Physical Intelligence

weights

open(public GCS bucket)

data

closed(10k+ hours of robot data, not released)

code

open(training, fine-tuning and inference)

license

Apache-2.0(repository license

The weights and the full training code are public under Apache-2.0, but the more than ten thousand hours of robot data the base models learned from are not. LeRobot's Hub copies of the same checkpoints are marked with Google's Gemma terms, which Physical Intelligence's own repository does not mention.

• https://cdn.jsdelivr.net/gh/Physical-Intelligence/openpi/LICENSE recorded 2026-09-26

  LICENSE: the Apache License, Version 2.0 text.

• https://cdn.jsdelivr.net/gh/Physical-Intelligence/openpi/README.md recorded 2026-09-26

  README: base checkpoints pi0_base, pi0_fast_base and pi05_base and expert checkpoints download from gs://openpi-assets; "pre-trained on 10k+ hours of robot data" with no dataset named; fine-tuning guide and configs; no separate license for the checkpoints.

• https://huggingface.co/api/models/lerobot/pi0_base recorded 2026-09-26

  Hub API for LeRobot's port lerobot/pi0_base: license gemma.

• https://ungh.cc/repos/Physical-Intelligence/openpi/files/main recorded 2026-09-26

  Repository tree: training and fine-tuning scripts and configs in the repository.

GitHub stars on openpi. Physical Intelligence publishes its checkpoints from a cloud bucket that reports no download count, and the Hub copies belong to LeRobot.

• https://repos.ecosyste.ms/api/v1/hosts/GitHub/repositories/Physical-Intelligence%2Fopenpi recorded 2026-09-26

  Repository record for Physical-Intelligence/openpi: 13,932 stargazers.

π0 is pretrained across many robots and has released fine-tunes that fold towels and uncap pens on a bimanual ALOHA rig, and π0.5 is built to generalize to environments outside its training. Physical Intelligence itself warns that adapting it to a new platform may not work.

• https://cdn.jsdelivr.net/gh/Physical-Intelligence/openpi/README.md recorded 2026-09-26

  README: base models "pre-trained on 10k+ hours of robot data"; expert checkpoints π0-ALOHA-towel, π0-ALOHA-tupperware, π0-ALOHA-pen-uncap, π0-DROID, π0.5-LIBERO, π0.5-DROID; π0.5 "with better open-world generalization"; "we do not expect every such attempt to be successful".