GPT-6 Astra Drives Unitree's G1 to Clean a Kitchen

GPT-6 Astra has been connected to Unitree's G1 humanoid robot, and in a demonstration the machine tidies up a kitchen, according to the Chinese outlet QbitAI. The interesting part is not the motion itself but the framing: GPT treats the robot's skills as tools it can invoke.

The "skills as tools" approach puts the large model in the role of orchestrator. It does not output joint torques directly; instead it reads the situation, selects an appropriate skill interface, and hands execution to the robot body. The model handles understanding and planning, and the hardware handles actuation.

The value of that abstraction layer is generalization. As long as the skill library covers the needed motions, new household tasks can be assembled through prompting and composition rather than fresh data collection and retraining.

A kitchen is a classic testbed for this kind of demo: many object types, variable placement, ordered steps, and contact-rich actions such as opening cabinets and setting down dishes all stress perception and planning.

Unitree's G1 supplies the body while GPT-6 Astra supplies the high-level decisions. The combination matters because general models are moving from chat assistants on a screen toward the orchestration layer that commands the physical world.

It is worth staying sober about the gap between a demo and deployment. Demos typically involve staged scenes and multiple attempts, and long-duration stability, failure recovery, and safety still need far more validation.

What to watch next: how open the skill library is, whether third-party hardware is supported, and how quickly this moves from scripted chores like tidying a kitchen toward genuinely open environments.