HomeBody: a vision model directly pilots a Unitree G1, without a learned action layer

Teams from Stanford and Caltech hook up GPT-6 Astra to five of a humanoid's motor skills and had him tidy up a kitchen he'd never seen.

The usual stack of humanoid autonomy stacks three floors: a vision-language model that reflects, a trained vision-language-action policy that translates intention into gestures, a low-level controller that executes. The HomeBody project, conducted jointly by Stanford and Caltech, removes the middle floor. GPT-6 Astra calls itself five composable skills from the Unitree G1 — navigate, grasp, pose, open a drawer, take from a drawer — with no learned action policy in between.

The binder is a persistent spatial memory. The robot first explores the place, and the model derives a digital twin from it in Isaac Sim which then serves as a reference for planning. In a kitchen never encountered, and without the slightest training data specific to this environment, the system tidies up the room and then finds an object seen earlier from a deliberately vague instruction.

The hardware set-up remains modest: perception and motion planning run on a single laptop equipped with an RTX 4090, while Astra runs remotely and simply sends skill requests with targets, then receives the result. The interest of the demonstration lies less in the raw performance than in what it suggests: some of the work that was entrusted to specialized politicians can now be transferred to the generalist model.