The robot running GPT-6 Astra cleans up a strange kitchen
Stanford / Caltech team has integrated GPT-6 Astra model with Unitree G1 robot. No pre-trained control policy was used to clean up an unknown kitchen.
A team from Stanford and Caltech plugged their GPT-6 Astra model straight into a Unitree G1 robot. They call the whole setup HomeBody. It explored a kitchen it had never seen before, built a 3D map of that kitchen inside Nvidia Isaac Sim (a simulator that lets the robot "rehearse" moves safely), cleared stuff off a table, and pulled the right items out of drawers. No one trained a separate control program just for that kitchen — instead the system picks from a library of ready skills (grab, move, open a drawer) and decides which one to use on the fly.
The team is upfront about the downsides: the model is slow, the finger motors overheat, and the compute bill is steep. That matches what I see every day working with AI agents — the more general your control model is, the more flexible it gets, but the slower and more expensive it runs in the moment. What I like is the direction: instead of retraining a robot for every new kitchen, you give a general model access to a skill library and let it pick. My agents in Claude Code work on a similar idea. This is still a research project with code on GitHub, not a product, but the direction is the right one.
Source: the-decoder.com
Free course
Stop reading about AI — start building with it
The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.
Start free →
Author
Evgenii Arsentev
PhD · Founder & CEO, Dad4Kids
Articles · Latest articles