openpi
Project record
Maintained by Physical Intelligence1
openpi is Physical Intelligence's repository of code and model checkpoints for its π0, π0-FAST, and π0.5 vision-language-action models for robot control. It provides base checkpoints pre-trained on more than 10,000 hours of robot data, fine-tuned checkpoints for the ALOHA, DROID, and LIBERO setups, and JAX and PyTorch code for inference and fine-tuning.1
- Repository: Repository (external site: github.com)
- License: License (Apache 2.0) (external site: github.com)
- Paper: π0 paper (external site: arxiv.org)
Availability and license
Overall availability
Code is on GitHub. Checkpoints download from a public Google Cloud Storage bucket (gs://openpi-assets), which the code fetches automatically.168
Availability is separate from permission: read the license before using or redistributing.
The repository's LICENSE file is Apache 2.0. The repository also contains Google's Gemma Terms of Use (LICENSE_GEMMA.txt, added September 2025) without saying which files or checkpoints it covers; the π0 paper says the model uses PaliGemma as its base. The README does not state separate license terms for the checkpoints.23471
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| Source codeIs the source code publicly readable? | Public | Model, training, and serving code in JAX, with PyTorch versions of π0 and π0.5.1 |
| DocumentationIs user documentation published? | Public | The README and docs folder cover installation, inference, fine-tuning, remote inference, and troubleshooting.1 |
| Training codeDoes it include code for training models? | Public | Fine-tuning scripts and configs are included, plus an approximate open-source version of the pipeline used to train π0-FAST-DROID. The README does not describe releasing the base-model pre-training pipeline.1 |
| Data informationAre the data it expects or ships with documented? | Partial | Fine-tuning examples use public LIBERO and DROID data. The π0 paper describes the pre-training mixture (mostly internal robot data plus OXE, Bridge v2, and DROID); no public release of the internal data was found.17 |
| ReproducibilityAre instructions for reproducing reported results published? | Partial | Step-by-step instructions cover LIBERO fine-tuning and evaluation and DROID training; reproducing the base-model pre-training is not covered.1 |
What it is useful for
The README documents fine-tuning the base models on a user's own robot data (converted to the LeRobot dataset format), serving a policy over a websocket for remote inference, and running the provided checkpoints on DROID and ALOHA robots. It describes the release as an experiment that may not transfer to every robot.1
Run and use notes
- The README estimates, assuming a single NVIDIA GPU, more than 8 GB of GPU memory for inference (example RTX 4090), more than 22.5 GB for LoRA fine-tuning, and more than 70 GB for full fine-tuning (A100 80GB or H100). It states that JAX inference keeps most weights in bfloat16 and training defaults to mixed precision (float32 weights, mostly bfloat16 activations). The repository is tested only on Ubuntu 22.04.1
Organization context
Provenance and derivatives
The π0 paper states that π0 uses PaliGemma, an open 3-billion-parameter vision-language model, as its base model and adds a flow-matching architecture for actions; the openpi model code pairs a PaliGemma backbone with a separate action expert.75
- Derived from: PaliGemma — Vision-language base model named in the π0 paper.
U.S. eligibility
Eligible · basis: U.S.-governed project
The repository is published by the Physical Intelligence team, per its README. The π0 paper gives the authors' affiliation as Physical Intelligence, San Francisco, California, USA, and The Robot Report describes the company as San Francisco-based; the physical-intelligence record documents its headquarters.178
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.