Physical Intelligence releases π0, a generalist robot policy
Trained across eight robot platforms and tasks including laundry-folding and box assembly, using flow matching to output motor commands up to 50 times a second.
- Models & capabilities
- Notable
Physical Intelligence, a robotics startup, unveiled π0 (pi-zero), a vision-language-action model intended to work as a single generalist controller across different robot bodies rather than being retrained for each new hardware platform or task. The model took in camera images and natural-language instructions and output low-level motor commands directly, at rates up to 50 times per second.
π0’s training combined three sources: internet-scale vision-language pretraining, existing open-source robot datasets, and Physical Intelligence’s own proprietary data collected across multiple robot platforms. The architecture used flow matching, a technique related to diffusion models, to generate continuous streams of motor commands while retaining the semantic understanding learned from web-scale pretraining. The company demonstrated the model on eight distinct robot configurations, including single-arm and bimanual UR5e and Franka arms, Trossen bimanual and mobile platforms, and a mobile Fibocom system, performing tasks such as folding laundry, bussing a table, assembling a box, making coffee, bagging groceries, loading a dishwasher and retrieving toast from a toaster.
Physical Intelligence said π0 outperformed comparable academic open-source models, including OpenVLA and Octo, on the tasks tested, and reported that models incorporating the full vision-language-model pretraining stage more than doubled performance over smaller variants trained without it. The October announcement was a research release rather than an open one: no weights or code accompanied it.
π0 became the foundation for Physical Intelligence’s subsequent work. The company open-sourced the model’s weights and code the following February, and went on to raise a $600 million Series B in late 2025 built on the same model family, part of a wider wave of venture investment treating general-purpose robot foundation models as a category distinct from both humanoid hardware and general-purpose language models.