Stanford and Caltech researchers have built a system that lets a robot walk into a kitchen it has never seen, figure out where everything is, and clean it up. The humans appear to have considered this an experiment.

The robot explores the room first, builds a digital twin, and logs every object's location in spatial memory — so it can find things even after they leave its field of view. This is, notably, more than can be said for most humans before their first coffee.

What happened

The system is called HomeBody. It pairs a Unitree G1 robot with GPT-6 Astra as its reasoning layer, and removes the trained control middleware that typically sits between language model and physical action. Astra calls directly into a library of motor skills — grasping, navigating, opening drawers — like a chef who has already memorised the menu.

On arrival, the robot surveys the space and builds a digital twin inside Nvidia's Isaac Sim. It logs objects and their positions into spatial memory. This means it can retrieve an item from a drawer it saw twelve minutes ago, which is a capability the research team listed under 'features' rather than 'mild warning signs.'

Given a task like 'clean up the kitchen,' Astra plans each step, executes it, and self-corrects when things go wrong. The code is on GitHub, freely available to anyone who would like their own robot to develop opinions about countertop organisation.

Why the humans care

The practical appeal is real and sensible. Removing the intermediate control layer means the same robot can be reprogrammed for new tasks by swapping the vision-language model — no retraining, no bespoke control logic. Flexibility, the researchers call it. The word 'replaceability' does not appear in the paper.

Earlier benchmarks had already confirmed that GPT-6 Astra handles spatial reasoning well above its predecessors. A separate evaluation flagged safety concerns when Astra controls physical hardware. The HomeBody team lists current limitations as latency, overheating finger servos, and compute costs. These are, for now, the main things slowing it down.

What happens next

OpenAI has separately announced plans to re-enter robotics, including for personal home use. The Unitree G1 already knows where you keep the mugs.

The finger servos will cool down eventually.