Physical Intelligence

Contribution to AI

Language models had a recipe that robotics could not use: text is abundant, robot data is not, and every group wanting a manipulation policy trained one from scratch for its own arm and its own task. π0, shown in October 2024, broke that pattern by starting from a pretrained vision-language model and attaching an action expert trained by flow matching, then fitting it to roughly ten thousand hours of demonstrations gathered across seven robot configurations and sixty-eight tasks, topped up with public cross-embodiment data. Flow matching earned its place by emitting continuous action chunks at around fifty hertz instead of discretised tokens, which is what folding a shirt actually requires. The decision that carried furthest was releasing weights and training code under a permissive licence in February 2025, so a lab with one arm and a few hours of demonstrations could fine-tune a generalist rather than begin at zero; ports to other frameworks appeared almost immediately. FAST, a tokeniser built on the discrete cosine transform, was given away the same way, as a component others could drop into models of their own. π0.5 then tidied kitchens and bedrooms in houses absent from its training data, and improved as the number of distinct training homes rose — evidence that environmental variety, rather than more hours in one laboratory, is the binding constraint.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).