Researchers from Stanford and Caltech developed HomeBody, a system enabling a Unitree G1 robot to autonomously tidy up an unfamiliar kitchen, showcasing GPT-6 Astra's effectiveness in robotics.
The system eliminates the need for a trained control layer between the language model and robot, instead using a swappable vision-language model, GPT Astra, to directly call an extensible skill library for grasping, navigating, or opening drawers.
HomeBody allows the robot to explore the room, build a digital twin in Nvidia's Isaac Sim, and log objects and locations in spatial memory, enabling it to find items even after they leave its field of view. For tasks like 'clean up the kitchen,' the language model plans each step and self-corrects on errors, demonstrating its adaptability in dynamic environments.
The system faces limitations such as GPT Astra's latency, overheating finger servos, and high compute costs, which may affect its practical deployment. The code is available on GitHub, allowing further development and testing by the research community.
"HomeBody drops the typical trained control layer between language model and robot," said Matthias Bastian, a researcher at Stanford. "Instead, a swappable vision-language model calls directly into an extensible skill library for grasping, navigating, or opening drawers."
The announcement follows OpenAI's announcement of plans to get back into robotics, including for personal use, highlighting the growing interest in integrating AI with physical tasks. The research underscores the potential of combining advanced language models with robotic systems for real-world applications.
Researchers did not specify the exact timeline for commercial deployment, and the system's high compute costs remain a challenge. The next step involves refining the model's efficiency and exploring ways to reduce hardware limitations.
Source: thedecoder
