openpi (Physical Intelligence)
Physical Intelligence's open-source code and weights for its pi0, pi0-FAST and pi0.5 vision-language-action robot models
Overview
openpi is the open-source release of Physical Intelligence, the San Francisco robot AI startup founded in 2024 by researchers including Karol Hausman, Sergey Levine and Chelsea Finn. Under Apache-2.0 it publishes code and checkpoints for three vision-language-action models: pi0, pi0-FAST and pi0.5, as base models plus fine-tunes for ALOHA, DROID (Franka) and LIBERO. It supports JAX and, since September 2025, PyTorch. Inference fits on a GPU with more than 8 GB, LoRA fine-tuning needs about 22.5 GB and full fine-tuning about 70 GB. openpi does not include the company's newer models (pi*0.6 from November 2025, pi0.7 from April 2026); for hosted access the company launched a separate Pi API ("Frontier Models for Every Robot") in beta in October 2026. Investors listed on its site include Bond, Jeff Bezos, Khosla, Lux, OpenAI, Redpoint, Sequoia, CapitalG and Thrive.
The Verdict
Who Should Use openpi?
Best For
- Robot learning researchers who want strong open VLA baselines
- Teams fine-tuning on their own arm data (ALOHA, Franka, UR5-style setups)
- Benchmarking on LIBERO and DROID
Not Ideal For
- Teams wanting Physical Intelligence's latest models (not in openpi)
- Users without a Linux machine and a large GPU for fine-tuning
- Plug-and-play production deployment without robotics expertise
What's Great
- Apache-2.0 code and weights, including ready fine-tunes
- Pretrained on large robot datasets from a leading lab
- Both JAX and PyTorch supported
- Modest inference hardware (8 GB+ GPU, RTX 4090 tested)
Watch Out For
- Tested on Ubuntu 22.04 only
- Large checkpoints; one write-up hit out-of-memory kills while loading
- In-the-wild tests found success sensitive to instruction wording, with mid-task freezes
- Newest models stay closed behind the Pi API
Pricing
View all features & details
Released Checkpoints
- pi0, pi0-FAST and pi0.5 base models
- pi0 and pi0-FAST fine-tuned on DROID
- pi0 ALOHA tasks (towel, tupperware, pen uncapping)
- pi0.5 LIBERO and DROID
Hardware
- Inference: more than 8 GB GPU memory
- LoRA fine-tuning: more than 22.5 GB
- Full fine-tuning: more than 70 GB (A100 80GB or H100)
Company
- Physical Intelligence, San Francisco, founded 2024
- CEO Karol Hausman
- Raised over $1B across several rounds (reported; figures vary by source)
How It Compares
| Feature | openpi | NVIDIA GR00T | Gemini Robotics |
|---|---|---|---|
| Open weights | Yes (Apache-2.0) | Yes (NVIDIA Open Model License) | No |
| Hosted API | Separate Pi API (beta) | No | Robotics-ER 2 in the Gemini API |
| Inference hardware | 8 GB+ GPU | 16 GB+ GPU | Cloud |
| Supported robots (examples) | ALOHA, DROID (Franka), LIBERO | Humanoids; LIBERO, DROID, WidowX | Partner humanoids and arms |
| Price | Free | Free | ER 2 free tier; paid per token |
Is this your tool? Add a badge