Physical Intelligence

Physical Intelligence develops AI foundation models that let different kinds of robots follow instructions and perform physical tasks. Its π model family combines vision, language and robot actions, supporting work such as folding laundry, clearing tables and assembling boxes. Robotics teams can prompt models for particular tasks or fine-tune them for specialized applications. Its original π0 model uses flow matching to turn a pretrained vision-language model’s understanding into continuous motor commands.

Founded in 2024, the company’s founders are CEO Karol Hausman, Sergey Levine, Brian Ichter, Chelsea Finn, Adnan Esmail, Quan Vuong and Lachy Groom. Its research addresses how to train robot control models across hardware and tasks. FAST, its action tokenizer, compresses robot movement sequences using discrete cosine transforms and byte-pair encoding, allowing transformer models to learn robot control through next-token prediction. The company released a tokenizer trained on one million real robot action sequences for researchers to train their own policies.

Physical Intelligence works with robot companies to adapt its models to their hardware and operating environments. In 2026, partners documented live deployments: Weave used its models for laundry folding at a San Francisco laundromat, while Ultra used them to package customer orders in warehouses. Ultra’s workflow combines model-driven operation with human intervention when needed, generating additional training data from those deployments.