Physical Intelligence π0

2024-10-31 · Publisher : Physical Intelligence · Domain : Robotics

Physical Intelligence unveils π0, a vision-language-action model that adds a flow-matching action head to a pretrained VLM to control very different robots: single arms, bimanual setups and mobile manipulators.

### Background On October 31, 2024, the Californian startup Physical Intelligence released π0 ("pi-zero"), its first robot foundation model. It is a vision-language-action (VLA) model: it takes images and a natural-language instruction and directly outputs motor commands. ### Specifications - **Architecture**: a flow-matching action head on top of a pretrained vision-language model, inheriting web-scale semantic knowledge. - **Robots**: single arms, dual-arm systems and mobile manipulators, with one model. - **Demonstrated tasks**: laundry folding, table clearing, box assembly. - **Uses**: zero-shot execution after pretraining, instruction following, new skills through fine-tuning. - **Publication**: arXiv 2410.24164, accepted at RSS 2025. Code and weights opened later, on February 4, 2025, in the openpi repository (Apache 2.0). ### Why it matters π0 was one of the first models to show that a single policy can chain long, dexterous tasks across different robots. It set the pattern many players followed in 2025: a VLM as the semantic "brain" plus a fast action expert. ### Key takeaway π0 marks the starting point of the race for generalist robot brains. Physical Intelligence then improved it with π0-FAST, π0.5, π*0.6 and π0.7.

Tags : VLA, Modèle de fondation, Manipulation, Flow matching, Physical Intelligence