Ï€0 (Pi-Zero) is a robot foundation model released in late 2024 by Physical Intelligence, a San Francisco AI startup focused on building general-purpose physical intelligence for robots. It uses a PaliGemma vision-language model (~3 billion parameters) as its backbone, augmented with a separate action expert that produces continuous robot actions via flow matching — a diffusion-style technique that allows smooth, high-frequency control up to 50 Hz. Unlike token-based VLA models such as OpenVLA and RT-2, this continuous-action design enables the dexterity needed for tasks like folding laundry, table bussing, and grocery bagging. Ï€0 was pre-trained on more than 10,000 hours of data spanning seven robot platforms and 68 tasks, plus public datasets like Open X-Embodiment. The model has since been ported to Hugging Face’s LeRobot library and open-sourced (with Ï€0-FAST and successors Ï€0.5 and Ï€*0.6 also released), making it one of the most accessible high-performance generalist policies currently available. Demonstrated robots include single-arm and bimanual ARX, Franka, Trossen WidowX, and mobile manipulators.
General-purpose ~3.3B-parameter Vision-Language-Action flow-matching model from Physical Intelligence, built on a PaliGemma VLM backbone. Outputs continuous action chunks at up to 50 Hz, enabling dexterous manipulation across many robot embodiments.
Ï€0 (Pi-Zero) is a robot foundation model released in late 2024 by Physical Intelligence, a San Francisco AI startup focused on building general-purpose physical intelligence for robots. It uses a PaliGemma vision-language model (~3 billion parameters) as its backbone, augmented with a separate action expert that produces continuous robot actions via flow matching — a diffusion-style technique that allows smooth, high-frequency control up to 50 Hz. Unlike token-based VLA models such as OpenVLA and RT-2, this continuous-action design enables the dexterity needed for tasks like folding laundry, table bussing, and grocery bagging. Ï€0 was pre-trained on more than 10,000 hours of data spanning seven robot platforms and 68 tasks, plus public datasets like Open X-Embodiment. The model has since been ported to Hugging Face’s LeRobot library and open-sourced (with Ï€0-FAST and successors Ï€0.5 and Ï€*0.6 also released), making it one of the most accessible high-performance generalist policies currently available. Demonstrated robots include single-arm and bimanual ARX, Franka, Trossen WidowX, and mobile manipulators.