pi-zero-pytorch

Why we included this project

Physical Intelligence's π₀ architecture pairs a pretrained vision-language model with flow matching so a robot can turn camera frames and language commands into continuous action sequences. This repo is a compact PyTorch implementation of that design, built on PaliGemma 2B as the backbone, and it is laid out in modular pieces you can actually retrain and adapt. Alongside the core model there is a basic training loop, an EFPO wrapper that gathers experience from an environment and learns from it for online-style updates, and a mock environment to exercise the whole pipeline before you touch real hardware. If you want to try the π₀ approach without working through the full official codebase, the README's worked examples get a first training run going quickly.