Physical-Intelligence/openpi ? reverse-engineered prompt

Build me a Python toolkit for robotics researchers that makes it easy to use open source vision language action models for robot control. I want people to be able to load pretrained π₀, π₀ FAST, and π₀.₅ checkpoints, run inference on supported robot setups, and fine tune the base models on their own robot datasets.

Please include clear examples for common platforms like DROID, ALOHA, and LIBERO, with simple scripts for downloading or pointing to checkpoints, preparing data, running evaluation, and starting fine tuning. Add setup instructions for Ubuntu with NVIDIA GPUs, including both a normal Python environment and a Docker option. Make the docs honest that these models may not work on every robot without adaptation.

Keep the project organized like a real research repo, with reusable model code, training and inference scripts, example configs, and beginner friendly README docs. Look up current docs online if you need to.