HomeBody Gives GPT Astra Spatial Memory and a Humanoid Body

HomeBody connects GPT Astra to spatial memory and a library of skills for a Unitree G1 humanoid. Researchers from Caltech and Stanford demonstrate multi-object cleanup and medicine retrieval from a drawer after an underspecified request in a previously unseen kitchen.

GPT Astra selects skills instead of driving motors directly

The system replaces a learned VLA middle layer with a VLM that calls reusable skills. GPT Astra chooses navigation, picking, placing or drawer opening and supplies a structured target. Specialized planners and controllers still handle low-level motion.

Before doing useful work, the robot explores the room with cameras, LiDAR, SLAM and joint positions. Astra helps build a digital twin from this data in Isaac Sim. Stored keyframes then let the system recall an object outside its current field of view. The authors say the deployment uses no environment-specific training data or additional policy learning.

The language model becomes a planner for physical routines

Mollick's observation about the surprising breadth of LLM applications lands precisely here. The language model does not send 250 Hz commands to the motors. It turns an ambiguous human intention into a sequence of available actions and chooses the next step from tool feedback after a failure.

The modularity matters for robotics. A new skill can be connected through a shared interface instead of retraining the whole system for one kitchen. The same advantage creates dependence on the map, object descriptions and the reliability of every component skill.

Autonomy rests on extensive preparation and fixed infrastructure

The demonstration does not begin in an empty home with one sentence. It requires exploration, map registration, Real2Sim reconstruction and a laptop with an RTX 4090. GPT Astra runs remotely, and reasoning latency pauses the robot between skills. The authors also cite limited reach, finger-servo overheating, and the time and API cost of the digital twin.

New rooms and off-script failures will settle the case

The next meaningful signal is repeated deployment in different rooms with unfamiliar drawers, obstacles and objects. End-to-end task success, setup time and safe termination after errors will matter. Those results will show whether a reusable skill library scales better than a policy trained for a specific environment.

Lilith's verdict

A humanoid that finds medicine in a closed drawer already looks like a coworker. While it still needs a digital twin, an RTX 4090 and patience with pauses before the shift, the floor manager can keep the keys in a pocket.