OpenPI

Physical Intelligence VLA foundation models for general robot control

License

Apache-2.0

Overview

Physical Intelligence open-source vision-language-action (VLA) foundation models for general robot control. Includes pi0, pi0-FAST, and pi0.5 models pre-trained on 10k+ hours of robot data. The first general-purpose robot foundation model, capable of controlling diverse robot platforms across multiple tasks including folding laundry, assembling boxes, bussing tables, and making coffee.

OpenPI is an open-source repository from Physical Intelligence containing vision-language-action (VLA) foundation models and training infrastructure for robotics. It provides base model checkpoints pre-trained on 10,000+ hours of robot data for fine-tuning and inference. Available models include pi0 (flow-based VLA), pi0-FAST (autoregressive with FAST action tokenizer), and pi0.5 (improved open-world generalization with knowledge insulation). The project achieves state-of-the-art results on benchmarks like LIBERO and supports diverse tasks including folding laundry, assembling boxes, bussing tables, and making coffee across multiple robot platforms.

The newest model in this repository is pi0.5, from September 2025. In October 2026 Physical Intelligence announced a Pi API beta that offers its latest models another way: you upload data, fine-tune on compute it manages, and run the result through hosted inference. Entry to the beta is by application. The post does not say those latest models will be published here, so the checkpoints in this repository are what you can run on your own hardware.

Why it matters

OpenPI represents the first general-purpose robot foundation model from Physical Intelligence. Its pi0 family of models demonstrates that a single VLA model can control diverse robots across many tasks, signaling a shift toward foundation models in robotics similar to what GPT did for language.

Problem it solves

Building robot policies from scratch for each task is expensive and slow. OpenPI provides pre-trained foundation models that can be fine-tuned for new tasks with minimal data, dramatically reducing the barrier to deploying capable robot behaviors.

Key takeaways

• First open-source general-purpose robot foundation model (pi0, pi0-FAST, pi0.5)
• Pre-trained on 10,000+ hours of diverse robot data across multiple embodiments
• Capable of complex multi-step tasks: folding laundry, assembling boxes, making coffee
• Vision-language-action architecture processes both visual input and language instructions
• Open weights and code enable community fine-tuning on custom robot setups

Who it is for

• Robotics researchers exploring foundation models
• Companies deploying manipulation robots
• ML engineers fine-tuning robot policies
• Research labs with limited robot demonstration data

Limitations

• Requires significant GPU resources for inference (large VLA models)
• Fine-tuning still needs some task-specific demonstration data
• Performance depends on similarity to pre-training robot embodiments
• Real-world deployment requires careful safety evaluation
• This repository is not where Physical Intelligence's latest models are offered: since October 2026 those come through its hosted Pi API beta, by application. The newest model here is pi0.5.