Blog
Chapter 3: An agentic robotics fable
We gave frontier models our real-robot API and let them write, test and fix their own policies. Thirteen findings on where these agentic policies beat end-to-end policies, where they fall short, and how the frontier models compare.
Chapter 2: Towards an agent-native CERN of AI robotics
We gave AI agents our datasets, training infrastructure and robot farm through an API, and challenged them to reproduce the core claims of ten recent robotics papers on real robots.
Chapter 1: The Reality Check
PAW-GEN-10 evaluates four open robot policies in ten real-world environments, 3,600 reproducible rollouts per policy, to measure how task-specific data and object placement shape success.