The project tests whether a general language model can drive a robot's low-level actions without a purpose-trained control layer. If a model can call skills like grasping and navigation directly, that changes how robot control stacks are built. The Decoder's report describes one kitchen-tidying demonstration, so it is a single result rather than proof the approach generalizes.
Stanford and Caltech put GPT-6 Astra in a humanoid to tidy an unfamiliar kitchen
The Decoder reports Stanford and Caltech researchers had a humanoid robot powered by GPT-6 Astra clean an unfamiliar kitchen using HomeBody, a system where the language model calls grasping and navigation skills directly instead of relying
