Stanford and Caltech Robot Tidies Unfamiliar Kitchen Using GPT-6 Astra as Its Only Brain
A new HomeBody system lets a language model call robot skills directly, skipping the trained control layer that usually sits between the two.
Researchers at Stanford and Caltech have built a humanoid robot that can walk into a kitchen it has never seen and put it in order, with GPT-6 Astra making the decisions. The system, called HomeBody, removes the purpose-built control layer that normally translates a language model's intentions into motor commands. Instead, the model calls modular skills such as grasping and navigating on its own, and the robot acts.
The setup inverts the usual division of labor in robotics. Most systems pair a language model with a separately trained policy that handles low-level execution, and that policy carries most of the engineering effort. HomeBody treats those skills as callable tools and leaves the sequencing to the model, so the intelligence that decides what to pick up and where to put it is the same intelligence that runs the rest of the stack.
The demonstration took place in an unfamiliar kitchen, which is the part that matters. A known room allows for maps, pre-programmed object locations and rehearsed motions; an unfamiliar one does not. The robot had to identify clutter, choose an order of operations and move between stations without a prior walkthrough, according to the team's account of the work.
Skipping the trained control layer is a bet that general-purpose models are now good enough at grounding language in physical action. If it holds up, teams could add new behaviors by adding skills and prompts rather than retraining a policy for every environment, which is slow and expensive. That would lower the cost of getting a robot to do useful work in a house, a lab or a stockroom.
The obvious question is reliability. A model that plans directly against hardware has less of a safety net when a grasp slips or a surface turns out to be fragile, and the published account does not yet say how often the robot finished the job cleanly across many trials. Watch for follow-up results on success rates, failure recovery and how much of the performance depends on the specific kitchen used in the demo.