Physical Intelligence openpi

What the evidence says

1. Documented by company source

   Physical Intelligence released the weights and code for π0 and its π0-FAST model on February 4, 2025, and π0.5, which it describes as an upgraded π0 with better open-world generalization, in September 2025. The repository lists base models for fine-tuning and versions fine-tuned for the DROID, ALOHA and LIBERO robot setups.

   Observed Oct 18, 2026 · Physical Intelligence (company source), Physical-Intelligence/openpi (company source)
2. Documented by company source

   The openpi code is licensed under Apache 2.0. The repository also includes Google's Gemma Terms of Use, and its models are built on Google's PaliGemma; Gemma's terms cover models derived from Gemma, including its prohibited-use policy and conditions on redistribution. The repository does not name a separate licence for the weights.

   Observed Oct 18, 2026 · Physical Intelligence (company source)
3. Independently verified

   In the PhAIL real-robot benchmark (May 2026), π0.5 and NVIDIA's GR00T N1.6 tied for best among four open models on a bin-picking task (13.8% and 13.3% of a human operator's throughput, statistically indistinguishable); the best model was roughly seven times slower per operation than a person. The author's company runs the benchmark, and annotation was not blinded.

   Observed Oct 18, 2026 · arXiv (Positronic Robotics) (independent reporting)
4. Independently verified

   NVIDIA's Isaac Lab Arena documents running openpi's π0 and π0.5 DROID checkpoints in simulation through openpi's policy server, and notes that "VLA models are improving but not yet fully robust under zero-shot distribution shift."

   Observed Oct 18, 2026 · NVIDIA (independent reporting)

Price

• DownloadFreeFree. Google's Gemma Terms of Use apply to the models. Checked Oct 18, 2026. Physical Intelligence (company source)