HomeBody: GPT-6 Astra runs a Unitree G1 kitchen tidy-up

HomeBody is a test of how far a general vision-language model can go in controlling a robot without a dedicated trained layer. The robot first explores the room, builds a digital twin in Nvidia's Isaac Sim (a simulation environment for robots), and logs objects and locations in spatial memory. That lets it find items even after they leave its field of view. For a task like cleaning up the kitchen, the language model plans each step and self-corrects on errors.

Earlier benchmarks had already shown GPT Astra's greatly improved spatial reasoning, while another flagged safety issues when Astra controls a robot. OpenAI has also announced plans to get back into robotics, including for personal use. HomeBody fits that trend by showing a language model can drive a real robot's actions.

What the approach yields in practice will hinge on whether latency, overheating finger servos, and high compute costs can be brought down. For robotics researchers, the GitHub code offers a way to test language-model-direct control, but everyday home use looks distant until those issues are resolved.