HomeBody是一款人形机器人系统，利用视觉语言模型（VLM）实现自主探索、记忆和物理执行的能力1。该系统通过空间目标将任务推理与物理执行连接，使机器人能够选择和执行多种动作1。
VLM根据当前视图、地图上下文、夹爪状态、回忆观察和前一步操作结果来选择技能和目标，然后通过结构化工具调用将任务分解给底层技能执行1。HomeBody支持的操作包括抓取、导航、放置和开抽屉等1。在抓取操作中，系统使用0–1000归一化的图像点坐标、双手选择、分割和深度估计，其中深度通过Fast-FoundationStereo从D435i立体摄像头估计1。导航以2D目标和朝向点（地图坐标，单位米）指定1。放置操作则指定使用哪只手、躯干坐标系中的3D释放目标和释放距离1。机器人的臂部规划器采用最小加速度样条曲线、反向运动学求解和扫过碰撞检查来实现精确的动作控制1。
HomeBody is a humanoid robot designed to perform autonomous exploration, memory retention, and physical task execution through the integration of visual language models (VLM) technology.1 The system connects task reasoning to physical action via spatial objectives, with the VLM selecting skills and targets based on the current visual input, map context, gripper state, recalled observations, and previous results.1
The robot's operational capabilities are executed through structured tool invocation that decomposes tasks into lower-level skill execution, including grasping, navigation, placing objects, and drawer opening.1 Grasping operations utilize normalized image point coordinates ranging from 0 to 1000, with dual-hand selection, segmentation, and depth estimation, while depth is estimated using Fast-FoundationStereo from D435i stereo cameras.1 Navigation is specified via 2D targets and orientation points expressed in map coordinates measured in meters, and placement operations designate which arm to use, the 3D release target in torso coordinates, and release distance.1 The arm planner employs minimum acceleration spline curves, inverse kinematics solving, and swept collision checking to execute these movements.1
评论
还没有评论，欢迎留下第一条。