Researchers Connect GPT-6 Astra to Unitree G1 Robot

Stanford University and California Institute of Technology researchers have introduced HomeBody, an open-source system that allows a Unitree G1 humanoid robot to autonomously tidy up, navigate, and retrieve items from drawers in an unfamiliar kitchen environment. By linking OpenAI's GPT-6 Astra vision-language model directly to the robot's physical capabilities, the system bypasses the traditional intermediate control layers that typically translate high-level instructions into physical actions.

Instead of relying on a dedicated, pre-trained control layer, HomeBody utilizes a swappable vision-language model to call directly into an extensible library of basic skills, such as grasping, navigating, and opening drawers. Before performing tasks, the Unitree G1 robot explores its environment to construct a digital twin within Nvidia's Isaac Sim platform. It logs the locations of various objects in a spatial memory database, which enables the robot to locate and retrieve items even after they are no longer in its immediate line of sight. When tasked with cleaning, GPT-6 Astra generates step-by-step plans and self-corrects when errors occur.

For robotics practitioners, this architecture simplifies deployment by removing the need to train complex end-to-end control policies for every new environment. However, the researchers noted several hardware and software bottlenecks during testing. These include high computational costs, significant latency from the GPT-6 Astra model, and physical limitations such as overheating finger servos on the Unitree G1.

The code for HomeBody has been made publicly available on GitHub. While previous benchmarks have highlighted the improved spatial reasoning capabilities of the Astra model, other evaluations have raised safety concerns regarding direct model control over physical hardware. This development comes as OpenAI signals its intention to re-enter the robotics space, particularly for personal and domestic applications.