Robot Running GPT-6 Astra Cleans Up an Unfamiliar Kitchen

Researchers at Stanford and Caltech connected the GPT-6 Astra model directly to a Unitree G1 robot. They call the resulting system HomeBody. The robot looked around an unfamiliar kitchen on its own. It built a digital copy of the kitchen in Nvidia Isaac Sim. Then it cleared items off the table and pulled the right objects out of drawers. Nobody had trained a separate control layer for this specific kitchen beforehand. Instead of that layer, the model calls on a ready-made set of skills — grasping, moving, opening drawers — and decides which one to use at each moment.

The authors are upfront about the downsides: the model itself is slow to respond, the finger servos overheat, and the compute cost is high. That matches what I see working with agents every day. The less task-specific the layer, the more flexible the system, but the slower and more expensive it gets in the moment. I like the underlying idea here. Instead of retraining the robot for every new kitchen, you let a general model call ready-made skills directly. That is the same principle my agents in Claude Code run on. For now this is a research prototype with open code on GitHub, not a finished product, but the direction is sound.