openpi overview and key facts (October 2026)

openpi is the open source repository of Physical Intelligence with the robot foundation models π0, π0-FAST and π0.5. Robotics engineers use the pretrained checkpoints to run or fine-tune general manipulation policies on their own robots.

The repository contains three model families: π0, a flow-based vision-language-action model; π0-FAST, an autoregressive variant using the FAST action tokenizer; and π0.5, an upgrade with better open-world generalization. Base checkpoints are pretrained on more than 10,000 hours of robot data, with examples for direct use and fine-tuning, including on the DROID dataset.

PyTorch support was added in September 2025. The README lists at least 8 GB of GPU memory for inference, 22.5 GB for LoRA fine-tuning and 70 GB for full fine-tuning, tested on Ubuntu 22.04. The repository had 14,016 GitHub stars and its latest push on August 24, 2026.

• Starts from a pretrained robot policyThe base checkpoints already encode manipulation skills learned from thousands of hours of robot data. Teams can test grasping and placing on their robot before collecting large datasets of their own.
• Fine-tunes on company dataScripts fine-tune the models on your own recorded demonstrations, with LoRA as a lighter option on a single high-end GPU. Automation teams can adapt a general policy to one production task at moderate hardware cost.
• Compares model approachesHaving flow-based and autoregressive variants in one codebase lets engineers compare speed and quality on the same task. This helps pick the approach that fits the robot's control rate.

• Three model variants from the same lab in one repository, all under Apache 2.0.
• Pretrained on more than 10,000 hours of robot data according to the README.
• The README states the models were developed for the lab's own robots, so results on other platforms may vary.

openpi is an open source project under the Apache 2.0 license and runs on your own hardware in any country.

openpi is research code, not a supported product: the README calls it an experiment and supports only Ubuntu 22.04 with NVIDIA GPUs. Your team runs, secures and maintains it. Plan data collection for your robot, GPU capacity for fine-tuning and a safety concept before any policy controls hardware near people.

No reviews yet. Be the first to share how your team uses it.

Every review shows its verification level. We read each one before it goes live. How we check reviews