Π*0.6 real robot that learns from experience via RL (and can make you a coffee) (www.physicalintelligence.company)

Researchers introduced Recap (RL with Experience & Corrections via Advantage‑conditioned Policies), a three-stage training recipe that turns a vision-language-action (VLA) model, π0.6, into an RL‑trained controller called π*0.6 that can reliably perform long-horizon real‑world tasks. Recap combines (1) demonstrations, (2) expert teleoperated corrections for the exact failure states the policy encounters, and (3) reinforcement learning from the robot’s own experience. Key to the method is learning a value function that predicts task progress (negative steps to completion) and conditioning the policy on the change in value (advantage). By keeping all data and annotating it with advantage signals, the policy can be trained to prefer high‑advantage actions and thus learn from “bad” trajectories rather than merely imitating them. The results are practical and striking: on hardest tasks like espresso making, throughput and success rates more than doubled and failure rates halved, enabling continuous, hours‑long runs—making espresso, folding novel laundry items, and assembling factory packaging boxes. Technically, Recap addresses compounding errors from imitation by providing targeted corrections and solving credit assignment via a learned value function and advantage‑conditioned policy extraction. The approach scales to large VLAs (π*0.6 is a refinement of π0.6) and demonstrates a viable path to robust, high‑throughput robotic manipulation in diverse, real‑world settings.