π0.5 (pi-zero-point-five)

What π0.5 is

π0.5 is a vision-language-action model published by Physical Intelligence on 22 April 2025, which the company describes as a model with open-world generalization ^[1]. A vision-language-action model (VLA) is a single neural network that reads camera images and a written instruction and outputs the motor commands a robot needs in order to act. The vision part reads the scene, the language part reads what is being asked, and the action part turns both into movement. π0.5 extends π0, the company's earlier model of the same family ^[1].

Why it was built

The company says the goal of π0.5 was generalization to new settings rather than new skills or high dexterity ^[1]. That distinction matters for how the work should be read. A robot that can perform one impressive trick in one laboratory is a demonstration; a robot that can be dropped into a room it has never seen and still do something useful is a different problem. Rooms differ in layout, in where objects are kept, in lighting and clutter, and in the small accidents — a spilled box, a chair pushed out of place — that no training set enumerates in advance. π0.5 was aimed at that second problem.

Co-training on heterogeneous data

Physical Intelligence says the main principle behind π0.5 is co-training on heterogeneous data ^[1]. Co-training means training one model on several kinds of data at the same time rather than on a single uniform dataset. For π0.5 the mixture included robot action data, verbal instructions, subtask commands, object detection annotations and multimodal web data ^[1]. These sources contribute different things. Robot action data ties perception to movement. Verbal instructions and subtask commands connect behaviour to language, including language at a coarser grain than a single motion. Object detection annotations attach names to what is visible. Multimodal web data brings in general knowledge about how objects and scenes tend to look and what words mean, without any robot involved.

On the robot side, the company says it trained π0.5 on data from in-the-wild mobile robots, in-the-wild static robots, in-office static robots and general robot data ^[1]. The spread is deliberate: robots that move through real environments see scenes that fixed laboratory arms never do, while fixed robots contribute large amounts of steadier manipulation experience.

Two levels of prediction

The model works at two levels at once. Physical Intelligence says π0.5 makes high-level inferences about the next semantic step, and also makes low-level predictions that produce motor commands for the robot's joints ^[1]. In practice this means deciding something like which part of the task comes next, then translating that decision into the continuous signals that drive the arm, the gripper and the base. Splitting the problem this way lets a language-shaped plan and a physically grounded controller live in the same model.

What it can do

The company said π0.5 can control a mobile manipulator to clean up an entirely new kitchen or bedroom ^[1]. It is not presented as reliable. Physical Intelligence says the model does not always succeed on the first try, but often shows a hint of the flexibility and resourcefulness a person would bring to a new challenge ^[1].

Where it sits

π0.5 is best understood as a step in the direction of robots that work in ordinary, uncontrolled spaces rather than a finished product. Its contribution is the claim that mixing robot experience with language, annotation and web data lets one model carry what it has learned into rooms it has never entered, and the acknowledgment that this generalization is still partial.