Stanford's HomeBody wires GPT Astra to a Unitree G1 humanoid

HomeBody skips the middle. A team from Stanford and Caltech has wired GPT Astra directly to a Unitree G1 humanoid's motor skill library, with no learned action policy in between. The project page from Stanford's TML describes the traditional humanoid autonomy stack as "System 2 VLM → System 1 VLA → System 0 controller," and replaces it with a frontier VLM calling a small set of composable skills against a persistent spatial memory.

The vocabulary is five moves: navigation, picking, placing, drawer opening, and picking from drawers. Before the robot acts, it explores. Its exploration data grounds a Real2Sim reconstruction of the space, so the VLM can reason about objects and locations it isn't currently looking at. "HomeBody uses the humanoid's own exploration data as grounding for the Real2Sim agent, including camera observations, measured SLAM geometry, joint poses and selected waypoints," the writeup states. Local compute runs on a Razer Blade laptop with an RTX 4090; GPT Astra runs remotely. Perception uses SAM 2.1 with SAMURAI memory selection and Fast-FoundationStereo for depth from a D435i stereo camera. Arm and hand commands go out at 250 Hz.

The two demonstrated tasks are the kind of thing you'd expect from a kitchen. One instruction reads: "Clean up all of the coffee bags and put them in the middle, and throw away all of the milk and orange juice cartons that have gone bad." Another: "I forgot my medicine, can you get it for me? Also throw out the bad carton while you are at it." The medicine task requires remembering where the drawer was, using the right hand to open it and the left to dispose of trash.

The page is candid about limits. Real2Sim reconstruction "introduces setup time and API costs." GPT Astra reasoning "creates pauses between skill executions." Task length is bounded by, among other things, "finger-servo overheating during extended operation," the deadpan detail of a robot pushed until its fingers get hot. Three researchers we track in our Who's Who directory posted the project link within a day of it appearing.