil.get_terminal_size()[0]
+
+        data_size = len(data)
+        items = []
+        items_len = []
+
+        # --------------------------------------------- Add numbers, margins and divider
+        for i in range(data_size):
+            number = f"{i}-" if numbering else ""
+            items.append(f"{divider}{number}{data[i]}")
+            items_len.append(len(items[-1]))
+
+        max_cols = np.clip((WIDTH + len(divider)) // min(items_len), 1, math.ceil(
+            data_size / max(min_per_col, 1)))  # width + len(divider) because it is not needed in last col
+
+        # --------------------------------------------- Check maximum number of columns, considering content width (min:1)
+        for i in range(max_cols, 0, -1):
+            cols_width = []
+            cols_items = []
+            table_width = 0
+            a, b = divmod(data_size, i)
+            for col in range(i):
+                start = a * col + min(b, col)
+                end = start + a + (1 if col < b else 0)
+                cols_items.append(items[start:end])
+                col_width = max(items_len[start:end])
+                cols_width.append(col_width)
+                table_width += col_width
+            if table_width <= WIDTH + len(divider):
+                break
+        table_width -= len(divider)
+
+        # --------------------------------------------- Print columns
+        print("=" * table_width)
+        for row in range(math.ceil(data_size / i)):
+            for col in range(i):
+                content = cols_items[col][row] if len(
+                    cols_items[col]) > row else divider  # print divider when there are no items
+                if col == 0:
+                    l = len(divider)
+                    print(end=f"{content[l:]:{alignment}{cols_width[col] - l}}")  # remove divider from 1st col
+                else:
+                    print(end=f"{content    :{alignment}{cols_width[col]}}")
+            print()
+        print("=" * table_width)
+
+        # --------------------------------------------- Prompt
+        if prompt is None:
+            return None
+
+        if numbering is None:
+            return None
+        else:
+            idx = UI.read_int(prompt, 0, data_size)
+            return idx, data[idx]
+
+
+class Cyclic_Callback(BaseCallback):
+    ''' Stable baselines custom callback '''
+
+    def __init__(self, freq, function):
+        super(Cyclic_Callback, self).__init__(1)
+        self.freq = freq
+        self.function = function
+
+    def _on_step(self) -> bool:
+        if self.n_calls % self.freq == 0:
+            self.function()
+        return True  # If the callback returns False, training is aborted early
+
+
+class Export_Callback(BaseCallback):
+    ''' Stable baselines custom callback '''
+
+    def __init__(self, freq, load_path, export_name):
+        super(Export_Callback, self).__init__(1)
+        self.freq = freq
+        self.load_path = load_path
+        self.export_name = export_name
+
+    def _on_step(self) -> bool:
+        if self.n_calls % self.freq == 0:
+            path = os.path.join(self.load_path, f"model_{self.num_timesteps}_steps.zip")
+            Train_Base.export_model(path, f"./scripts/gyms/export/{self.export_name}")
+        return True  # If the callback returns False, training is aborted early
+
+
+

scripts/commons/UI.py

@@ -0,0 +1,302 @@
+from itertools import zip_longest
+from math import inf
+import math
+import numpy as np
+import shutil
+
+class UI():
+    console_width = 80
+    console_height = 24
+    
+    @staticmethod
+    def read_particle(prompt, str_options, dtype=str, interval=[-inf,inf]):
+        ''' 
+        Read particle from user from a given dtype or from a str_options list
+
+        Parameters
+        ----------
+        prompt : `str`
+            prompt to show user before reading input
+        str_options : `list`
+            list of str options (in addition to dtype if dtype is not str)
+        dtype : `class`
+            if dtype is str, then user must choose a value from str_options, otherwise it can also send a dtype value
+        interval : `list`
+            [>=min,<max] interval for numeric dtypes
+        
+        Returns
+        -------
+        choice : `int` or dtype
+            index of str_options (int) or value (dtype)
+        is_str_option : `bool`
+            True if `choice` is an index from str_options
+        '''
+        # Check if user has no choice
+        if dtype is str and len(str_options) == 1:
+            print(prompt, str_options[0], sep="")
+            return 0, True
+        elif dtype is int and interval[0] == interval[1]-1:
+            print(prompt, interval[0], sep="")
+            return interval[0], False
+
+        while True:
+            inp = input(prompt)
+            if inp in str_options: 
+                return str_options.index(inp), True
+
+            if dtype is not str:
+                try:
+                    inp = dtype(inp)
+                    if inp >= interval[0] and inp < interval[1]:
+                        return inp, False
+                except:
+                    pass
+            
+            print("Error: illegal input! Options:", str_options, f" or  {dtype}" if dtype != str else "")
+
+    @staticmethod
+    def read_int(prompt, min, max):
+        ''' 
+        Read int from user in a given interval
+        :param prompt: prompt to show user before reading input
+        :param min: minimum input (inclusive)
+        :param max: maximum input (exclusive)
+        :return: choice
+        '''
+        while True:
+            inp = input(prompt)
+            try:
+                inp = int(inp)
+                assert inp >= min and inp < max
+                return inp
+            except:
+                print(f"Error: illegal input! Choose number between {min} and {max-1}")
+
+    @staticmethod
+    def print_table(data, titles=None, alignment=None, cols_width=None, cols_per_title=None, margins=None, numbering=None, prompt=None):
+        '''
+        Print table
+        
+        Parameters
+        ----------
+        data : `list`
+            list of columns, where each column is a list of items
+        titles : `list`
+            list of titles for each column, default is `None` (no titles)
+        alignment : `list`
+            list of alignments per column (excluding titles), default is `None` (left alignment for all cols)
+        cols_width : `list`
+            list of widths per column, default is `None` (fit to content)
+            Positive values indicate a fixed column width
+            Zero indicates that the column will fit its content
+        cols_per_title : `list`
+            maximum number of subcolumns per title, default is `None` (1 subcolumn per title)
+        margins : `list`
+            number of added leading and trailing spaces per column, default is `None` (margin=2 for all columns)
+        numbering : `list`
+            list of booleans per columns, indicating whether to assign numbers to each option
+        prompt : `str`
+            the prompt string, if given, is printed after the table before reading input
+
+        Returns
+        -------
+        index : `int`
+            returns global index of selected item (relative to table)
+        col_index : `int`
+            returns local index of selected item (relative to column)
+        column : `int`
+            returns number of column of selected item (starts at 0)
+        * if `numbering` or `prompt` are `None`, `None` is returned
+        
+
+        Example
+        -------
+        titles = ["Name","Age"]
+        data = [[John,Graciete], [30,50]]
+        alignment = ["<","^"]               # 1st column is left-aligned, 2nd is centered
+        cols_width = [10,5]                # 1st column's width=10, 2nd column's width=5
+        margins = [3,3]                    
+        numbering = [True,False]           # prints: [0-John,1-Graciete][30,50]
+        prompt = "Choose a person:"
+        '''
+        
+        #--------------------------------------------- parameters
+        cols_no = len(data)
+
+        if alignment is None:
+            alignment = ["<"]*cols_no
+
+        if cols_width is None:
+            cols_width = [0]*cols_no
+
+        if numbering is None:
+            numbering = [False]*cols_no
+            any_numbering = False
+        else:
+            any_numbering = True
+
+        if margins is None:
+            margins = [2]*cols_no
+
+        # Fit column to content + margin, if required
+        subcol = [] # subcolumn length and widths
+        for i in range(cols_no):
+            subcol.append([[],[]])
+            if cols_width[i] == 0:
+                numbering_width = 4 if numbering[i] else 0
+                if cols_per_title is None or cols_per_title[i] < 2:
+                    cols_width[i] = max([len(str(item))+numbering_width for item in data[i]]) + margins[i]*2
+                else:
+                    subcol[i][0] = math.ceil(len(data[i])/cols_per_title[i]) # subcolumn maximum length
+                    cols_per_title[i] = math.ceil(len(data[i])/subcol[i][0]) # reduce number of columns as needed
+                    cols_width[i] = margins[i]*(1+cols_per_title[i]) - (1 if numbering[i] else 0) # remove one if numbering, same as when printing
+                    for j in range(cols_per_title[i]):
+                        subcol_data_width = max([len(str(item))+numbering_width for item in data[i][j*subcol[i][0]:j*subcol[i][0]+subcol[i][0]]])
+                        cols_width[i] += subcol_data_width     # add subcolumn data width to column width
+                        subcol[i][1].append(subcol_data_width) # save subcolumn data width
+                        
+                if titles is not None: # expand to acomodate titles if needed
+                    cols_width[i] = max(cols_width[i], len(titles[i]) + margins[i]*2  )
+
+        if any_numbering:
+            no_of_items=0
+            cumulative_item_per_col=[0] # useful for getting the local index
+            for i in range(cols_no):
+                assert type(data[i]) == list, "In function 'print_table', 'data' must be a list of lists!"
+
+                if numbering[i]:
+                    data[i] = [f"{n+no_of_items:3}-{d}" for n,d in enumerate(data[i])]
+                    no_of_items+=len(data[i])
+                cumulative_item_per_col.append(no_of_items)
+
+        table_width = sum(cols_width)+cols_no-1
+
+        #--------------------------------------------- col titles
+        print(f'{"="*table_width}')
+        if titles is not None:
+            for i in range(cols_no):
+                print(f'{titles[i]:^{cols_width[i]}}', end='|' if i < cols_no - 1 else '')
+            print()
+            for i in range(cols_no):
+                print(f'{"-"*cols_width[i]}', end='+' if i < cols_no - 1 else '')
+            print()
+
+        #--------------------------------------------- merge subcolumns
+        if cols_per_title is not None:
+            for i,col in enumerate(data):
+                if cols_per_title[i] < 2:
+                    continue
+                for k in range(subcol[i][0]): # create merged items
+                    col[k] = (" "*margins[i]).join( f'{col[item]:{alignment[i]}{subcol[i][1][subcol_idx]}}' 
+                                                    for subcol_idx, item in enumerate(range(k,len(col),subcol[i][0])) )
+                del col[subcol[i][0]:] # delete repeated items
+        
+        #--------------------------------------------- col items
+        for line in zip_longest(*data):       
+            for i,item in enumerate(line):
+                l_margin = margins[i]-1 if numbering[i] else margins[i] # adjust margins when there are numbered options
+                item = "" if item is None else f'{" "*l_margin}{item}{" "*margins[i]}' # add margins
+                print(f'{item:{alignment[i]}{cols_width[i]}}', end='')
+                if i < cols_no - 1:
+                    print(end='|')
+            print(end="\n")
+        print(f'{"="*table_width}')
+
+        #--------------------------------------------- prompt
+        if prompt is None:
+            return None
+
+        if not any_numbering:
+            print(prompt)
+            return None
+
+        index = UI.read_int(prompt, 0, no_of_items)
+
+        for i,n in enumerate(cumulative_item_per_col):
+            if index < n:
+                return index, index-cumulative_item_per_col[i-1], i-1
+
+        raise ValueError('Failed to catch illegal input')
+
+
+    @staticmethod
+    def print_list(data, numbering=True, prompt=None, divider=" | ", alignment="<", min_per_col=6):
+        '''
+        Print list - prints list, using as many columns as possible
+        
+        Parameters
+        ----------
+        data : `list`
+            list of items
+        numbering : `bool`
+            assigns number to each option
+        prompt : `str`
+            the prompt string, if given, is printed after the table before reading input
+        divider : `str`
+            string that divides columns
+        alignment : `str`
+            f-string style alignment ( '<', '>', '^' )
+        min_per_col : int
+            avoid splitting columns with fewer items
+        
+        Returns
+        -------
+        item : `int`, item
+            returns tuple with global index of selected item and the item object,
+            or `None` (if `numbering` or `prompt` are `None`)
+
+        '''
+        
+        WIDTH = shutil.get_terminal_size()[0]
+
+        data_size = len(data)   
+        items = []
+        items_len = []
+
+        #--------------------------------------------- Add numbers, margins and divider
+        for i in range(data_size):
+            number = f"{i}-" if numbering else ""
+            items.append( f"{divider}{number}{data[i]}" )
+            items_len.append( len(items[-1]) )
+
+        max_cols = np.clip((WIDTH+len(divider)) // min(items_len),1,math.ceil(data_size/max(min_per_col,1))) # width + len(divider) because it is not needed in last col
+
+        #--------------------------------------------- Check maximum number of columns, considering content width (min:1)
+        for i in range(max_cols,0,-1):
+            cols_width = []
+            cols_items = []
+            table_width = 0
+            a,b = divmod(data_size,i)
+            for col in range(i):
+                start = a*col + min(b,col)
+                end = start+a+(1 if col<b else 0)
+                cols_items.append( items[start:end] )
+                col_width = max(items_len[start:end])
+                cols_width.append( col_width )
+                table_width += col_width
+            if table_width <= WIDTH+len(divider):
+                break
+        table_width -= len(divider)
+        
+        #--------------------------------------------- Print columns
+        print("="*table_width)
+        for row in range(math.ceil(data_size / i)):
+            for col in range(i):
+                content = cols_items[col][row] if len(cols_items[col]) > row else divider # print divider when there are no items
+                if col == 0:
+                    l = len(divider)
+                    print(end=f"{content[l:]:{alignment}{cols_width[col]-l}}")  # remove divider from 1st col
+                else:
+                    print(end=f"{content    :{alignment}{cols_width[col]  }}") 
+            print()  
+        print("="*table_width)
+
+        #--------------------------------------------- Prompt
+        if prompt is None:
+            return None
+
+        if numbering is None:
+            return None
+        else:
+            idx = UI.read_int( prompt, 0, data_size )
+            return idx, data[idx]

scripts/gyms/Walk.py

@@ -0,0 +1,640 @@
+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.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 = 1000
+        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.reference_joint_nominal_position = self.joint_nominal_position.copy()
+        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
+
+        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})"
+        )
+
+    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 calibrate_train_sim_flip_from_neutral(self, neutral_joint_positions):
+        updated_flip = self.train_sim_flip.copy()
+        changed = []
+
+        for idx, (reference_value, observed_value) in enumerate(
+                zip(self.reference_joint_nominal_position, neutral_joint_positions)
+        ):
+            if idx >= 10:
+                continue
+            if abs(reference_value) < 0.15 or abs(observed_value) < 0.15:
+                continue
+
+            inferred_flip = 1.0 if np.sign(reference_value) == np.sign(observed_value) else -1.0
+            if updated_flip[idx] != inferred_flip:
+                changed.append((idx, updated_flip[idx], inferred_flip))
+                updated_flip[idx] = inferred_flip
+
+        self.train_sim_flip = updated_flip
+
+        if changed:
+            self.debug_log(
+                "[FlipDebug] "
+                f"changes={[(idx, old, new) for idx, old, new in changed]}"
+            )
+
+    def is_reliable_neutral_pose(self, neutral_joint_positions):
+        leg_positions = neutral_joint_positions[11:]
+        leg_norm = float(np.linalg.norm(leg_positions))
+        leg_max = float(np.max(np.abs(leg_positions)))
+        height = float(self.Player.world.global_position[2])
+
+        reliable = (
+                leg_norm > 0.8
+                and leg_max > 0.35
+                and 0.12 < height < 0.8
+        )
+
+        return reliable, leg_norm, leg_max, height
+
+    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}"
+        )
+
+    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 = np.random.uniform(10, 20)  # randomize target distance
+        length2 = np.random.uniform(10, 20)  # randomize target distance
+        length3 = np.random.uniform(10, 20)  # 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
+
+        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=0)
+            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(10):
+            finished = self.Player.skills_manager.execute("Neutral")
+            self.sync()
+            if finished:
+                finished_count += 1
+                if finished_count >= 3:  # 假设需要连续3次完成才算成功
+                    break
+
+        # neutral_joint_positions = np.deg2rad(
+        #     [self.Player.robot.motor_positions[motor] for motor in self.Player.robot.ROBOT_MOTORS]
+        # )
+        # reliable_neutral, neutral_leg_norm, neutral_leg_max, neutral_height = self.is_reliable_neutral_pose(neutral_joint_positions)
+
+        # if self.auto_calibrate_train_sim_flip and reliable_neutral and not self.flip_calibrated_once:
+        #     self.calibrate_train_sim_flip_from_neutral(neutral_joint_positions)
+        #     self.flip_calibrated_once = True
+        # if self.calibrate_nominal_from_neutral and reliable_neutral and not self.nominal_calibrated_once:
+        #     self.joint_nominal_position = neutral_joint_positions * self.train_sim_flip
+        #     self.nominal_calibrated_once = True
+        # self.debug_log(
+        #     "[ResetDebug] "
+        #     f"neutral_pos={np.round(self.Player.world.global_position, 3).tolist()} "
+        #     f"shoulders={np.round(neutral_joint_positions[2:10], 3).tolist()} "
+        #     f"legs={np.round(neutral_joint_positions[11:], 3).tolist()} "
+        #     f"flip={self.train_sim_flip.tolist()} "
+        #     f"nominal_legs={np.round(self.joint_nominal_position[11:], 3).tolist()} "
+        #     f"calibrated_once={(self.flip_calibrated_once, self.nominal_calibrated_once)} "
+        #     f"reliable_neutral={reliable_neutral} "
+        #     f"leg_norm={neutral_leg_norm:.3f} leg_max={neutral_leg_max:.3f} height={neutral_height:.3f}"
+        # )
+
+        # reset_action_noise = np.random.uniform(-0.015, 0.015, size=(len(self.Player.robot.ROBOT_MOTORS),))
+        # self.target_joint_positions = (self.joint_nominal_position + reset_action_noise) * 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=25, kd=0.6
+        #     )
+
+        # memory variables
+        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)
+        point1 = self.initial_position + np.array([length1, 0])
+        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, point2, point3]
+        self.target_position = self.point_list[self.waypoint_index]
+
+        return self.observe(True), {}
+
+    def render(self, mode='human', close=False):
+        return
+
+    def compute_reward(self, previous_pos, current_pos, action):
+        eps = 1e-6
+        dt = 0.05
+
+        velocity = current_pos - previous_pos
+        speed_step = float(np.linalg.norm(velocity))
+        speed = speed_step / dt
+
+        direction_to_target = self.target_position - current_pos
+        prev_direction_to_target = self.target_position - previous_pos
+        distance_to_target = float(np.linalg.norm(direction_to_target))
+        prev_distance_to_target = float(np.linalg.norm(prev_direction_to_target))
+
+        # Progress toward waypoint (secondary signal)
+        progress = prev_distance_to_target - distance_to_target
+        progress_reward = np.clip(progress * 2.0, -1.5, 2.5)
+
+        # Forward speed and lateral drift
+        forward_dir = direction_to_target / max(distance_to_target, eps)
+        forward_speed = float(np.dot(velocity, forward_dir)) / dt
+        target_speed = 1.0
+        speed_error = forward_speed - target_speed
+        speed_reward = 3.0 * math.exp(-1.5 * (speed_error ** 2))
+
+        lateral_vec = velocity - forward_dir * np.dot(velocity, forward_dir)
+        lateral_speed = float(np.linalg.norm(lateral_vec)) / dt
+        lateral_penalty = -0.6 * np.clip(lateral_speed, 0.0, 2.0)
+
+        # Heading alignment (small shaping term)
+        if speed_step > 1e-4 and distance_to_target > 1e-4:
+            directional_alignment = np.dot(velocity, direction_to_target) / (speed_step * distance_to_target)
+            directional_alignment = float(np.clip(directional_alignment, -1.0, 1.0))
+            direction_reward = max(0.0, directional_alignment) * 0.3
+        else:
+            direction_reward = 0.0
+
+        alive_bonus = 0.05
+
+        # Height and posture
+        height = float(self.Player.world.global_position[2])
+        if 0.8 <= height <= 1.05:
+            height_reward = 1.0
+        elif 0.40 <= height <= 1.20:
+            height_reward = -1.0
+        else:
+            height_reward = -6.0
+
+        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]))
+        posture_penalty = -2.2 * (tilt_mag ** 2)
+
+        motionless_penalty = -1.5 if speed < 0.1 else 0.0
+
+        # Waypoint bonus
+        waypoint_bonus = 0.0
+        if distance_to_target < 0.5:
+            waypoint_bonus = 25.0
+            if self.waypoint_index < len(self.point_list) - 1:
+                self.waypoint_index += 1
+                self.target_position = self.point_list[self.waypoint_index]
+            else:
+                waypoint_bonus = 100.0
+                self.route_completed = True
+
+        # Effort + smoothness
+        action_magnitude = float(np.linalg.norm(action[11:]))
+        action_penalty = -0.05 * action_magnitude
+        action_delta = action - self.last_action_for_reward
+        smoothness_penalty = -0.02 * float(np.linalg.norm(action_delta[11:]))
+
+        return (
+                progress_reward
+                + speed_reward
+                + lateral_penalty
+                + direction_reward
+                + alive_bonus
+                + height_reward
+                + posture_penalty
+                + motionless_penalty
+                + waypoint_bonus
+                + action_penalty
+                + smoothness_penalty
+        )
+
+    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=25, kd=0.6
+            )
+
+        self.sync()  # run simulation step
+        self.step_counter += 1
+
+        # if 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 = 512  # RolloutBuffer is of size (n_steps_per_env * n_envs)
+        minibatch_size = 128  # should be a factor of (n_steps_per_env * n_envs)
+        total_steps = 30000000
+        learning_rate = 2e-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=0.01,  # Entropy coefficient for exploration
+                    # clip_range=0.2,  # PPO clipping parameter
+                    # gae_lambda=0.95,  # GAE lambda
+                    # gamma=0.99  # Discount factor
+                )
+
+            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({})
+    # trainer.test({"model_file": "scripts/gyms/logs/Walk_R0_003/best_model.zip",
+    #               "folder_dir": "Walk_R0_003",})