Physical Intelligence unveils π*0.6 VLA, which learns from autonomous experience

Physical Intelligence on Nov. 17, 2025, announced its VLA model π*0.6. The company developed a new training method called Recap and said reinforcement learning using autonomous experience more than doubled throughput on the most challenging tasks and reduced the failure rate by more than half. In demonstrations, the robot made espresso from 5:30 to 23:30, folded 50 types of previously unseen laundry without interruption for several hours in a new home, and successfully assembled and labeled 59 cardboard boxes for chocolate packaging at a real factory.

The announcement directly addresses a challenge in real-world robotics: learning from demonstrations alone can leave robots prone to stopping when small operating errors compound, making long, continuous tasks and adaptation to new objects difficult. Physical Intelligence said its training method uses the robot’s own attempts, in addition to human corrections. The demonstrations showed it could continue without interruption through tasks under varying conditions, including handling 50 types of previously unseen laundry in a new home and assembling and labeling 59 packaging boxes at a real factory.