Researchers from Stanford and Caltech have demonstrated HomeBody, a system that lets a Unitree G1 humanoid navigate an unfamiliar kitchen, put items away and retrieve objects from drawers. The experiment connects GPT-6 Astra directly to reusable robot skills instead of relying on a control policy trained specifically for the room.
The robot first explores the space, builds a digital twin in Nvidia Isaac Sim and stores the locations of objects in spatial memory. A vision-language model then plans a request such as “clean up the kitchen” and calls modules for navigation, grasping or opening drawers. Because locations persist in memory, the system can look for an item after it has left the cameras' view and revise a plan when a step fails.
This modular design could make one planning model useful across different robots and tasks, while keeping low-level movement in specialized skills. The researchers also published the code on GitHub.
The demonstration is not a general-purpose household product. The team reports latency from the model, high computing costs and overheating finger servos. Earlier evaluations have also raised safety concerns when Astra controls a physical robot, so the work shows a promising architecture rather than dependable unsupervised operation.