HomeBody: A Humanoid That Explores, Remembers, and Acts on Its Own
Summary
The paper introduces HomeBody, a humanoid system that can explore unseen environments, build a persistent spatial memory, and execute long-horizon tasks by invoking a plug-and-play library of composable skills, rather than relying on a fixed action sequence. It grounds high-level planning in a Real2Sim digital twin built from the robot's own exploration data, enabling robust reasoning across ego-view geometry and a shared map. The work covers architecture, implementation details, and limitations, including compute requirements and latency implications.