π0 (pi-zero)

π0 (pi-zero) is Physical Intelligence's first generalist robot foundation model, released on October 31, 2024. It is a vision-language-action (VLA) model that combines internet-scale vision-language pretraining with robot action outputs via flow matching (a variant of diffusion models). π0 is trained on the largest robot interaction dataset to date, including open-source data from the Open X-Embodiment dataset and Physical Intelligence's own diverse dataset of dexterous tasks across 8 distinct robot types. The model inherits semantic knowledge from a pre-trained 3 billion parameter VLM and adapts it for real-time dexterous robot control at up to 50 Hz. The model can perform a wide variety of zero-shot tasks including folding laundry, bussing tables, bagging groceries, packing items into envelopes, routing cables, assembling boxes, plugging in power plugs, and picking up trash. It can be directly prompted for desired tasks or fine-tuned for specialized applications. π0 was open-sourced in February 2025 with weights and code released. Subsequent versions include π0.5 (open-world generalization for mobile manipulators), π*0.6 (RL-improved), and π0.7 (steerable model with emergent capabilities). The paper describes π0 as a "first step toward artificial physical intelligence" — enabling users to simply ask robots to perform any task, similar to how LLMs and chatbot assistants work.