AI + 机器人实战教程 — 从0到1搭建具身智能Agent

2026年具身智能Embodied AI全面爆发本文手把手教你用ROS2 LangGraph VLM搭建具身智能机器人Agent实现“看懂世界 自主行动”。包含硬件选型、完整代码、仿真环境和真实部署方案。跟着做你就能做出自己的AI机器人一、2026具身智能为什么这么火2026年AI从“云端对话”走向“物理世界行动”。具身智能让机器人真正理解物理规律、自主完成任务成为机器人、工业、服务等领域核心技术。核心趋势世界模型 VLM视觉语言模型多Agent协同控制合成数据训练 真实世界微调机会掌握具身智能 抓住机器人时代红利。二、硬件选型推荐入门级Raspberry Pi 5 Hailo加速卡进阶级NVIDIA Jetson Orin Nano / Unitree Go2开发平台专业级全尺寸人形机器人平台如宇树H1、智元机器人三、完整实战教程AI视觉抓取机器人目标机器人看到物体后自主规划路径并抓取。Step 1环境搭建ROS2# 安装ROS2 Humble sudo apt install ros-humble-desktop # 安装必要包 pip install ultralytics langgraph opencv-pythonStep 2核心代码VLM LangGraph Agentimport cv2 from ultralytics import YOLO from langgraph.graph import StateGraph, END from langchain_openai import ChatOpenAI llm ChatOpenAI(modelqwen-vl-max) class RobotState(dict): image: any objects: list plan: str action: str def vision_module(state): model YOLO(yolov8n.pt) results model(state[image]) objects [r.names[int(cls)] for r in results[0].boxes.cls] return {objects: objects} def planner_agent(state): prompt f看到物体{state[objects]}请规划抓取动作。 res llm.invoke(prompt) return {plan: res.content} def executor_agent(state): # 调用机器人控制接口 print(f执行动作{state[plan]}) return {action: 抓取完成} workflow StateGraph(RobotState) workflow.add_node(vision, vision_module) workflow.add_node(planner, planner_agent) workflow.add_node(executor, executor_agent) workflow.set_entry_point(vision) workflow.add_edge(vision, planner) workflow.add_edge(planner, executor) workflow.add_edge(executor, END) app workflow.compile() # 运行 frame cv2.imread(scene.jpg) result app.invoke({image: frame})运行效果输入摄像头画面Agent自动识别、规划、执行。四、ROS2仿真环境搭建详细版ROS2是具身智能开发的核心框架仿真环境可以让你在没有真实硬件的情况下快速验证算法。1. 安装Gazebo仿真器# Ubuntu 22.04 ROS2 Humble sudo apt update sudo apt install ros-humble-gazebo-ros-pkgs ros-humble-gazebo-ros2-control -y # 验证安装 gazebo --version常用命令启动空世界gazebo启动带机器人的世界ros2 launch gazebo_ros gazebo.launch.py2. 使用MoveIt!规划机械臂MoveIt! 是ROS2中最强大的机械臂运动规划库。安装sudo apt install ros-humble-moveit ros-humble-moveit-servo -y创建机械臂项目# 创建工作空间 mkdir -p ~/ros2_ws/src cd ~/ros2_ws/src ros2 pkg create --build-type ament_cmake my_robot_moveit # 使用MoveIt Setup Assistant生成配置文件 ros2 launch moveit_setup_assistant setup_assistant.launch.py核心规划代码示例moveit_planning.pyimport rclpy from moveit_commander import MoveGroupCommander, PlanningSceneInterface from geometry_msgs.msg import Pose rclpy.init() move_group MoveGroupCommander(arm) # 设置目标位姿 target_pose Pose() target_pose.position.x 0.3 target_pose.position.y 0.2 target_pose.position.z 0.5 target_pose.orientation.w 1.0 move_group.set_pose_target(target_pose) success move_group.go(waitTrue) print(规划是否成功, success) rclpy.shutdown()3. 结合NVIDIA Isaac Sim进行高保真仿真Isaac Sim是目前最强大的机器人仿真平台尤其适合具身智能。安装步骤NVIDIA官方安装Omniverse Launcher下载Isaac Sim 2023.1版本安装ROS2 Bridge扩展Isaac Sim ROS2桥接配置# 启动Isaac Sim仿真 ./isaac_sim.sh # 启动ROS2 Bridge ros2 run isaac_ros_bridge isaac_ros_bridge优势物理仿真极度真实支持光照、碰撞、传感器模拟可直接训练具身智能策略。五、真实硬件部署详细版1. Jetson Orin Nano上部署量化模型# 安装TensorRT sudo apt install tensorrt # 模型量化FP16 python3 -m torch2trt model.pth --fp16 -o model.engine # 推理加速测试 trtexec --loadEnginemodel.engine --shapesinput:1x3x640x640部署YOLOv8示例from ultralytics import YOLO import tensorrt as trt model YOLO(yolov8n.engine) # 使用TensorRT引擎 results model(test_image.jpg, devicecuda)2. 使用ROS2 Bridge连接机械臂# 安装机械臂驱动 ros2 pkg create my_robot_driver # 示例驱动节点 class ArmDriver(Node): def __init__(self): super().__init__(arm_driver) self.publisher self.create_publisher(JointState, joint_states, 10) def publish_joint_states(self): msg JointState() msg.position [0.5, 1.2, -0.8] # 示例关节角度 self.publisher.publish(msg)配置修改urdf文件描述机械臂结构使用ros2_control硬件接口3. 实现闭环控制Vision Planning Execution# 完整闭环控制节点 class EmbodiedAgent(Node): def __init__(self): super().__init__(embodied_agent) self.vision_sub self.create_subscription(Image, camera/image, self.vision_callback, 10) self.arm_client self.create_client(FollowJointTrajectory, /arm_controller/follow_joint_trajectory) def vision_callback(self, msg): # 视觉识别 detections self.detect_objects(msg) # 规划 plan self.plan_grasp(detections) # 执行 self.execute_plan(plan)闭环控制流程摄像头获取图像VLM / YOLO识别物体MoveIt!规划抓取轨迹执行器执行动作反馈调整闭环