π*0.6 (pi-star-zero-point-six)

What it is

π*0.6 is a robot model released by Physical Intelligence on 17 November 2025 ^[1]. It is a vision-language-action model: it takes in camera images and a language instruction and produces the motor commands that drive the robot. What separates it from a purely imitation-trained model is where its training data comes from — not only human demonstrations, but also human corrections and the robot's own autonomous experience ^[1]. The published model is the company's π0.6 trained with a method called Recap ^[1].

The problem it targets

Physical Intelligence frames the motivation in terms of consistency rather than capability. With imitation learning alone, the company says, making a robot succeed half the time is not hard; making it succeed every time is very hard ^[1]. The gap comes from compounding errors. A small mistake moves the robot into a state that was never in its training data, so there is nothing there for it to imitate, and the next mistake is bigger than the first ^[1]. A demonstration dataset covers the trajectories a person happened to show; it does not cover what the robot does after it drifts away from them. Tasks that run for hours make this costly, because every extra minute is another chance to drift.

How Recap works

Recap stands for "RL with Experience & Corrections via Advantage-conditioned Policies" ^[1]. It puts three kinds of data together. Demonstrations show roughly what the task looks like. Corrections are the interventions a person makes when the robot goes wrong. Autonomous experience is what the robot produces by attempting the task itself ^[1]. The model is therefore coached as well as shown.

The "advantage-conditioned" part describes how that mixture becomes a training signal. In reinforcement learning, the advantage of an action measures how much better it is than what the policy would do on average. Feeding that value to the policy as an input condition means the network can be asked for actions of a given quality rather than for its average behaviour, so it learns to prefer what worked and to set aside what did not. The practical effect is that demonstrations, corrections and self-collected experience can be used together inside one model instead of being kept apart.

Results

The company reports three long-running tests ^[1]. The model made espresso drinks from 5:30 in the morning to 11:30 at night without interruption ^[1]. It folded 50 different novel laundry items in a new home ^[1]. It assembled and labelled 59 boxes used for packaging chocolates, in a real factory ^[1]. All three tasks, the company says, are performed with over a 90% success rate ^[1]. Training with Recap also more than doubles throughput on some of the hardest tasks and can decrease failure rates by 2 times or more ^[1].

Where it sits

The release is an argument about data rather than about hardware. A robot that only imitates is bounded by the situations someone demonstrated; adding corrections and self-collected experience extends what it has seen to the states it actually gets into while running. The three demonstrations — an espresso station, a home laundry pile, a factory packaging line — are chosen to show the same method over very different objects and settings, and to show it holding up over hours rather than over a single attempt ^[1]. What π*0.6 claims is not that a robot can do these jobs once, but that the same policy can keep doing them while the mistakes that would normally accumulate are held down ^[1].