Physical AI Lab

Developing capable autonomous systems with less data and computation.

We investigate how to make progress in physical AI without relying on proprietary datasets or vast compute budgets. Our research develops efficient algorithms for robot learning, including methods that learn through interaction inside world models. We work closely with KE:SAI, Kyutai, and the Empirical Inference department. By releasing open datasets, models, and reproducible training recipes, we aim to lower the resources required for physical AI research and enable a broader community to contribute to the research frontier.