Physical Intelligence · 2024
π0: one policy, many robot bodies
π0 (pi-zero) is a generalist vision-language-action model from Physical Intelligence. Its bet is a single foundation model that transfers across different robots — and a flow-matching action head that produces smooth, dexterous motion.
What it is
A backbone plus an action expert
π0 takes a pre-trained vision-language model — which supplies broad semantic understanding — and attaches a flow-matching action expert that generates continuous motor commands at high frequency. The VLM knows what; the action expert produces how, smoothly.
Crucially, π0 is trained on data from many robot embodiments and a wide range of tasks. The goal is generalization: a single policy that a new robot can inherit, rather than a bespoke controller trained from scratch for each machine.
Why it matters
Toward a robot foundation model
π0 is one of the clearest expressions of the "foundation model for robots" thesis: pre-train once, transfer everywhere. Its focus on dexterous, real-world tasks — laundry, table clearing, packing — pushes VLA capability toward the messy manipulation that real deployments demand. Follow-up work (π0.5) extends this toward open-world generalization.
A generalist policy still starts blank about your world. The skill transfers; the context does not. That context — your home, your people, your way of working — is the knowledge layer benned builds on top of the action model.
FAQ
Common questions
What is π0 (pi0)?
A generalist vision-language-action model from Physical Intelligence (2024). It pairs a pre-trained vision-language model with a flow-matching action expert for high-frequency control, trained across many robot embodiments.
What is flow matching in π0?
A generative technique that produces smooth, continuous action trajectories at high frequency — well suited to dexterous fine-motor tasks — instead of emitting discrete action tokens.
Related
Last updated: July 2026