HomeBody is a humanoid robot designed for long-horizon loco-manipulation tasks in unstructured environments like kitchens. It combines spatial memory with composable skills to perform complex actions such as tidying up, retrieving items, and discarding waste—even when objects are out of view. The system uses exploration to build an internal spatial model from ego-view data including iPhone video, D435i camera, LiDAR scans with SLAM, joint poses, and waypoints selected by Astra.

This data grounds a Real2Sim process in Isaac Sim, creating a geometrically, semantically, and visually accurate digital twin for reasoning beyond immediate perception. HomeBody executes tasks like gathering coffee bags, placing them on an island, and removing spoiled cartons by coordinating walking, grasping, and manipulation across both arms. It uses stored keyframes to locate hidden objects, such as medicine in a drawer, and retrieves them using bimanual coordination.

Skill libraries for navigation, picking, placing, and drawer opening are composable, with local retries and visual feedback enabling the VLM to revise decisions after failures. The robot adapts dynamically to changing views and incomplete tasks through feedback loops between action execution and spatial memory updates.