Physical Intelligence

What Physical Intelligence is

Physical Intelligence, whose name is often written with the Greek letter π, describes itself as a group of engineers, scientists, roboticists and company builders developing foundation models and learning algorithms to power robots^[1]. Its stated aim is to bring general-purpose AI into the physical world^[1].

That description places the company in a particular corner of robotics. The usual way to make a robot do something is to write a program for that machine and that task; change the robot, the objects or the room, and the program often has to be rewritten. The approach implied by "generalist policy" is different: train one large model on data from many robots and many kinds of manipulation, then let that model control new robots in new settings. A policy, in this context, is simply the mapping from what the robot senses to what it does next.

A timeline of published work

The company's first generalist policy, π0, was developed over the eight months before it was published on 31 October 2024^[2]. Publishing the policy itself, rather than only a description of a product, became a pattern: much of what is publicly known about the company comes from a steady sequence of technical releases.

On 16 January 2025 the company published FAST, a robot action tokenizer, which it said trained generalist policies five times faster than previous models^[1]. A tokenizer in this setting turns a robot's continuous actions into discrete pieces a model can predict one at a time, in much the way a language model predicts words.

On 4 February 2025 the company released the weights and code for π0 and for a π0-FAST autoregressive model^[1]. Releasing weights and code means other researchers can download the trained model and run it themselves, instead of only reading about it.

On 22 April 2025 came π0.5, which the company said extends π0 and enables open-world generalization^[1][3]. "Open-world" refers to the hope that a policy trained in some environments will still work in rooms and with objects it has not seen before, rather than only in the setting it was trained on.

On 9 June 2025 the company published a real-time action chunking system for large vision-language-action models^[1]. Vision-language-action models, usually shortened to VLA, combine visual input, language instructions and motor output in one network; action chunking means predicting a short sequence of future actions at once, which helps a large model keep up with the speed a robot needs.

Later releases moved toward learning from feedback and from people. On 17 November 2025 the company published π*0.6, describing a method for training generalist policies with reinforcement learning^[1]. Reinforcement learning is training by reward: the model tries things, and behaviour that leads to better outcomes is made more likely. On 16 December 2025 the company published work on human-to-robot transfer in vision-language-action models^[1], a line of research concerned with using data collected from people to improve what robots can do.

The team

The company's team page lists Chelsea Finn, Sergey Levine and Karol Hausman among its staff^[1].

Where it sits

What distinguishes the company's public record is the combination of two things: a research programme aimed at general-purpose robot learning rather than single-task automation, and a habit of releasing models, tokenizers and weights so that others can inspect and build on them^[1][2]. The sequence from π0 to π0.5 and π*0.6, with FAST and action chunking in between, shows a group working through the practical parts of that programme in turn: how to represent actions, how to train faster, how to generalize beyond the training environment, how to run in real time, and how to improve behaviour after initial training^[1][3].