GPT-6 Astra Powers HomeBody Robot in Unfamiliar Kitchen Cleanup Test

Researchers from Stanford and Caltech built HomeBody, a system that lets a Unitree G1 robot work autonomously in an unfamiliar kitchen. In the test described by The Decoder, the robot navigated the space, tidied up, and fetched items from drawers.

The key takeaway is that GPT-6 Astra was used as the system’s vision-language model, helping plan the robot’s actions in a real-world household setting.

HomeBody drops the typical trained control layer between the language model and the robot. Instead, GPT Astra calls directly into an extensible skill library for actions such as grasping, navigating, and opening drawers.

That setup points to a more modular approach: the language model handles planning while reusable robot skills handle physical execution.

Before carrying out tasks, the robot explores the room, builds a digital twin in Nvidia’s Isaac Sim, and logs objects and locations in spatial memory. This helps it find items even after they leave its field of view.

For broad instructions such as “clean up the kitchen,” the model plans steps and self-corrects when errors occur.

The system still faces practical limits, including Astra’s latency, overheating finger servos, and high compute costs. The article also notes that the code is available on GitHub.

The test follows earlier reports on Astra’s improved spatial reasoning and separate safety concerns when Astra controls a robot. The Decoder also notes that OpenAI has announced plans to return to robotics, including personal-use applications.