diff --git a/src/Neuro_Mujoco/main.py b/src/Neuro_Mujoco/main.py index 890b1ee711..d3f8d0f536 100644 --- a/src/Neuro_Mujoco/main.py +++ b/src/Neuro_Mujoco/main.py @@ -217,6 +217,7 @@ def visualize_model(model_path: str, use_ros_mode: bool = False, policy_model_pa obs_dimension = model.nq + model.nv # 观测维度 = 关节位置数 + 关节速度数 action_dimension = model.nu control_range = None # 执行器控制范围(min, max) + obs_tensor = None # 观测张量(提前初始化) if policy_model_path and TORCH_AVAILABLE and action_dimension > 0: try: @@ -234,6 +235,9 @@ def visualize_model(model_path: str, use_ros_mode: bool = False, policy_model_pa else: control_range.append((-1.0, 1.0)) control_range = np.array(control_range) + + # 【优化点1】提前初始化固定形状的观测张量,避免循环内重复创建 + obs_tensor = torch.zeros(1, obs_dimension, dtype=torch.float32, device="cpu") except Exception as e: logger.error(f"策略模型加载失败:{str(e)}", exc_info=True) policy_net = None @@ -302,10 +306,7 @@ def control_cmd_callback(msg: Float32MultiArray): # ===================== 可视化主循环 ===================== logger.info("启动MuJoCo可视化窗口 | 退出:ESC键 | 交互:鼠标拖拽(旋转)/滚轮(缩放)") try: - with viewer.launch_passive(model, data) as viewer_instance: - # 预分配观测张量(复用避免重复内存分配) - obs_tensor = torch.zeros(1, obs_dimension, dtype=torch.float32) if policy_net else None - + with viewer.launch_passive(model, data) as viewer_instance: while viewer_instance.is_running() and (not use_ros_mode or not rospy.is_shutdown()): # 【核心优先级规则】控制指令优先级:ROS外部指令 > 策略网络推理 > 无控制输入 if use_ros_mode and ros_ctrl_cmd is not None: @@ -314,16 +315,11 @@ def control_cmd_callback(msg: Float32MultiArray): # 提取观测向量:关节位置 + 关节速度 observation = np.concatenate([data.qpos, data.qvel]) - # 复用张量,避免重复创建 - if obs_tensor is None: - obs_tensor = torch.tensor(observation, dtype=torch.float32).unsqueeze(0) - else: - obs_tensor[0] = torch.from_numpy(observation) + obs_tensor[0] = torch.from_numpy(observation).to(obs_tensor.dtype) # 策略推理得到归一化动作([-1,1]) normalized_action = policy_net(obs_tensor).squeeze().numpy() - # 【关键安全机制】线性映射归一化动作到执行器物理极限,并强制裁剪防止超限损坏模拟 if control_range is not None: # 线性映射:[-1,1] → [ctrl_min, ctrl_max] normalized_action = control_range[:, 0] + (control_range[:, 1] - control_range[:, 0]) * (normalized_action + 1) / 2