Init

deploy/deploy_mujoco/configs/go2.yaml

@@ -0,0 +1,61 @@
+policy_path: "{LEGGED_GYM_ROOT_DIR}/deploy/pre_train/go2/go2_cts_150k.pt"
+
+# xml_path: "{LEGGED_GYM_ROOT_DIR}/resources/robots/go2/flat.xml"
+# xml_path: "{LEGGED_GYM_ROOT_DIR}/resources/robots/go2/race_track.xml"  # change go2 init pos to pos="-5 2 0.445" in resources/robots/go2/go2.xml
+xml_path: "{LEGGED_GYM_ROOT_DIR}/resources/robots/go2/stairs.xml"
+
+# Total simulation time
+simulation_duration: 60000000.0
+# Simulation time step
+simulation_dt: 0.002
+# Controller update frequency (meets the requirement of simulation_dt * controll_decimation=0.02; 50Hz)
+control_decimation: 10
+
+kps: [20.0, 20.0, 20.0, 20.0, 20.0, 20.0, 20.0, 20.0, 20.0, 20.0, 20.0, 20.0] 
+kds: [0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5, 0.5] 
+
+default_angles: [0.1,  0.8,  -1.5,  -0.1, 0.8, -1.5, 
+                 0.1,  1.0,  -1.5,  -0.1, 1.0, -1.5] 
+
+lin_vel_scale: 2.0
+ang_vel_scale: 0.25 
+dof_pos_scale: 1.0 
+dof_vel_scale: 0.05 
+action_scale: 0.25 
+cmd_scale: [2.0, 2.0, 0.25]
+num_actions: 12
+num_obs: 45
+max_cmd: [2.0, 1.5, 2.5]
+
+cmd_init: [1.0, 0.0, 0.0]
+
+mujoco_joint_names: [
+  "FL_hip_joint",
+  "FL_thigh_joint",
+  "FL_calf_joint",
+  "FR_hip_joint",
+  "FR_thigh_joint",
+  "FR_calf_joint",
+  "RL_hip_joint",
+  "RL_thigh_joint",
+  "RL_calf_joint",
+  "RR_hip_joint",
+  "RR_thigh_joint",
+  "RR_calf_joint"
+]
+
+# IsaacGym is same as Mujoco joint names
+model_joint_names: [
+  "FL_hip_joint",
+  "FL_thigh_joint",
+  "FL_calf_joint",
+  "FR_hip_joint",
+  "FR_thigh_joint",
+  "FR_calf_joint",
+  "RL_hip_joint",
+  "RL_thigh_joint",
+  "RL_calf_joint",
+  "RR_hip_joint",
+  "RR_thigh_joint",
+  "RR_calf_joint"
+]

deploy/deploy_mujoco/deploy_go2.py

@@ -0,0 +1,251 @@
+import time
+import mujoco.viewer
+import mujoco
+import numpy as np
+from legged_gym import LEGGED_GYM_ROOT_DIR
+import torch
+import yaml
+import os
+import imageio
+from pathlib import Path
+from argparse import ArgumentParser
+import pygame
+from matplotlib import pyplot as plt
+
+
+def get_gravity_orientation(quaternion):
+    qw = quaternion[0]
+    qx = quaternion[1]
+    qy = quaternion[2]
+    qz = quaternion[3]
+
+    gravity_orientation = np.zeros(3)
+
+    gravity_orientation[0] = 2 * (-qz * qx + qw * qy)
+    gravity_orientation[1] = -2 * (qz * qy + qw * qx)
+    gravity_orientation[2] = 1 - 2 * (qw * qw + qz * qz)
+
+    return gravity_orientation
+
+
+def pd_control(target_q, q, kp, target_dq, dq, kd):
+    """Calculates torques from position commands"""
+    return (target_q - q) * kp + (target_dq - dq) * kd
+
+def get_xbox_command(joystick, max_cmd):
+    pygame.event.pump()
+    dead_zone = 0.1
+    lx = joystick.get_axis(0)
+    ly = joystick.get_axis(1)
+    rx = joystick.get_axis(3)
+    if abs(lx) < dead_zone: lx = 0
+    if abs(ly) < dead_zone: ly = 0
+    if abs(rx) < dead_zone: rx = 0
+    cmd_x = -ly * max_cmd[0]
+    cmd_y = -lx * max_cmd[1]
+    cmd_yaw = -rx * max_cmd[2]
+    return np.array([cmd_x, cmd_y, cmd_yaw], dtype=np.float32)
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--save-video", action="store_true", help="Whether to save video of the simulation.")
+    parser.add_argument("--visualize-moe-weights", action="store_true", help="Whether to visualize mixture of experts weights.")
+    parser.add_argument("--save-moe-latent", action="store_true", help="Whether to save mixture of experts latent vectors.")
+    args = parser.parse_args()
+    save_video = args.save_video
+    visualize_moe_weights = args.visualize_moe_weights
+    save_moe_latent = args.save_moe_latent
+    config_file = "go2.yaml"
+
+    pygame.init()
+    use_joystick = False
+    joystick = None
+    if pygame.joystick.get_count() > 0:
+        joystick = pygame.joystick.Joystick(0)
+        joystick.init()
+        use_joystick = True
+        print(f"Detected Joystick: {joystick.get_name()}")
+    else:
+        print("No Joystick detected. Using default commands from config.")
+
+    with open(f"{LEGGED_GYM_ROOT_DIR}/deploy/deploy_mujoco/configs/{config_file}", "r") as f:
+        config = yaml.load(f, Loader=yaml.FullLoader)
+        policy_path = config["policy_path"].replace("{LEGGED_GYM_ROOT_DIR}", LEGGED_GYM_ROOT_DIR)
+        xml_path = config["xml_path"].replace("{LEGGED_GYM_ROOT_DIR}", LEGGED_GYM_ROOT_DIR)
+
+        simulation_duration = config["simulation_duration"]
+        simulation_dt = config["simulation_dt"]
+        control_decimation = config["control_decimation"]
+
+        kps = np.array(config["kps"], dtype=np.float32)
+        kds = np.array(config["kds"], dtype=np.float32)
+
+        default_angles = np.array(config["default_angles"], dtype=np.float32)
+
+        lin_vel_scale = config["lin_vel_scale"]
+        ang_vel_scale = config["ang_vel_scale"]
+        dof_pos_scale = config["dof_pos_scale"]
+        dof_vel_scale = config["dof_vel_scale"]
+        action_scale = config["action_scale"]
+        cmd_scale = np.array(config["cmd_scale"], dtype=np.float32)
+
+        num_actions = config["num_actions"]
+        num_obs = config["num_obs"]
+
+        cmd = np.array(config["cmd_init"], dtype=np.float32)
+
+        idx_model2mj = idx_mj2model = list(range(num_actions))
+        if 'mujoco_joint_names' in config and 'model_joint_names' in config:
+            mujoco_joint_names = config["mujoco_joint_names"]
+            model_joint_names = config["model_joint_names"]
+            idx_model2mj = [model_joint_names.index(joint) for joint in mujoco_joint_names]
+            idx_mj2model = [mujoco_joint_names.index(joint) for joint in model_joint_names]
+
+    video_save_dir = str(Path(__file__).parent / "videos")
+    os.makedirs(video_save_dir, exist_ok=True)
+
+    model_name = os.path.basename(policy_path).split('.')[0]
+    cmd_str = f"cmd_{cmd[0]}_{cmd[1]}_{cmd[2]}"
+
+    # define context variables
+    action = np.zeros(num_actions, dtype=np.float32)
+    last_action = np.zeros(num_actions, dtype=np.float32)
+    target_dof_pos = default_angles.copy()
+    obs = np.zeros(num_obs, dtype=np.float32)
+
+    counter = 0
+
+    # Load robot model
+    m = mujoco.MjModel.from_xml_path(xml_path)
+    d = mujoco.MjData(m)
+    m.opt.timestep = simulation_dt
+
+    renderer = mujoco.Renderer(m, height=360, width=640)
+    
+    # load policy
+    policy = torch.jit.load(policy_path)
+
+    if save_video:
+        video_filename = f"{model_name}_{cmd_str}.mp4"
+        video_path = os.path.join(video_save_dir, video_filename)
+        print(f"Video recording will be saved to: {video_path}")
+        video_fps = 50
+        sim_fps = 1.0 / m.opt.timestep
+        frame_skip = int(sim_fps / video_fps)
+        if frame_skip < 1:
+            frame_skip = 1
+        writer = imageio.get_writer(video_path, fps=video_fps)
+        print(f"Sim FPS: {sim_fps:.2f}, Video FPS: {video_fps}, Frame Skip: {frame_skip}, Save at: {video_path}")
+
+    if visualize_moe_weights:
+        plt.ion()
+        fig, ax = plt.subplots(figsize=(5,3))
+        ax.set_title(f"Command: Vx={cmd[0]:.2f}, Vy={cmd[1]:.2f}, Wz={cmd[2]:.2f}")
+        bars = None
+    
+    if save_moe_latent:
+        latent_save_dir = str(Path(__file__).parent / "data_latents")
+        os.makedirs(latent_save_dir, exist_ok=True)
+        latent_filename = f"{model_name}_{cmd_str}_latents.npy"
+        latent_path = os.path.join(latent_save_dir, latent_filename)
+        all_latents = []
+
+    with mujoco.viewer.launch_passive(m, d) as viewer:
+
+        # set viewer.camera to follow robot
+        viewer.cam.type = mujoco.mjtCamera.mjCAMERA_TRACKING
+        viewer.cam.trackbodyid = 1
+        viewer.cam.distance = 3.0
+        viewer.cam.elevation = -30.0
+        viewer.cam.azimuth = 0.0
+
+        # Close the viewer automatically after simulation_duration wall-seconds.
+        start = time.time()
+        while viewer.is_running() and time.time() - start < simulation_duration:
+            step_start = time.time()
+
+            if use_joystick and counter % control_decimation == 0:
+                cmd = get_xbox_command(joystick, config["max_cmd"])
+                print(f"Cmd: Vx={cmd[0]:.2f}, Vy={cmd[1]:.2f}, Wz={cmd[2]:.2f}", end='\r')
+
+            tau = pd_control(target_dof_pos, d.qpos[7:], kps, np.zeros_like(kds), d.qvel[6:], kds)
+            d.ctrl[:] = tau
+            # mj_step can be replaced with code that also evaluates
+            # a policy and applies a control signal before stepping the physics.
+            mujoco.mj_step(m, d)
+
+            if save_video and counter % frame_skip == 0:
+                try:
+                    renderer.update_scene(d, camera=viewer.cam)
+                    frame = renderer.render()
+                    writer.append_data(frame)
+                except Exception as e:
+                    print(f"Error rendering frame: {e}")
+
+            counter += 1
+            if counter % control_decimation == 0:
+                # Apply control signal here.
+
+                # create observation
+                qj = d.qpos[7:]
+                dqj = d.qvel[6:]
+                quat = d.qpos[3:7]
+                lin_vel = d.qvel[:3]
+                ang_vel = d.qvel[3:6]
+
+                qj = (qj - default_angles) * dof_pos_scale
+
+                dqj = dqj * dof_vel_scale
+                gravity_orientation = get_gravity_orientation(quat)
+                lin_vel = lin_vel * lin_vel_scale
+                ang_vel = ang_vel * ang_vel_scale
+
+                obs[:3] = ang_vel
+                obs[3:6] = gravity_orientation
+                obs[6:9] = cmd * cmd_scale
+                obs[9 : 9 + num_actions] = qj[idx_mj2model]
+                obs[9 + num_actions : 9 + 2 * num_actions] = dqj[idx_mj2model]
+                obs[9 + 2 * num_actions : 9 + 3 * num_actions] = action[idx_mj2model]
+                obs_tensor = torch.from_numpy(obs).unsqueeze(0)
+                # policy inference
+                last_action = action
+                result = policy(obs_tensor)
+                if isinstance(result, tuple):
+                    action, (weights, latent) = result  # moe
+                    action = action.detach().numpy().squeeze()[idx_model2mj]
+                    weights = weights.detach().numpy().squeeze()
+                    latent = latent.detach().numpy().squeeze()
+                    if visualize_moe_weights:
+                        if bars is None:
+                            x = np.arange(len(weights))
+                            bars = ax.bar(x, weights)
+                            ax.set_ylim(0, 1)
+                        else:
+                            for bar, w in zip(bars, weights):
+                                bar.set_height(w)
+                        
+                        plt.draw()
+                        plt.pause(0.001) # 这会造成大约 1ms 的延迟
+                    if save_moe_latent:
+                        all_latents.append(latent)
+                else:
+                    action = result.detach().cpu().numpy().squeeze()[idx_model2mj]
+                # transform action to target_dof_pos
+                target_dof_pos = action * action_scale + default_angles
+
+            # Pick up changes to the physics state, apply perturbations, update options from GUI.
+            viewer.sync()
+
+            # Rudimentary time keeping, will drift relative to wall clock.
+            # time_until_next_step = m.opt.timestep - (time.time() - step_start) - 0.1
+            # if time_until_next_step > 0:
+            #     time.sleep(time_until_next_step)
+
+    # writer.close()
+    if save_video:
+        print(f"Video saved successfully to {video_path}")
+        writer.close()
+    if save_moe_latent and len(all_latents) > 0:
+        all_latents = np.array(all_latents)
+        np.save(latent_path, all_latents)
+        print(f"Latent vectors saved successfully to {latent_path}")

deploy/deploy_mujoco/deploy_go2_lab.py

@@ -0,0 +1,126 @@
+import time
+
+import mujoco.viewer
+import mujoco
+import numpy as np
+from legged_gym import LEGGED_GYM_ROOT_DIR
+import torch
+import yaml
+
+
+def get_gravity_orientation(quaternion):
+    qw = quaternion[0]
+    qx = quaternion[1]
+    qy = quaternion[2]
+    qz = quaternion[3]
+
+    gravity_orientation = np.zeros(3)
+
+    gravity_orientation[0] = 2 * (-qz * qx + qw * qy)
+    gravity_orientation[1] = -2 * (qz * qy + qw * qx)
+    gravity_orientation[2] = 1 - 2 * (qw * qw + qz * qz)
+
+    return gravity_orientation
+
+
+def pd_control(target_q, q, kp, target_dq, dq, kd):
+    """Calculates torques from position commands"""
+    return (target_q - q) * kp + (target_dq - dq) * kd
+
+
+if __name__ == "__main__":
+    # get config file name from command line
+    import argparse
+
+    parser = argparse.ArgumentParser()
+    parser.add_argument("config_file", type=str, help="config file name in the config folder")
+    args = parser.parse_args()
+    config_file = args.config_file
+    with open(f"{LEGGED_GYM_ROOT_DIR}/deploy/deploy_mujoco/configs/{config_file}", "r") as f:
+        config = yaml.load(f, Loader=yaml.FullLoader)
+        policy_path = config["policy_path"].replace("{LEGGED_GYM_ROOT_DIR}", LEGGED_GYM_ROOT_DIR)
+        xml_path = config["xml_path"].replace("{LEGGED_GYM_ROOT_DIR}", LEGGED_GYM_ROOT_DIR)
+
+        simulation_duration = config["simulation_duration"]
+        simulation_dt = config["simulation_dt"]
+        control_decimation = config["control_decimation"]
+
+        kps = np.array(config["kps"], dtype=np.float32)
+        kds = np.array(config["kds"], dtype=np.float32)
+
+        default_angles = np.array(config["default_angles"], dtype=np.float32)
+
+        joint_ids_map = config["joint_ids_map"]
+        ang_vel_scale = config["ang_vel_scale"]
+        dof_pos_scale = config["dof_pos_scale"]
+        dof_vel_scale = config["dof_vel_scale"]
+        action_scale = config["action_scale"]
+        cmd_scale = np.array(config["cmd_scale"], dtype=np.float32)
+
+        num_actions = config["num_actions"]
+        num_obs = config["num_obs"]
+
+        cmd = np.array(config["cmd_init"], dtype=np.float32)
+
+    # define context variables
+    action = np.zeros(num_actions, dtype=np.float32)
+    target_dof_pos = default_angles.copy()
+    obs = np.zeros(num_obs, dtype=np.float32)
+
+    counter = 0
+
+    # Load robot model
+    m = mujoco.MjModel.from_xml_path(xml_path)
+    d = mujoco.MjData(m)
+    m.opt.timestep = simulation_dt
+
+    # load policy
+    policy = torch.jit.load(policy_path)
+
+    with mujoco.viewer.launch_passive(m, d) as viewer:
+        # Close the viewer automatically after simulation_duration wall-seconds.
+        start = time.time()
+        while viewer.is_running() and time.time() - start < simulation_duration:
+            step_start = time.time()
+            temp = target_dof_pos[[0,4,8,1,5,9,2,6,10,3,7,11]]
+            tau = pd_control(temp, d.qpos[7:], kps, np.zeros_like(kds), d.qvel[6:], kds)
+            d.ctrl[:] = tau
+            # mj_step can be replaced with code that also evaluates
+            # a policy and applies a control signal before stepping the physics.
+            mujoco.mj_step(m, d)
+
+            counter += 1
+            if counter % control_decimation == 0:
+                # Apply control signal here.
+
+                # create observation
+                qj = d.qpos[7:]
+                dqj = d.qvel[6:]
+                quat = d.qpos[3:7]
+                ang_vel = d.qvel[3:6]
+
+                qj = (qj - default_angles) * dof_pos_scale
+
+                dqj = dqj * dof_vel_scale
+                gravity_orientation = get_gravity_orientation(quat)
+                ang_vel = ang_vel * ang_vel_scale
+
+                obs[:3] = ang_vel
+                obs[3:6] = gravity_orientation
+                obs[6:9] = cmd * cmd_scale
+                obs[9 : 9 + num_actions] = qj[joint_ids_map]
+                obs[9 + num_actions : 9 + 2 * num_actions] = dqj[joint_ids_map]
+                obs[9 + 2 * num_actions : 9 + 3 * num_actions] = action
+                obs_tensor = torch.from_numpy(obs).unsqueeze(0)
+                # policy inference
+                action = policy(obs_tensor).detach().numpy().squeeze()
+                # transform action to target_dof_pos
+                target_dof_pos = action * action_scale + default_angles[joint_ids_map]
+
+            # Pick up changes to the physics state, apply perturbations, update options from GUI.
+            viewer.sync()
+
+            # Rudimentary time keeping, will drift relative to wall clock.
+            time_until_next_step = m.opt.timestep - (time.time() - step_start)
+            if time_until_next_step > 0:
+                time.sleep(time_until_next_step)

deploy/deploy_mujoco/deploy_go2_moe.py

@@ -0,0 +1,272 @@
+import time
+import mujoco.viewer
+import mujoco
+import numpy as np
+from legged_gym import LEGGED_GYM_ROOT_DIR
+import torch
+import yaml
+import os
+import imageio
+from pathlib import Path
+from argparse import ArgumentParser
+import pygame
+# from matplotlib import pyplot as plt # 移除 matplotlib
+
+def get_gravity_orientation(quaternion):
+    qw = quaternion[0]
+    qx = quaternion[1]
+    qy = quaternion[2]
+    qz = quaternion[3]
+
+    gravity_orientation = np.zeros(3)
+
+    gravity_orientation[0] = 2 * (-qz * qx + qw * qy)
+    gravity_orientation[1] = -2 * (qz * qy + qw * qx)
+    gravity_orientation[2] = 1 - 2 * (qw * qw + qz * qz)
+
+    return gravity_orientation
+
+
+def pd_control(target_q, q, kp, target_dq, dq, kd):
+    """Calculates torques from position commands"""
+    return (target_q - q) * kp + (target_dq - dq) * kd
+
+def get_xbox_command(joystick, max_cmd):
+    # 注意：如果开启了 Pygame 显示窗口，这里 event.pump 也是必要的
+    pygame.event.pump()
+    dead_zone = 0.1
+    if joystick is not None:
+        lx = joystick.get_axis(0)
+        ly = joystick.get_axis(1)
+        rx = joystick.get_axis(3)
+        if abs(lx) < dead_zone: lx = 0
+        if abs(ly) < dead_zone: ly = 0
+        if abs(rx) < dead_zone: rx = 0
+        cmd_x = -ly * max_cmd[0]
+        cmd_y = -lx * max_cmd[1]
+        cmd_yaw = -rx * max_cmd[2]
+        return np.array([cmd_x, cmd_y, cmd_yaw], dtype=np.float32)
+    return np.zeros(3, dtype=np.float32)
+
+def draw_moe_weights(screen, weights, width, height):
+    """使用 Pygame 绘制 MoE 权重"""
+    screen.fill((255, 255, 255)) # 白底
+    
+    num_experts = len(weights)
+    if num_experts == 0:
+        return
+
+    # 设置边距
+    margin = 5
+    bar_width = (width - 2 * margin) / num_experts
+    max_bar_height = height - 2 * margin
+
+    for i, w in enumerate(weights):
+        # 限制 w 在 [0, 1] 之间用于显示
+        w_clamped = max(0.0, min(1.0, w))
+        bar_height = int(w_clamped * max_bar_height)
+        
+        # 计算矩形位置 (Pygame 坐标原点在左上角)
+        # left, top, width, height
+        x = margin + i * bar_width
+        y = height - margin - bar_height # 从底部向上长
+        
+        # 绘制矩形 (蓝色)
+        # 在 bar 之间留一点空隙 (width - 2)
+        pygame.draw.rect(screen, (50, 100, 255), (x, y, bar_width - 2, bar_height))
+        
+    pygame.display.flip()
+
+if __name__ == "__main__":
+    parser = ArgumentParser()
+    parser.add_argument("--save-video", action="store_true", help="Whether to save video of the simulation.")
+    parser.add_argument("--visualize-moe-weights", action="store_true", help="Whether to visualize mixture of experts weights.")
+    args = parser.parse_args()
+    save_video = args.save_video
+    visualize_moe_weights = args.visualize_moe_weights
+    config_file = "go2.yaml"
+
+    # Pygame 初始化
+    pygame.init()
+    
+    use_joystick = False
+    joystick = None
+    if pygame.joystick.get_count() > 0:
+        joystick = pygame.joystick.Joystick(0)
+        joystick.init()
+        use_joystick = True
+        print(f"Detected Joystick: {joystick.get_name()}")
+    else:
+        print("No Joystick detected. Using default commands from config.")
+
+    # 如果需要可视化权重，设置 Pygame 窗口
+    screen = None
+    win_width, win_height = 400, 200
+    if visualize_moe_weights:
+        # 创建一个独立的窗口用于显示权重
+        screen = pygame.display.set_mode((win_width, win_height))
+        pygame.display.set_caption("MoE Weights Visualization")
+
+    with open(f"{LEGGED_GYM_ROOT_DIR}/deploy/deploy_mujoco/configs/{config_file}", "r") as f:
+        config = yaml.load(f, Loader=yaml.FullLoader)
+        policy_path = config["policy_path"].replace("{LEGGED_GYM_ROOT_DIR}", LEGGED_GYM_ROOT_DIR)
+        xml_path = config["xml_path"].replace("{LEGGED_GYM_ROOT_DIR}", LEGGED_GYM_ROOT_DIR)
+
+        simulation_duration = config["simulation_duration"]
+        simulation_dt = config["simulation_dt"]
+        control_decimation = config["control_decimation"]
+
+        kps = np.array(config["kps"], dtype=np.float32)
+        kds = np.array(config["kds"], dtype=np.float32)
+
+        default_angles = np.array(config["default_angles"], dtype=np.float32)
+
+        lin_vel_scale = config["lin_vel_scale"]
+        ang_vel_scale = config["ang_vel_scale"]
+        dof_pos_scale = config["dof_pos_scale"]
+        dof_vel_scale = config["dof_vel_scale"]
+        action_scale = config["action_scale"]
+        cmd_scale = np.array(config["cmd_scale"], dtype=np.float32)
+
+        num_actions = config["num_actions"]
+        num_obs = config["num_obs"]
+
+        cmd = np.array(config["cmd_init"], dtype=np.float32)
+
+        idx_model2mj = idx_mj2model = list(range(num_actions))
+        if 'mujoco_joint_names' in config and 'model_joint_names' in config:
+            mujoco_joint_names = config["mujoco_joint_names"]
+            model_joint_names = config["model_joint_names"]
+            idx_model2mj = [model_joint_names.index(joint) for joint in mujoco_joint_names]
+            idx_mj2model = [mujoco_joint_names.index(joint) for joint in model_joint_names]
+
+    video_save_dir = str(Path(__file__).parent / "videos")
+    os.makedirs(video_save_dir, exist_ok=True)
+
+    model_name = os.path.basename(policy_path).split('.')[0]
+    cmd_str = f"cmd_{cmd[0]}_{cmd[1]}_{cmd[2]}"
+    video_filename = f"{model_name}_{cmd_str}.mp4"
+    video_path = os.path.join(video_save_dir, video_filename)
+    print(f"Video recording will be saved to: {video_path}")
+
+    # define context variables
+    action = np.zeros(num_actions, dtype=np.float32)
+    last_action = np.zeros(num_actions, dtype=np.float32)
+    target_dof_pos = default_angles.copy()
+    obs = np.zeros(num_obs, dtype=np.float32)
+
+    counter = 0
+
+    # Load robot model
+    m = mujoco.MjModel.from_xml_path(xml_path)
+    d = mujoco.MjData(m)
+    m.opt.timestep = simulation_dt
+
+    renderer = mujoco.Renderer(m, height=360, width=640)
+    
+    # load policy
+    policy = torch.jit.load(policy_path)
+
+    if save_video:
+        video_fps = 50
+        sim_fps = 1.0 / m.opt.timestep
+        frame_skip = int(sim_fps / video_fps)
+        if frame_skip < 1:
+            frame_skip = 1
+        writer = imageio.get_writer(video_path, fps=video_fps)
+        print(f"Sim FPS: {sim_fps:.2f}, Video FPS: {video_fps}, Frame Skip: {frame_skip}, Save at: {video_path}")
+
+    # 移除了 plt 初始化逻辑
+
+    with mujoco.viewer.launch_passive(m, d) as viewer:
+
+        # set viewer.camera to follow robot
+        viewer.cam.type = mujoco.mjtCamera.mjCAMERA_TRACKING
+        viewer.cam.trackbodyid = 1
+        viewer.cam.distance = 3.0
+        viewer.cam.elevation = -30.0
+        viewer.cam.azimuth = 0.0
+
+        # Close the viewer automatically after simulation_duration wall-seconds.
+        start = time.time()
+        while viewer.is_running() and time.time() - start < simulation_duration:
+            step_start = time.time()
+
+            if use_joystick and counter % control_decimation == 0:
+                cmd = get_xbox_command(joystick, config["max_cmd"])
+                print(f"Cmd: Vx={cmd[0]:.2f}, Vy={cmd[1]:.2f}, Wz={cmd[2]:.2f}", end='\r')
+            elif visualize_moe_weights and counter % control_decimation == 0:
+                 # 如果没有手柄但开了可视化，也需要 pump 事件，防止窗口卡死
+                 pygame.event.pump()
+
+            tau = pd_control(target_dof_pos, d.qpos[7:], kps, np.zeros_like(kds), d.qvel[6:], kds)
+            d.ctrl[:] = tau
+            
+            mujoco.mj_step(m, d)
+
+            if save_video and counter % frame_skip == 0:
+                try:
+                    renderer.update_scene(d, camera=viewer.cam)
+                    frame = renderer.render()
+                    writer.append_data(frame)
+                except Exception as e:
+                    print(f"Error rendering frame: {e}")
+
+            counter += 1
+            if counter % control_decimation == 0:
+                # Apply control signal here.
+
+                # create observation
+                qj = d.qpos[7:]
+                dqj = d.qvel[6:]
+                quat = d.qpos[3:7]
+                lin_vel = d.qvel[:3]
+                ang_vel = d.qvel[3:6]
+
+                qj = (qj - default_angles) * dof_pos_scale
+
+                dqj = dqj * dof_vel_scale
+                gravity_orientation = get_gravity_orientation(quat)
+                lin_vel = lin_vel * lin_vel_scale
+                ang_vel = ang_vel * ang_vel_scale
+
+                obs[:3] = ang_vel
+                obs[3:6] = gravity_orientation
+                obs[6:9] = cmd * cmd_scale
+                obs[9 : 9 + num_actions] = qj[idx_mj2model]
+                obs[9 + num_actions : 9 + 2 * num_actions] = dqj[idx_mj2model]
+                obs[9 + 2 * num_actions : 9 + 3 * num_actions] = action[idx_mj2model]
+                obs_tensor = torch.from_numpy(obs).unsqueeze(0)
+                # policy inference
+                last_action = action
+                result = policy(obs_tensor)
+                
+                # 处理 MoE 和 绘图
+                if isinstance(result, tuple):
+                    action, weights = result  # moe
+                    action = action.detach().numpy().squeeze()[idx_model2mj]
+                    weights = weights.detach().numpy().squeeze()
+                    
+                    if visualize_moe_weights and screen is not None:
+                        draw_moe_weights(screen, weights, win_width, win_height)
+                        
+                else:
+                    action = result.detach().numpy().squeeze()[idx_model2mj]
+                    
+                # transform action to target_dof_pos
+                target_dof_pos = action * action_scale + default_angles
+
+            # Pick up changes to the physics state, apply perturbations, update options from GUI.
+            viewer.sync()
+            
+            # 如果需要严格同步时间，可以解开下面的注释
+            # time_until_next_step = m.opt.timestep - (time.time() - step_start)
+            # if time_until_next_step > 0:
+            #     time.sleep(time_until_next_step)
+
+    if save_video:
+        writer.close()
+    
+    # 退出时清理 Pygame
+    pygame.quit()
+    print(f"Video saved successfully to {video_path}")