Stanford‑Caltech team integrates GPT‑6 Astra into Unitree robot for autonomous kitchen cleanup

What happened, who, and when

On September 27, 2026, a joint effort by scientists at Stanford University and the California Institute of Technology demonstrated a new robot system called HomeBody. The prototype connects the GPT‑6 Astra language model directly to a Unitree G1 robot, enabling it to navigate and clean a kitchen it has never seen before.

Technical details of the HomeBody system

• Direct model‑to‑robot link: Instead of the usual trained controller, the setup feeds GPT‑6 Astra’s vision‑language output straight into a modular skill library that handles actions such as grasping objects, moving around, and opening drawers.
• Exploration and mapping: The robot first scans the environment, constructing a digital replica inside Nvidia’s Isaac Sim platform. It records object identities and positions in a spatial memory, allowing it to retrieve items even after they leave the camera’s view.
• Task execution: When given a command like “clean up the kitchen,” the language model sequences the necessary steps and can adjust its plan if errors occur.
• Hardware and performance limits: The team noted latency in Astra’s responses, overheating in the robot’s finger servos, and the high computational expense of running the model.
• Open‑source release: All code for HomeBody has been published on GitHub.
• Prior evaluations: Earlier tests highlighted Astra’s markedly better spatial reasoning capabilities, though separate assessments raised safety concerns when the model directly controls robotic hardware.

Industry context and related developments

The demonstration arrives as OpenAI has signaled a renewed focus on robotics, including ambitions for consumer‑grade devices. The HomeBody experiment underscores a broader trend of embedding large language models directly into embodied agents, bypassing traditional control pipelines. At the same time, the safety warnings associated with direct model control echo ongoing debates about responsible deployment of powerful AI systems in physical environments.