Stanford robot explores, tidies unfamiliar kitchen on its own
Researchers built HomeBody, software that lets a Unitree G1 humanoid map a strange kitchen, then fetch items and correct its own mistakes using a vision-language model the team calls Astra.
Stanford and Caltech researchers built HomeBody, software that lets a Unitree G1 humanoid explore an unfamiliar kitchen and remember what it finds. It can then tidy the room up on its own, according to the project page.
The robot first explores a room and builds a 3D digital twin in Nvidia's Isaac Sim, storing where objects are in spatial memory, the researchers said. It then calls on a vision-language model The Decoder identified as GPT-6 Astra to plan each step. That model fetches items from drawers and corrects itself when something goes wrong.
Astra plans and re-plans in real time rather than following a pre-trained control policy, according to the project page. The setup runs locally on an RTX 4090 laptop while querying the vision-language model remotely, and uses Meta's SAM 2.1 tracker to track objects as the robot moves.
The Decoder reported the system still struggles with the model's latency, overheating finger servos and heavy compute costs. The team, which includes researchers from Stanford and Caltech, published its code on GitHub and says a paper is forthcoming.