EEmbodied AI Hub
ProjectFeaturedAdvanced

openpi

Open training and inference code around π0 from Physical Intelligence for generalist robot policies.

Why: Frontier open implementation of generalist robot policies

Official / sourcePaperBack to Project
VLAImitation learningOpen sourceReal robot

Overview

openpi is the open training and inference codebase around Physical Intelligence’s π0 generalist robot policies. It is the practical entry for teams that want to run or finetune π0-style models rather than only reading the blog post.

Expect heavier GPU requirements than tiny ACT models, careful data formatting, and a workflow closer to foundation-model finetuning than classic IL scripts. Pair with the π0 model card for weights and license terms.

Who it is for

Teams finetuning generalist VLA/policy models; labs following PI releases.

Key highlights

  • Official-adjacent code for π0-style policies
  • Training and inference entrypoints
  • Bridge between research release and reproducible runs

When to use

  • You want to finetune π0 on your robot data
  • You study generalist policy transfer

When not to use

  • Beginner first project with one GPU and no VLA experience

Getting started

  1. 1Read the π0 model card and license before downloading weights.
  2. 2Set up openpi env and run inference on a sample episode.
  3. 3Prepare your dataset in the expected format; start with a small finetune.
  4. 4Evaluate on held-out real tasks, not only training loss.

Install / setup

git clone https://github.com/Physical-Intelligence/openpi.git

Papers & reading

How it compares

Caveats & pitfalls

  • Hardware and license constraints can block commercial use.
  • APIs may lag blog announcements.

Content reviewed 2026-07-23