Physical Intelligence

Excited to share one of the projects we've been working on at Physical Intelligence that directly unlocks deploying robots into real production environments. For robots to work in manufacturing, they need to be fast, accurate, and precise, all at once, and especially in the moments that matter most... The broad approach to a task is tractable. The hard part is what happens at contact in fine manipulation. These are the moments where small errors in position or force amplify into failures, where the robot slows to compensate, and where a hype demo can be separated from a real deployment. We developed a method called RLT (Reinforcement Learning Tokens) that improves exactly these contact-critical phases without retraining the full model. The VLA exposes a compact RL token, a summary of its internal state, and a tiny actor-critic trains on top of it directly on the robot. The robot practices the hardest phase of a task and gets both faster and more reliable after each attempt. We tested on four canonical tasks: screwdriving, zip tie fastening, ethernet insertion, and power cord insertion, seeing up to 3x improvement in speed and reliability across all four. On ethernet insertion, the RL policy ran at twice the speed of expert human teleoperation after just two hours of practice on the real robot. What this unlocks for manufacturing: robots don't need to arrive perfect. They can deploy with broad competence and refine the speed, accuracy, and precision of the hardest steps in place, adapting to the specific fixtures and tolerances of their actual environment. The robot learns on the job, and the deployment site becomes part of the training loop. Had a lot of fun being part of this one. Blog and paper: https://lnkd.in/gHVHQPgG