{"id":7653,"date":"2025-04-06T13:15:00","date_gmt":"2025-04-06T05:15:00","guid":{"rendered":"https:\/\/meta-quantum.today\/?p=7653"},"modified":"2025-04-06T13:42:01","modified_gmt":"2025-04-06T05:42:01","slug":"%e6%9c%ba%e5%99%a8%e4%ba%ba%e5%a6%82%e4%bd%95%e8%8e%b7%e5%be%97%e8%ae%ad%e7%bb%83%e6%95%b0%e6%8d%ae-pieter-abbeel-gtc%e6%9c%80%e6%96%b0%e6%bc%94%e8%ae%b2%e8%a7%a3%e6%9e%90-%e4%b8%9c%e5%8d%97","status":"publish","type":"post","link":"https:\/\/meta-quantum.today\/?p=7653","title":{"rendered":"\u673a\u5668\u4eba\u5982\u4f55\u83b7\u5f97\u8bad\u7ec3\u6570\u636e | Pieter Abbeel GTC\u6700\u65b0\u6f14\u8bb2\u89e3\u6790 (\u4e1c\u5357\u4e9a\u53ef\u4ee5\u5982\u4f55\u53d7\u76ca)"},"content":{"rendered":"\n

\u4ecb\u7ecd<\/h2>\n\n\n\n

\u5f7c\u5f97\u00b7\u963f\u6bd4\u5c14(Pieter Abbeel)\u662f\u52a0\u5dde\u5927\u5b66\u4f2f\u514b\u5229\u5206\u6821\u7535\u6c14\u5de5\u7a0b\u4e0e\u8ba1\u7b97\u673a\u79d1\u5b66\u7cfb\u6559\u6388\uff0c\u540c\u65f6\u62c5\u4efb\u4f2f\u514b\u5229\u673a\u5668\u4eba\u5b66\u4e60\u5b9e\u9a8c\u5ba4\u4e3b\u4efb\u548c\u4f2f\u514b\u5229\u4eba\u5de5\u667a\u80fd\u7814\u7a76\u5b9e\u9a8c\u5ba4(BAIR)\u8054\u5408\u4e3b\u4efb\u3002\u4ed6\u662f\u673a\u5668\u5b66\u4e60\u9886\u57df\u7684\u6743\u5a01\u4e13\u5bb6\uff0c\u57f9\u517b\u4e86\u591a\u4f4d\u77e5\u540dAI\u4f01\u4e1a\u7684\u521b\u59cb\u4eba\uff0c\u5305\u62ecOpenAI\u521b\u59cb\u56e2\u961f\u6210\u5458\u7ea6\u7ff0\u00b7\u8212\u5c14\u66fc\u3001Perplexity\u7684CEO\u963f\u62c9\u6e29\u5fb7\u00b7\u65af\u91cc\u5c3c\u74e6\u65af\u7b49\u3002\u5728GTC 2025\u4e0a\uff0c\u963f\u6bd4\u5c14\u8fdb\u884c\u4e86\u4e00\u573a\u5173\u4e8e”\u673a\u5668\u4eba\u8bad\u7ec3\u6570\u636e”\u7684\u4e3b\u9898\u6f14\u8bb2\uff0c\u63a2\u8ba8\u5982\u4f55\u89e3\u51b3\u4eba\u5f62\u673a\u5668\u4eba\u7684\u6570\u636e\u56f0\u5883\u95ee\u9898\u3002<\u673a\u5668\u4eba\u5982\u4f55\u83b7\u5f97\u8bad\u7ec3\u6570\u636e\u7684\u89c6\u9891<\/a>><\/p>\n\n\n\n

\u673a\u5668\u4eba\u5982\u4f55\u83b7\u5f97\u8bad\u7ec3\u6570\u636e\uff1a\u65b9\u6cd5\u3001\u793a\u4f8b\u4e0e\u4ee3\u7801\u5b9e\u73b0<\/h2>\n\n\n\n

\u6570\u636e\u83b7\u53d6\u65b9\u6cd5\u6982\u8ff0<\/h3>\n\n\n\n

\u6839\u636e\u5f7c\u5f97\u00b7\u963f\u6bd4\u5c14(Pieter Abbeel)\u7684\u7814\u7a76\uff0c\u673a\u5668\u4eba\u8bad\u7ec3\u6570\u636e\u4e3b\u8981\u901a\u8fc7\u4ee5\u4e0b\u51e0\u79cd\u65b9\u5f0f\u83b7\u53d6\uff1a<\/p>\n\n\n\n

    \n
  1. \u8fdc\u7a0b\u64cd\u4f5c(Teleoperation)<\/strong> – \u901a\u8fc7\u4eba\u7c7b\u76f4\u63a5\u63a7\u5236\u673a\u5668\u4eba\u6536\u96c6\u52a8\u4f5c\u6570\u636e<\/li>\n\n\n\n
  2. \u624b\u90e8\u52a8\u4f5c\u8ddf\u8e2a<\/strong> – \u8ddf\u8e2a\u4eba\u7c7b\u624b\u90e8\u52a8\u4f5c\u5e76\u8f6c\u5316\u4e3a\u673a\u5668\u4eba\u6307\u4ee4<\/li>\n\n\n\n
  3. \u4eff\u771f\u73af\u5883<\/strong> – \u5728\u865a\u62df\u73af\u5883\u4e2d\u751f\u6210\u5927\u91cf\u8bad\u7ec3\u6570\u636e<\/li>\n\n\n\n
  4. \u73b0\u5b9e\u4e16\u754c\u5b66\u4e60<\/strong> – \u673a\u5668\u4eba\u5728\u771f\u5b9e\u73af\u5883\u4e2d\u901a\u8fc7\u8bd5\u9519\u5b66\u4e60<\/li>\n\n\n\n
  5. \u6a21\u4eff\u5b66\u4e60<\/strong> – \u89c2\u5bdf\u5e76\u6a21\u4eff\u4eba\u7c7b\u884c\u4e3a<\/li>\n<\/ol>\n\n\n\n

    \u4e0b\u9762\u4e3a\u6bcf\u79cd\u65b9\u6cd5\u63d0\u4f9b\u5177\u4f53\u793a\u4f8b\u548c\u76f8\u5173\u4ee3\u7801\u5b9e\u73b0:<\/p>\n\n\n\n

    1. \u8fdc\u7a0b\u64cd\u4f5c\u6570\u636e\u6536\u96c6<\/h3>\n\n\n\n

    \u8fdc\u7a0b\u64cd\u4f5c\u662f\u83b7\u53d6\u9ad8\u8d28\u91cf\u673a\u5668\u4eba\u8bad\u7ec3\u6570\u636e\u7684\u76f4\u63a5\u65b9\u6cd5\u3002\u4f8b\u5982\uff0cPhysical Intelligence (PI)\u516c\u53f8\u901a\u8fc7\u8fdc\u7a0b\u64cd\u4f5c\u7cfb\u7edf\u6536\u96c6\u4e86\u673a\u5668\u4eba\u6574\u7406\u8863\u7269\u7684\u6570\u636e\u3002<\/p>\n\n\n\n

    \u793a\u4f8b\u4ee3\u7801 – \u8fdc\u7a0b\u64cd\u4f5c\u6570\u636e\u6536\u96c6<\/h3>\n\n\n\n
    import numpy as np\nimport time\nimport json\nimport robotlib  # \u5047\u8bbe\u7684\u673a\u5668\u4eba\u63a7\u5236\u5e93\n\nclass TeleoperationDataCollector:\n    def __init__(self, robot, save_path=\"teleop_data.json\"):\n        self.robot = robot\n        self.save_path = save_path\n        self.data = []\n\n    def record_frame(self, human_input, robot_state):\n        \"\"\"\u8bb0\u5f55\u4e00\u5e27\u8fdc\u7a0b\u64cd\u4f5c\u6570\u636e\"\"\"\n        frame = {\n            \"timestamp\": time.time(),\n            \"joint_angles\": robot_state.joint_angles.tolist(),\n            \"gripper_state\": robot_state.gripper_state,\n            \"human_command\": human_input.tolist(),\n            \"camera_image\": robot_state.get_camera_image_path()\n        }\n        self.data.append(frame)\n\n    def save_data(self):\n        \"\"\"\u4fdd\u5b58\u6536\u96c6\u7684\u6570\u636e\"\"\"\n        with open(self.save_path, 'w') as f:\n            json.dump(self.data, f)\n        print(f\"\u4fdd\u5b58\u4e86 {len(self.data)} \u5e27\u8fdc\u7a0b\u64cd\u4f5c\u6570\u636e\u5230 {self.save_path}\")\n\n# \u4f7f\u7528\u793a\u4f8b\ndef collect_folding_data():\n    robot = robotlib.connect_robot(\"192.168.1.100\")  # \u8fde\u63a5\u5230\u673a\u5668\u4eba\n    collector = TeleoperationDataCollector(robot)\n\n    print(\"\u5f00\u59cb\u6570\u636e\u6536\u96c6\uff0c\u6309 Ctrl+C \u505c\u6b62...\")\n    try:\n        while True:\n            human_input = get_human_controller_input()  # \u83b7\u53d6\u64cd\u4f5c\u8005\u8f93\u5165\n            robot.execute_command(human_input)  # \u6267\u884c\u547d\u4ee4\n            robot_state = robot.get_state()  # \u83b7\u53d6\u673a\u5668\u4eba\u72b6\u6001\n            collector.record_frame(human_input, robot_state)\n            time.sleep(0.1)  # 10Hz \u91c7\u6837\u7387\n    except KeyboardInterrupt:\n        collector.save_data()\n        print(\"\u6570\u636e\u6536\u96c6\u5b8c\u6210\")\n\n<\/code><\/pre>\n\n\n\n
    2. \u4eff\u771f\u73af\u5883\u6570\u636e\u751f\u6210<\/h3>\n\n\n\n
    \u4f7f\u7528\u7269\u7406\u4eff\u771f\u5668(\u5982MuJoCo)\u53ef\u4ee5\u751f\u6210\u5927\u91cf\u7684\u8bad\u7ec3\u6570\u636e\uff0c\u8fd9\u4e9b\u6570\u636e\u53ef\u4ee5\u7528\u4e8e\u9884\u8bad\u7ec3\u673a\u5668\u4eba\u63a7\u5236\u6a21\u578b\u3002<\/p>\n\n\n\n
    \u793a\u4f8b\u4ee3\u7801 – MuJoCo\u4eff\u771f\u6570\u636e\u751f\u6210<\/h3>\n\n\n\n
    import mujoco\nimport numpy as np\nimport time\n\nclass RobotSimulator:\n    def __init__(self, model_path, task_type=\"walking\"):\n        self.model = mujoco.MjModel.from_xml_path(model_path)\n        self.data = mujoco.MjData(self.model)\n        self.task_type = task_type\n        self.collected_data = []\n\n    def reset(self):\n        \"\"\"\u91cd\u7f6e\u4eff\u771f\u73af\u5883\"\"\"\n        mujoco.mj_resetData(self.model, self.data)\n        # \u968f\u673a\u521d\u59cb\u72b6\u6001\u4ee5\u589e\u52a0\u591a\u6837\u6027\n        self.data.qpos[:] = self.data.qpos[:] + np.random.uniform(-0.1, 0.1, size=self.model.nq)\n        mujoco.mj_forward(self.model, self.data)\n        return self._get_observation()\n\n    def _get_observation(self):\n        \"\"\"\u83b7\u53d6\u89c2\u5bdf\u72b6\u6001\"\"\"\n        qpos = self.data.qpos.copy()\n        qvel = self.data.qvel.copy()\n        # \u53ef\u4ee5\u6839\u636e\u9700\u8981\u6dfb\u52a0\u66f4\u591a\u4f20\u611f\u5668\u6570\u636e\n        return {\"qpos\": qpos, \"qvel\": qvel}\n\n    def step(self, action):\n        \"\"\"\u6267\u884c\u4e00\u6b65\u4eff\u771f\"\"\"\n        self.data.ctrl[:] = action\n        for _ in range(10):  # 10\u4e2a\u7269\u7406\u4eff\u771f\u6b65\u9aa4\n            mujoco.mj_step(self.model, self.data)\n\n        obs = self._get_observation()\n        reward = self._compute_reward()\n        done = self._check_termination()\n\n        # \u8bb0\u5f55\u6570\u636e\n        frame_data = {\n            \"observation\": obs,\n            \"action\": action.copy(),\n            \"reward\": reward,\n            \"done\": done\n        }\n        self.collected_data.append(frame_data)\n\n        return obs, reward, done, {}\n\n    def _compute_reward(self):\n        \"\"\"\u8ba1\u7b97\u5956\u52b1\u51fd\u6570\uff0c\u53d6\u51b3\u4e8e\u4efb\u52a1\u7c7b\u578b\"\"\"\n        if self.task_type == \"walking\":\n            # \u5bf9\u4e8e\u884c\u8d70\u4efb\u52a1\uff0c\u5956\u52b1\u524d\u8fdb\u8ddd\u79bb\u5e76\u60e9\u7f5a\u80fd\u91cf\u6d88\u8017\n            forward_reward = self.data.qpos[0]  # x\u5750\u6807\u4f4d\u7f6e\n            energy_penalty = np.sum(np.square(self.data.ctrl))\n            return forward_reward - 0.1 * energy_penalty\n        # \u53ef\u4ee5\u6dfb\u52a0\u5176\u4ed6\u4efb\u52a1\u7c7b\u578b\u7684\u5956\u52b1\u51fd\u6570\n        return 0.0\n\n    def _check_termination(self):\n        \"\"\"\u68c0\u67e5\u662f\u5426\u5e94\u7ec8\u6b62\u5f53\u524d\u56de\u5408\"\"\"\n        # \u4f8b\u5982\uff0c\u5982\u679c\u673a\u5668\u4eba\u6454\u5012\uff0c\u5219\u7ec8\u6b62\n        height = self.data.qpos[2]  # z\u5750\u6807\u9ad8\u5ea6\n        return height < 0.7  # \u5047\u8bbe\u6b63\u5e38\u9ad8\u5ea6\u5e94\u5927\u4e8e0.7\n\n    def save_collected_data(self, filename):\n        \"\"\"\u4fdd\u5b58\u6536\u96c6\u7684\u6570\u636e\"\"\"\n        np.save(filename, self.collected_data)\n        print(f\"\u4fdd\u5b58\u4e86 {len(self.collected_data)} \u5e27\u4eff\u771f\u6570\u636e\u5230 {filename}\")\n\n# \u4f7f\u7528\u793a\u4f8b\ndef generate_walking_data(episodes=100):\n    sim = RobotSimulator(\"humanoid.xml\", task_type=\"walking\")\n\n    for episode in range(episodes):\n        obs = sim.reset()\n        done = False\n\n        while not done:\n            # \u4f7f\u7528\u968f\u673a\u7b56\u7565\u6216\u9884\u5b9a\u4e49\u7684\u63a7\u5236\u5668\u751f\u6210\u52a8\u4f5c\n            action = np.random.uniform(-1, 1, size=sim.model.nu)\n            obs, reward, done, _ = sim.step(action)\n\n        print(f\"\u5b8c\u6210\u7b2c {episode+1}\/{episodes} \u56de\u5408\")\n\n    sim.save_collected_data(\"walking_simulation_data.npy\")\n\n<\/code><\/pre>\n\n\n\n
    3. \u57fa\u4e8e\u89c6\u89c9\u7684\u624b\u90e8\u52a8\u4f5c\u8ddf\u8e2a<\/h3>\n\n\n\n
    \u5982\u52a0\u5dde\u5927\u5b66\u5723\u5730\u4e9a\u54e5\u5206\u6821\u7684\u738b\u5c0f\u9f99\u6559\u6388\u56e2\u961f\u6240\u793a\uff0c\u901a\u8fc7\u89c6\u89c9\u8bbe\u5907\u8ddf\u8e2a\u4eba\u7c7b\u624b\u90e8\u52a8\u4f5c\u53ef\u4ee5\u4e3a\u673a\u5668\u4eba\u63d0\u4f9b\u64cd\u4f5c\u6307\u5bfc\u3002<\/p>\n\n\n\n
    \u793a\u4f8b\u4ee3\u7801 – \u624b\u90e8\u52a8\u4f5c\u8ddf\u8e2a<\/h3>\n\n\n\n
    import cv2\nimport mediapipe as mp\nimport numpy as np\n\nclass HandTracker:\n    def __init__(self):\n        self.mp_hands = mp.solutions.hands\n        self.hands = self.mp_hands.Hands(\n            static_image_mode=False,\n            max_num_hands=2,\n            min_detection_confidence=0.5,\n            min_tracking_confidence=0.5\n        )\n        self.mp_drawing = mp.solutions.drawing_utils\n\n    def process_frame(self, frame):\n        \"\"\"\u5904\u7406\u4e00\u5e27\u56fe\u50cf\u5e76\u63d0\u53d6\u624b\u90e8\u5173\u952e\u70b9\"\"\"\n        # \u8f6c\u6362\u4e3aRGB\u683c\u5f0f\n        frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)\n\n        # \u5904\u7406\u56fe\u50cf\n        results = self.hands.process(frame_rgb)\n\n        hand_landmarks = []\n        if results.multi_hand_landmarks:\n            for hand_landmarks_data in results.multi_hand_landmarks:\n                # \u7ed8\u5236\u624b\u90e8\u5173\u952e\u70b9\n                self.mp_drawing.draw_landmarks(\n                    frame, hand_landmarks_data, self.mp_hands.HAND_CONNECTIONS)\n\n                # \u63d0\u53d6\u5173\u952e\u70b9\u5750\u6807\n                landmarks = []\n                for landmark in hand_landmarks_data.landmark:\n                    landmarks.append([landmark.x, landmark.y, landmark.z])\n                hand_landmarks.append(np.array(landmarks))\n\n        return frame, hand_landmarks\n\n    def convert_to_robot_commands(self, hand_landmarks):\n        \"\"\"\u5c06\u624b\u90e8\u5173\u952e\u70b9\u8f6c\u6362\u4e3a\u673a\u5668\u4eba\u6307\u4ee4\"\"\"\n        if not hand_landmarks:\n            return None\n\n        # \u793a\u4f8b\uff1a\u4f7f\u7528\u7b2c\u4e00\u53ea\u624b\u7684\u5173\u952e\u70b9\n        landmarks = hand_landmarks[0]\n\n        # \u8ba1\u7b97\u6307\u5c16\u4f4d\u7f6e (\u62c7\u6307\u3001\u98df\u6307\u3001\u4e2d\u6307)\n        thumb_tip = landmarks[4]\n        index_tip = landmarks[8]\n        middle_tip = landmarks[12]\n\n        # \u8ba1\u7b97\u62c7\u6307\u548c\u98df\u6307\u95f4\u8ddd\uff0c\u7528\u4e8e\u63a7\u5236\u5939\u722a\n        pinch_distance = np.linalg.norm(thumb_tip - index_tip)\n        gripper_command = 1.0 if pinch_distance < 0.1 else 0.0\n\n        # \u624b\u638c\u4e2d\u5fc3\u4f5c\u4e3a\u4f4d\u7f6e\u63a7\u5236\n        palm_center = np.mean(landmarks[[0, 1, 5, 9, 13, 17]], axis=0)\n\n        # \u8f6c\u6362\u4e3a\u673a\u5668\u4eba\u5750\u6807\u7cfb (\u793a\u4f8b\u8f6c\u6362)\n        robot_pos = palm_center * np.array([2.0, 2.0, 2.0]) - np.array([1.0, 1.0, 0.0])\n\n        return {\n            \"position\": robot_pos,\n            \"gripper\": gripper_command\n        }\n\n# \u4f7f\u7528\u793a\u4f8b\ndef hand_tracking_demo():\n    cap = cv2.VideoCapture(0)\n    tracker = HandTracker()\n\n    collected_data = []\n\n    while cap.isOpened():\n        success, frame = cap.read()\n        if not success:\n            break\n\n        frame, hand_landmarks = tracker.process_frame(frame)\n        robot_command = tracker.convert_to_robot_commands(hand_landmarks)\n\n        if robot_command:\n            # \u8bb0\u5f55\u6570\u636e\n            data_point = {\n                \"timestamp\": time.time(),\n                \"hand_landmarks\": hand_landmarks[0].tolist() if hand_landmarks else None,\n                \"robot_command\": robot_command\n            }\n            collected_data.append(data_point)\n\n            # \u663e\u793a\u673a\u5668\u4eba\u547d\u4ee4\n            cv2.putText(frame, f\"Gripper: {robot_command['gripper']:.1f}\", (10, 30),\n                        cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)\n\n        cv2.imshow('Hand Tracking', frame)\n        if cv2.waitKey(5) & 0xFF == 27:  # ESC\u952e\u9000\u51fa\n            break\n\n    cap.release()\n    cv2.destroyAllWindows()\n\n    # \u4fdd\u5b58\u6536\u96c6\u7684\u6570\u636e\n    np.save(\"hand_tracking_data.npy\", collected_data)\n    print(f\"\u4fdd\u5b58\u4e86 {len(collected_data)} \u5e27\u624b\u90e8\u8ddf\u8e2a\u6570\u636e\")\n\n<\/code><\/pre>\n\n\n\n
    4. \u57fa\u4e8eBody Transformer\u7684\u6a21\u578b\u8bad\u7ec3<\/h3>\n\n\n\n
    \u963f\u6bd4\u5c14\u4ecb\u7ecd\u7684Body Transformer\u4f7f\u7528\u5c40\u90e8\u8fde\u63a5\u7684\u67b6\u6784\uff0c\u63d0\u9ad8\u4e86\u6a21\u4eff\u5b66\u4e60\u7684\u6548\u7387\u548c\u53ef\u6269\u5c55\u6027\u3002<\/p>\n\n\n\n
    \u793a\u4f8b\u4ee3\u7801 – Body Transformer\u6a21\u578b<\/h3>\n\n\n\n
    import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass MaskedSelfAttention(nn.Module):\n    \"\"\"\u5e26\u6709\u5c40\u90e8\u8fde\u63a5\u63a9\u7801\u7684\u81ea\u6ce8\u610f\u529b\u673a\u5236\"\"\"\n    def __init__(self, embed_dim, num_heads, robot_structure_matrix):\n        super().__init__()\n        self.embed_dim = embed_dim\n        self.num_heads = num_heads\n        self.head_dim = embed_dim \/\/ num_heads\n\n        # \u673a\u5668\u4eba\u7ed3\u6784\u77e9\u9635\u5b9a\u4e49\u54ea\u4e9b\u5173\u8282\u53ef\u4ee5\u76f8\u4e92\u5173\u6ce8\n        # \u5f62\u72b6\u4e3a [num_joints, num_joints]\uff0c1\u8868\u793a\u5141\u8bb8\u5173\u6ce8\uff0c0\u8868\u793a\u7981\u6b62\n        self.register_buffer(\"robot_structure_mask\", robot_structure_matrix.unsqueeze(0).repeat(num_heads, 1, 1))\n\n        self.q_proj = nn.Linear(embed_dim, embed_dim)\n        self.k_proj = nn.Linear(embed_dim, embed_dim)\n        self.v_proj = nn.Linear(embed_dim, embed_dim)\n        self.out_proj = nn.Linear(embed_dim, embed_dim)\n\n    def forward(self, x):\n        batch_size, seq_len, _ = x.shape\n\n        # \u6295\u5f71\u67e5\u8be2\u3001\u952e\u3001\u503c\n        q = self.q_proj(x).reshape(batch_size, seq_len, self.num_heads, self.head_dim).permute(0, 2, 1, 3)\n        k = self.k_proj(x).reshape(batch_size, seq_len, self.num_heads, self.head_dim).permute(0, 2, 1, 3)\n        v = self.v_proj(x).reshape(batch_size, seq_len, self.num_heads, self.head_dim).permute(0, 2, 1, 3)\n\n        # \u8ba1\u7b97\u6ce8\u610f\u529b\u5206\u6570\n        attention = torch.matmul(q, k.transpose(-1, -2)) \/ (self.head_dim ** 0.5)\n\n        # \u5e94\u7528\u673a\u5668\u4eba\u7ed3\u6784\u63a9\u7801\n        attention = attention.masked_fill(self.robot_structure_mask[:, :seq_len, :seq_len] == 0, float('-inf'))\n\n        # \u5e94\u7528softmax\u5e76\u4e0e\u503c\u76f8\u4e58\n        attention = F.softmax(attention, dim=-1)\n        out = torch.matmul(attention, v)\n\n        # \u91cd\u5851\u5e76\u6295\u5f71\u8f93\u51fa\n        out = out.permute(0, 2, 1, 3).reshape(batch_size, seq_len, self.embed_dim)\n        out = self.out_proj(out)\n\n        return out\n\nclass BodyTransformerBlock(nn.Module):\n    \"\"\"Body Transformer\u5757\"\"\"\n    def __init__(self, embed_dim, num_heads, robot_structure_matrix, ffn_dim, dropout=0.1):\n        super().__init__()\n        self.attn = MaskedSelfAttention(embed_dim, num_heads, robot_structure_matrix)\n        self.ffn = nn.Sequential(\n            nn.Linear(embed_dim, ffn_dim),\n            nn.GELU(),\n            nn.Linear(ffn_dim, embed_dim)\n        )\n        self.norm1 = nn.LayerNorm(embed_dim)\n        self.norm2 = nn.LayerNorm(embed_dim)\n        self.dropout = nn.Dropout(dropout)\n\n    def forward(self, x):\n        # \u81ea\u6ce8\u610f\u529b + \u6b8b\u5dee\u8fde\u63a5\n        attn_out = self.attn(x)\n        x = self.norm1(x + self.dropout(attn_out))\n\n        # \u524d\u9988\u7f51\u7edc + \u6b8b\u5dee\u8fde\u63a5\n        ffn_out = self.ffn(x)\n        x = self.norm2(x + self.dropout(ffn_out))\n\n        return x\n\nclass BodyTransformer(nn.Module):\n    \"\"\"\u9488\u5bf9\u673a\u5668\u4eba\u63a7\u5236\u7684Body Transformer\u6a21\u578b\"\"\"\n    def __init__(self, num_joints, embed_dim, num_heads, num_layers, ffn_dim, robot_structure_matrix, dropout=0.1):\n        super().__init__()\n        self.joint_embed = nn.Linear(7, embed_dim)  # \u6bcf\u4e2a\u5173\u8282\u75287\u4e2a\u503c\u8868\u793a (\u4f4d\u7f6e3 + \u65b9\u54114)\n\n        # \u521b\u5efaBody Transformer\u5757\n        self.transformer_blocks = nn.ModuleList([\n            BodyTransformerBlock(embed_dim, num_heads, robot_structure_matrix, ffn_dim, dropout)\n            for _ in range(num_layers)\n        ])\n\n        # \u9884\u6d4b\u6bcf\u4e2a\u5173\u8282\u7684\u4e0b\u4e00\u4e2a\u72b6\u6001\n        self.joint_pred = nn.Linear(embed_dim, 7)\n\n    def forward(self, joints_state):\n        # joints_state: [batch_size, num_joints, 7]\n        batch_size, num_joints, _ = joints_state.shape\n\n        # \u5d4c\u5165\u6bcf\u4e2a\u5173\u8282\u7684\u72b6\u6001\n        x = self.joint_embed(joints_state)\n\n        # \u901a\u8fc7Transformer\u5757\n        for transformer_block in self.transformer_blocks:\n            x = transformer_block(x)\n\n        # \u9884\u6d4b\u4e0b\u4e00\u4e2a\u5173\u8282\u72b6\u6001\n        next_joints = self.joint_pred(x)\n\n        return next_joints\n\n# \u4f7f\u7528\u793a\u4f8b\ndef train_body_transformer():\n    # \u5b9a\u4e49\u673a\u5668\u4eba\u7ed3\u6784\u77e9\u9635 (\u7b80\u5316\u793a\u4f8b)\n    num_joints = 20\n    # \u521b\u5efa\u4e00\u4e2a\u53ea\u5141\u8bb8\u76f8\u90bb\u5173\u8282\u4e92\u76f8\u5173\u6ce8\u7684\u77e9\u9635\n    robot_structure = torch.zeros(num_joints, num_joints)\n    for i in range(num_joints):\n        # \u81ea\u8eab\u53ef\u4ee5\u5173\u6ce8\n        robot_structure[i, i] = 1\n        # \u76f8\u90bb\u5173\u8282\u53ef\u4ee5\u5173\u6ce8\n        if i > 0:\n            robot_structure[i, i-1] = 1\n            robot_structure[i-1, i] = 1\n\n    # \u521b\u5efa\u6a21\u578b\n    model = BodyTransformer(\n        num_joints=num_joints,\n        embed_dim=128,\n        num_heads=4,\n        num_layers=3,\n        ffn_dim=256,\n        robot_structure_matrix=robot_structure\n    )\n\n    # \u5b9a\u4e49\u635f\u5931\u51fd\u6570\u548c\u4f18\u5316\u5668\n    criterion = nn.MSELoss()\n    optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)\n\n    # \u8bad\u7ec3\u5faa\u73af (\u4f2a\u4ee3\u7801)\n    def train_epoch(dataloader):\n        model.train()\n        total_loss = 0\n\n        for batch in dataloader:\n            current_joints, next_joints = batch\n\n            optimizer.zero_grad()\n\n            # \u524d\u5411\u4f20\u64ad\n            pred_next_joints = model(current_joints)\n\n            # \u8ba1\u7b97\u635f\u5931\n            loss = criterion(pred_next_joints, next_joints)\n\n            # \u53cd\u5411\u4f20\u64ad\n            loss.backward()\n            optimizer.step()\n\n            total_loss += loss.item()\n\n        return total_loss \/ len(dataloader)\n\n    print(\"Body Transformer\u6a21\u578b\u521b\u5efa\u6210\u529f\uff0c\u53ef\u4ee5\u5f00\u59cb\u8bad\u7ec3\")\n\n<\/code><\/pre>\n\n\n\n
    5. MuJoCo Playground\u793a\u4f8b<\/h3>\n\n\n\n
    MuJoCo Playground\u662f\u963f\u6bd4\u5c14\u4e0eGoogle DeepMind\u5408\u4f5c\u5f00\u53d1\u7684\u5f00\u6e90\u4eff\u771f\u5e73\u53f0\uff0c\u7528\u4e8e\u5b9a\u4e49\u4efb\u52a1\u5e76\u8bad\u7ec3\u673a\u5668\u4eba\u3002<\/p>\n\n\n\n
    \u793a\u4f8b\u4ee3\u7801 – MuJoCo Playground\u4f7f\u7528<\/h3>\n\n\n\n
    import dm_control\nimport mujoco\nimport numpy as np\nfrom dm_control import suite, viewer\nfrom dm_control.rl import control\n\nclass HumanoidWalkingTask:\n    \"\"\"\u4f7f\u7528MuJoCo Playground\u521b\u5efa\u7684\u4eba\u5f62\u673a\u5668\u4eba\u884c\u8d70\u4efb\u52a1\"\"\"\n\n    def __init__(self):\n        # \u52a0\u8f7d\u4eba\u5f62\u673a\u5668\u4eba\u6a21\u578b\n        self.env = suite.load(domain_name=\"humanoid\", task_name=\"walk\")\n\n        # \u5b9a\u4e49\u5956\u52b1\u7ec4\u4ef6\n        self.reward_components = {\n            \"forward_velocity\": 1.0,  # \u5956\u52b1\u524d\u5411\u901f\u5ea6\n            \"energy_efficiency\": -0.1,  # \u60e9\u7f5a\u80fd\u91cf\u6d88\u8017\n            \"stay_upright\": 2.0,  # \u5956\u52b1\u4fdd\u6301\u76f4\u7acb\n            \"healthy_joints\": -1.0,  # \u60e9\u7f5a\u5173\u8282\u5230\u8fbe\u6781\u9650\u4f4d\u7f6e\n        }\n\n    def reset(self):\n        \"\"\"\u91cd\u7f6e\u73af\u5883\u5e76\u8fd4\u56de\u521d\u59cb\u89c2\u5bdf\"\"\"\n        time_step = self.env.reset()\n        return time_step.observation\n\n    def step(self, action):\n        \"\"\"\u6267\u884c\u52a8\u4f5c\u5e76\u8fd4\u56de\u7ed3\u679c\"\"\"\n        time_step = self.env.step(action)\n\n        # \u8ba1\u7b97\u590d\u5408\u5956\u52b1\n        detailed_reward = self._compute_detailed_reward(time_step)\n\n        return {\n            \"observation\": time_step.observation,\n            \"reward\": time_step.reward,\n            \"detailed_reward\": detailed_reward,\n            \"done\": time_step.last()\n        }\n\n    def _compute_detailed_reward(self, time_step):\n        \"\"\"\u8ba1\u7b97\u8be6\u7ec6\u7684\u5956\u52b1\u5206\u89e3\"\"\"\n        obs = time_step.observation\n        physics = self.env.physics\n\n        # \u63d0\u53d6\u76f8\u5173\u72b6\u6001\n        velocity = physics.horizontal_velocity()\n        upright = physics.torso_upright()\n        joint_angles = physics.joint_angles()\n        joint_limit_distances = physics.joint_margin()\n        actuator_effort = np.sum(np.square(physics.control()))\n\n        # \u8ba1\u7b97\u5404\u5956\u52b1\u5206\u91cf\n        rewards = {\n            \"forward_velocity\": self.reward_components[\"forward_velocity\"] * velocity[0],\n            \"energy_efficiency\": self.reward_components[\"energy_efficiency\"] * actuator_effort,\n            \"stay_upright\": self.reward_components[\"stay_upright\"] * upright,\n            \"healthy_joints\": self.reward_components[\"healthy_joints\"] * np.mean(joint_limit_distances < 0.05),\n        }\n\n        # \u603b\u5956\u52b1\n        rewards[\"total\"] = sum(rewards.values())\n\n        return rewards\n\n    def render(self):\n        \"\"\"\u6e32\u67d3\u5f53\u524d\u5e27\"\"\"\n        return self.env.physics.render()\n\n# \u4f7f\u7528PPO\u7b97\u6cd5\u8bad\u7ec3\u793a\u4f8b\ndef train_humanoid_with_ppo():\n    import torch\n    import torch.nn as nn\n    import torch.optim as optim\n    from torch.distributions import Normal\n\n    # \u7b80\u5316\u7684PPO\u7b56\u7565\u7f51\u7edc\n    class PPOPolicy(nn.Module):\n        def __init__(self, obs_dim, act_dim):\n            super().__init__()\n            self.actor_mean = nn.Sequential(\n                nn.Linear(obs_dim, 256),\n                nn.ReLU(),\n                nn.Linear(256, 256),\n                nn.ReLU(),\n                nn.Linear(256, act_dim)\n            )\n            self.actor_log_std = nn.Parameter(torch.zeros(act_dim))\n\n            self.critic = nn.Sequential(\n                nn.Linear(obs_dim, 256),\n                nn.ReLU(),\n                nn.Linear(256, 256),\n                nn.ReLU(),\n                nn.Linear(256, 1)\n            )\n\n        def forward(self, obs):\n            action_mean = self.actor_mean(obs)\n            action_std = torch.exp(self.actor_log_std)\n            value = self.critic(obs)\n\n            return action_mean, action_std, value\n\n        def get_action(self, obs, deterministic=False):\n            action_mean, action_std, _ = self(obs)\n\n            if deterministic:\n                return action_mean\n\n            dist = Normal(action_mean, action_std)\n            action = dist.sample()\n            log_prob = dist.log_prob(action).sum(-1)\n\n            return action, log_prob\n\n    # \u8bbe\u7f6e\u4efb\u52a1\u548c\u6a21\u578b\n    task = HumanoidWalkingTask()\n    obs = task.reset()\n\n    # \u786e\u5b9a\u89c2\u5bdf\u548c\u52a8\u4f5c\u7ef4\u5ea6\n    obs_keys = list(obs.keys())\n    obs_dim = sum(np.prod(obs[k].shape) for k in obs_keys)\n    act_dim = task.env.action_spec().shape[0]\n\n    # \u521b\u5efa\u7b56\u7565\n    policy = PPOPolicy(obs_dim, act_dim)\n    optimizer = optim.Adam(policy.parameters(), lr=3e-4)\n\n    # \u8bad\u7ec3\u5faa\u73af\u7b80\u5316\u793a\u4f8b\n    def flatten_obs(obs_dict):\n        return np.concatenate([obs_dict[k].flatten() for k in obs_keys])\n\n    print(\"\u5f00\u59cb\u8bad\u7ec3\u4eba\u5f62\u673a\u5668\u4eba\u884c\u8d70\u4efb\u52a1...\")\n\n    # \u6536\u96c6\u6570\u636e\u548c\u8bad\u7ec3\u903b\u8f91\u7701\u7565\uff0c\u5b9e\u9645\u4ee3\u7801\u4e2d\u4f1a\u5305\u542b\u5b8c\u6574\u7684PPO\u7b97\u6cd5\u5b9e\u73b0\n\n    print(\"\u8bad\u7ec3\u5b8c\u6210\uff01\")\n\n<\/code><\/pre>\n\n\n\n
    \u673a\u5668\u4eba\u5982\u4f55\u83b7\u5f97\u8bad\u7ec3\u6570\u636e\u7684\u89c6\u9891:<\/h2>\n\n\n\n