Researchers plug GPT-6 Astra directly into a robot and let it clean up an unfamiliar kitchen
decoder ·
Researchers from Stanford and Caltech have developed HomeBody, a system that enables a Unitree G1 humanoid robot to autonomously clean an unfamiliar kitchen using the GPT-6 Astra language model. This approach eliminates the traditional trained control layer, allowing the vision-language model to directly access a library of modular skills such as grasping and navigation. The robot first explores the environment to create a digital twin in Nvidia's Isaac Sim, logging objects and locations in spatial memory to track items even when they are out of view. When given commands like cleaning the kitchen, the language model plans each step individually and self-corrects errors during execution. Despite these advancements, the system faces limitations including latency issues, overheating finger servos, and high computational costs. The project highlights significant improvements in spatial reasoning for large language models while also raising safety concerns regarding direct robot control. OpenAI has previously announced intentions to re-enter the robotics market with a focus on personal use applications.
decoder