Learning policies across robot bodies
Physical Intelligence develops models that turn images, language instructions and robot state into motor actions. Its π0 work combines a pretrained vision-language backbone with a flow-matching action model, trained across robot platforms and then adapted to particular tasks. The contribution is a reusable learned policy rather than a single robot design. [1]
The company released π0 code, weights and adaptation examples in February 2025. These are experimental tools that require hardware integration and evaluation. The openpi repository's code license and its separate Gemma terms should also be distinguished; a code release does not make every upstream model component unrestricted. [2][3][4]
Research results and hosted access
The April 2026 π0.7 research investigated composing skills for tasks absent from training. The model accepts richer context, including language, visual subgoals and task annotations. Its reported generalization comes from the company's experiments; the research post and paper do not by themselves establish downloadable weights. [5]
An October 2026 API beta offered another route: customers upload robot data, fine-tune on managed infrastructure and access models through hosted inference. Early access and application-based beta availability are narrower than a generally available service for every robot. [6]
Sources
- π0: Our First Generalist Policy
Physical Intelligence · Oct 31, 2024
- Open Sourcing π0
Physical Intelligence · Feb 4, 2025
- Physical-Intelligence/openpi
Physical Intelligence · retrieved Oct 23, 2026
- Gemma Terms of Use
Google · Feb 21, 2024
- π0.7: a Steerable Model with Emergent Capabilities
Physical Intelligence · Apr 16, 2026
- Frontier Models for Every Robot: Announcing our API beta
Physical Intelligence · Oct 13, 2026