Scientists connect GPT-6 Astra straight to a robot and have it tidy an unfamiliar kitchen
Stanford and Caltech researchers introduced HomeBody, a system enabling GPT-6 Astra to directly control a Unitree G1 humanoid without a trained control layer. The robot explores unfamiliar kitchens, builds a digital twin in Nvidia Isaac Sim, uses spatial memory, and self-corrects during tasks like tidying up. Latency, overheating servos, and compute costs remain challenges, but the demo hints at OpenAI's robotics ambitions.
Key Takeaways
- Researchers at Stanford and Caltech introduced HomeBody, a system letting a Unitree G1 humanoid robot tidy an unfamiliar kitchen with GPT-6 Astra controlling the hardware directly — no trained control layer in between.
- The robot first explores the space and builds a digital twin in Nvidia Isaac Sim with spatial memory, so it can locate items even after they leave its view, and it can plan and self-correct on open-ended tasks like "clean up the kitchen."
- Current bottlenecks include the model's response latency, overheating finger servos, and high compute costs; the code is available on GitHub.
- Earlier studies showed Astra's improved spatial reasoning, but also raised safety concerns about putting the model in charge of physical robots — benchmark gains don't guarantee real-world safety.
- The project aligns with OpenAI's announced return to robotics, but latency, overheating, and cost must be solved before such direct-control systems become consumer products.
Researchers at Stanford and Caltech have unveiled HomeBody, a robot system that allows a Unitree G1 humanoid to explore an unfamiliar kitchen, tidy up, and retrieve items from drawers entirely on its own. The experiment is notable for what it removes: the trained control layer that usually sits between a language model and a robot's hardware. Instead, GPT-6 Astra — OpenAI's latest model — calls a skill library directly, a sign that large language models may be moving closer to direct robot control.
How HomeBody Works
HomeBody replaces conventional learned control policies with a swappable vision-language model (VLM), a type of AI that understands both images and text. In this setup, GPT-6 Astra taps into an extensible library of skills covering grasping, navigation, and opening drawers.
Before attempting any task, the Unitree G1 first explores the room and constructs a digital twin — a simulated replica — inside Nvidia's Isaac Sim platform. Objects and their locations are recorded in spatial memory, which means the robot can still find an item after it moves out of view. For an open-ended instruction like "clean up the kitchen," Astra plans each step and self-corrects when something goes wrong.
Limitations and Open Questions
The system is not without problems. The researchers point to Astra's response latency, overheating finger servos on the robot's hands, and high compute costs as current bottlenecks. The HomeBody code has been released on GitHub for others to build on.
The results also arrive alongside mixed earlier findings. Prior benchmarks showed Astra delivering significantly improved spatial reasoning, but separate research flagged safety concerns when the model is put in charge of a physical robot — a reminder that benchmark gains do not automatically translate into safe real-world behavior.
Why It Matters
The project lands at a moment of renewed industry interest. OpenAI has already announced plans to return to robotics, including robots intended for personal use. HomeBody offers an early glimpse of what that might look like: a general-purpose model orchestrating a commercial humanoid through everyday household chores.
The current stage is clearly experimental. Latency, hardware overheating, and compute expense stand between this demo and any consumer product. What to watch next is whether OpenAI's robotics push builds on this direct-control approach — and whether the safety concerns raised in earlier studies get addressed before such systems leave the lab.
Comments
No comments yet. Be the first to comment.