Tier 3 Practical open-weight
What the publisher claims: open-source, open weights
The one-liner: Open code and weights for a robot policy, with Gemma-licensed weights and closed pretraining data.
The six questions
Base and fine-tuned checkpoints are downloadable from the project public bucket.
Code is Apache 2.0, but the weights derive from PaliGemma and carry Gemma terms, so redistribution conditions are less clean than the repo licence implies.
Runs on a single GPU with 8GB or more, and fine-tuning scripts are included.
The paper documents the architecture and action expert, but the 10k+ hours of proprietary robot pretraining data are not released.
Fine-tuning on your own data is supported, but the base model itself cannot be reproduced.
Physical-Intelligence/openpi shows 304 merged PRs, all from external contributors — a remarkably open research release under corporate direction.
License
Repository code Apache 2.0; checkpoints are PaliGemma and Gemma derivatives and carry Gemma terms rather than plain Apache 2.0 — license source ↗
- Weight terms are set by the Gemma licence, not the repo licence
- The release is presented by the publisher as an experiment
What is actually released
- code in JAX and PyTorch
- base and fine-tuned checkpoints
- fine-tuning code
- inference examples and policy server
Why it matters
Open robotics foundation models are real; the binding constraints are data provenance and backbone licensing, not access.
Sources
- https://huggingface.co/api/models/lerobot/pi0_base
- https://raw.githubusercontent.com/Physical-Intelligence/openpi/main/LICENSE
- https://raw.githubusercontent.com/Physical-Intelligence/openpi/main/README.md
- https://www.pi.website/blog/openpi
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.