Unitree G1 Robot Controlled by AI GPT-6 Astra

Researchers from Stanford and Caltech have demonstrated the HomeBody system, in which a Unitree G1 robot performs household tasks in unfamiliar rooms without prior training for a specific space.

Control is carried out using the OpenAI GPT-6 Astra model. The architecture lacks an intermediate VLA layer: the model independently selects the necessary skill to complete the task.

The system has mastered five basic actions: movement, picking up an item, placing it, opening a drawer, and retrieving objects.

The process begins with the robot navigating the room using cameras and lidar. Based on this data, Astra creates a digital twin of the space in the NVIDIA Isaac Sim simulator.

The system's spatial memory retains information about object locations even if they are out of direct view. This allows the robot to recall where a specific object was spotted after receiving a command.

The authors do not provide success statistics, including the percentage of successful attempts or the total number of tests conducted.

Key limitations include pauses for the model to think, the risk of servo overheating in the fingers, and the high cost of API usage.