Repolode
Submit
Physical-Intelligence

Physical-Intelligence/openpi

Open-source robotics vision-language-action (VLA) models from Physical Intelligence, with pre-trained checkpoints and fine-tuning examples.

View on GitHub
Stars
14,311
Forks
2,571
Open issues
363
Last commit
2 months ago

What is it?

openpi is a repository of open-source robotics models and packages published by the Physical Intelligence team. It provides three VLA model families: π₀, a flow-based model; π₀-FAST, an autoregressive model built on the FAST action tokenizer; and π₀.₅, an upgraded π₀ with better open-world generalization. Base checkpoints pre-trained on over 10k hours of robot data are available, along with fine-tuned checkpoints for platforms such as DROID and ALOHA.

The repo covers inference and fine-tuning. Users can run a checkpoint with a few lines of Python, serve a policy over websocket for remote inference, or fine-tune on their own data after converting it to the LeRobot dataset format. JAX is the original implementation, and a PyTorch implementation of π₀ and π₀.₅ is also included, with some features such as π₀-FAST and LoRA training not yet supported in PyTorch.

Hardware requirements are listed by mode: inference needs more than 8 GB of GPU memory, LoRA fine-tuning more than 22.5 GB, and full fine-tuning more than 70 GB. The project is tested on Ubuntu 22.04 only.

Key features

  • 01Three VLA model families: π₀, π₀-FAST, and π₀.₅
  • 02Base checkpoints pre-trained on over 10k hours of robot data
  • 03Fine-tuned checkpoints for DROID and ALOHA platforms
  • 04Remote policy serving over websocket
  • 05JAX and PyTorch implementations for π₀ and π₀.₅

Use cases

  • Running pre-trained robot policies on DROID or ALOHA hardware
  • Fine-tuning a base VLA model on your own robot dataset
  • Offloading policy inference to a remote GPU server

Quick start

  1. 1

    Clone the repo with submodules

    git clone --recurse-submodules git@github.com:Physical-Intelligence/openpi.git
  2. 2

    Install uv, then set up the Python environment

    GIT_LFS_SKIP_SMUDGE=1 uv sync
  3. 3

    Install the openpi package in editable mode

    GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .
  4. 4

    Run inference with a pre-trained checkpoint using the Python example in the README (requires an NVIDIA GPU)

Things to know

  • Requires an NVIDIA GPU; inference needs more than 8 GB of memory.
  • Officially tested only on Ubuntu 22.04; other operating systems are not supported.
  • The π₀ models were developed for the authors' own robots and may not work on your platform.

Topics

#robotics#vision-language-action#imitation-learning#jax#pytorch#fine-tuning

FAQ

What is openpi?
Open-source robotics vision-language-action (VLA) models from Physical Intelligence, with pre-trained checkpoints and fine-tuning examples. openpi is a repository of open-source robotics models and packages published by the Physical Intelligence team. It provides three VLA model families: π₀, a flow-based model; π₀-FAST, an autoregressive model built on the FAST action tokenizer; and π₀.₅, an upgraded π₀ with better open-world generalization. Base checkpoints pre-trained on over 10k hours of robot data are available, along with fine-tuned checkpoints for platforms such as DROID and ALOHA.
Is openpi free to use? What license does it have?
openpi is open source under the Apache-2.0 license. Check the license terms before using it in commercial products.
Is openpi still maintained?
Updated within 6 months. The last commit to openpi was on Aug 24, 2026.
What can I use openpi for?
Running pre-trained robot policies on DROID or ALOHA hardware Fine-tuning a base VLA model on your own robot dataset Offloading policy inference to a remote GPU server
How do I install openpi?
Clone the repo with submodules: git clone --recurse-submodules git@github.com:Physical-Intelligence/openpi.git Install uv, then set up the Python environment: GIT_LFS_SKIP_SMUDGE=1 uv sync Install the openpi package in editable mode: GIT_LFS_SKIP_SMUDGE=1 uv pip install -e . Run inference with a pre-trained checkpoint using the Python example in the README (requires an NVIDIA GPU)
Where can I get openpi?
The source code and installation instructions are on GitHub: https://github.com/Physical-Intelligence/openpi

Summary generated by AI from the project's README. GitHub stats refresh daily.

Related projects