Pi-0.5

Pi-0.5 is Physical Intelligence's follow-up to its Pi-0 robot foundation model, shifting the focus from dexterity to open-world generalization. The vision-language-action model co-trains across heterogeneous tasks, multiple robots, high-level semantic prediction, and web data, enabling mobile manipulators to clean kitchens and bedrooms they never saw during training.

The model was published in April 2025 with weights released through the openpi project, making it one of the few frontier robot policies with publicly available checkpoints. It builds on a PaliGemma backbone with a flow-matching action expert, and later work added a RECAP reinforcement-learning recipe that the company reports doubled task throughput.