Stanford's HomeBody uses GPT-6 Astra to drive a Unitree G1 robot for housework
A Unitree G1 first explores a kitchen to build spatial memory, then GPT-6 Astra plans multi-step tasks like navigation, picking and opening drawers without task-specific training data.
GPT-6 Astra+Unitree G1->home-humanoid Stanford’s HomeBody is testing a different way to build humanoid intelligence. A Unitree G1 first explores an unfamiliar kitchen, builds persistent spatial memory and a Real2Sim digital twin, then uses GPT-6 Astra to plan and compose skills like Navigate, Pick, Place and Open Drawer. It can handle multi-step tasks across the room without environment-specific training data or additional policy learning. There’s no learned VLA sitting between the VLM and the robot’s skills. The VLM decides what to do. Skills handle how to do it. Execution results are fed back to the model, allowing it to revise the next step when something fails. It inevitably makes one reflect--> As general models become better at spatial reasoning, planning and generalization, how much more do we actually need to build for the robot stack?