A humanoid in the kitchen and Stanford's shortcut: controlling a robot without action models

In September 2026 Stanford's TML lab presented HomeBody (a Caltech researcher is also among the authors), a research system that hands control of a humanoid robot directly to a frontier vision-language model. As documented on the project page and reported by The Decoder on 27 September, the approach skips the intermediate layer usually filled by a trained action model (VLA). Instead, the OpenAI model — called "GPT Astra" on Stanford's page and "GPT-6 Astra" by The Decoder — directly calls a library of five existing motor skills, coordinating map navigation, picking and placing objects, opening drawers and grasping from an open drawer.

Before acting, the Unitree G1 explores its surroundings to build a digital twin in Nvidia Isaac Sim through a real2sim process, storing the 3D position of objects in a spatial memory. This internal map lets it retrieve objects even when they leave its immediate field of view. Onboard computing runs on a laptop with a single RTX 4090 GPU, which handles arm and hand control at 250 Hz and AMO policy updates at 50 Hz; the OpenAI model, by contrast, is queried via API. In the qualitative demonstration released by the researchers, the robot tidied a kitchen it had never seen before from a generic request, with no training specific to that setting.

The system nonetheless shows the constraints typical of technologies that are not yet mature. The authors point out that GPT Astra's reasoning latency introduces pauses between skills, and that overall task duration is also limited by the humanoid's reach, its manipulation capabilities and hardware endurance, with particular reference to overheating in the finger servomotors. Reconstructing the environment also takes setup time and involves API costs that have not been quantified. At the time of checking, the code on the official GitHub repository is not yet public, no licence is stated, there is no linked arXiv paper, and it is unclear whether the test was repeated in more than one kitchen. No statistics on success rate or number of trials are available; the 95% figure circulating online does not concern HomeBody: it appears to come from a third-party evaluation of I2RT YAM robotic arms, which we could not verify against a primary source.

Skipping the intermediate motor translation and relying directly on a frontier intelligence is a fascinating experiment, but it reminds us that physics runs on a different clock from pure computation. If the action has to freeze while waiting for the model to answer over an API, fluid movement is still a distant goal. — Olya