HomeBody: A humanoid that explores, remembers, and acts on its own
8 hours ago
- HomeBody uses spatial targets to connect task reasoning to physical execution, with the VLM selecting skills and targets based on current context.
- The VLM passes selections via structured tool calls, leaving low-level motion planning to the skills themselves.
- For picking, the VLM provides a normalized image point and hand selection; the system then uses segmentation, depth estimation, and analytic grasp prediction.
- The arm planner builds a spline reference with minimum-jerk timing, solves inverse kinematics, and checks for collision clearance.
- Other skills include navigation (2D goal and facing point in map coordinates), placing (3D release target and distance), and coordinated drawer opening combining alignment, hooking, and walking.