Physical Intelligence π*0.6 and Recap
2025-11-18 · Publisher : Physical Intelligence · Domain : Robotics
Physical Intelligence unveils π*0.6 and the Recap method, letting a VLA improve through reinforcement learning from its own deployments and human operator corrections.
What changes
- Recap: reinforcement learning through advantage-conditioned policies.
- Mixes demonstrations, autonomous data and teleoperated corrections.
- Folds laundry in real homes, assembles boxes, makes espresso.
- Throughput more than doubled and failures roughly halved on the hardest tasks, per the authors.
### Background On November 18, 2025, Physical Intelligence posted on arXiv its work on π*0.6 ("pi-star 0.6") and the Recap method (RL with Experience and Corrections via Advantage-conditioned Policies). The question: how can a vision-language-action model keep improving once deployed in the real world? ### Specifications - **Recap method**: reinforcement learning adapted to VLAs through policies conditioned on the "advantage" (the gap between an action's outcome and the expected outcome). - **Heterogeneous data**: demonstrations, autonomously collected data and expert teleoperated interventions during execution. - **π*0.6**: a generalist VLA pretrained with offline RL, then specialized through on-robot data collection. - **Tasks**: folding laundry in real homes, reliably assembling boxes, making espresso with a professional machine. - **Results**: on some of the hardest tasks, throughput more than doubles and the failure rate roughly halves (results published by the authors). ### Why it matters Imitation alone plateaus. Recap shows a robot can learn from its own experience and human corrections, paving the way for robots that improve in production. ### Key takeaway Physical Intelligence adds learning from experience to its model lineage; π0.7 followed in April 2026.
Links
Tags : VLA, Apprentissage par renforcement, Physical Intelligence, Déploiement réel