π*₀.₆: a VLA That Learns From Experience

Fig. 1: RECAP enables VLAs to improve performance during deployment.Our model incorporates advantage conditioning, allowing it to ingest diverse types of real-world experience: human demonstrations, autonomous rollouts and online corrections.Learning from real-world interaction via advantage conditioning improves the model's understanding of how actions influence performance, leading to a faster and more successful policy.Abstract-Vision-language-action (VLA) models offer a promising path toward general-purpose robots, but achieving the reliability and speed required for practical deployment remains challenging.We present a general-purpose method, RL with Experience and Corrections via Advantage-conditioned Policies (RECAP) that improves the efficiency and reliability of VLA policies by utilizing their real-world experience.Our method introduces value-based advantage conditioning during both pretraining and post-training phases, enabling VLA policies to ingest highly heterogeneous real-world experience, including human demonstrations, policy rollouts, and online correction data.We show that the ω → 0.6 model, trained with RECAP, achieves hourslong deployment of folding diverse laundry in real homes, reliably assembles boxes in a factory, and makes espresso drinks using a professional espresso machine.On some of the hardest tasks, RECAP more than doubles task throughput and roughly halves the task failure rate.