roboparty/ufo

UFO is a research framework for training humanoid robots to move using a type of machine learning called unsupervised reinforcement learning. Instead of hand coding specific walking or running motions, the system learns general movement skills from motion data and can later be pointed at a goal or a motion clip to reproduce. The project supports two named training methods referred to as FB and TeCH, along with tools to import existing motion data, run tracking or goal based inference, and export a trained model in the ONNX format so it can run outside the original training code. The most complete and well tested setup is for the Unitree G1 humanoid robot. Support for other robots exists but is described as experimental, and adding a new robot requires supplying its physical description as a MuJoCo XML file, optionally a URDF file, and motion data that has already been adapted to that specific robot's body. The framework does not automatically adapt motion recorded for one robot or for a human to a different robot, and a trained model cannot simply be reused on a robot with a different body or set of joints. Installation involves cloning the repository, installing the uv Python package manager, and syncing the project's dependencies. A guided quick start walks through downloading processed motion data for the G1 robot, running a short smoke test to confirm everything works, then running full scale training across multiple GPUs. A separate branch of the project handles deploying trained models onto a real G1 robot and teleoperating it. This project is intended for robotics researchers and engineers working on humanoid control, not hobbyists or general developers. It assumes access to GPUs for training and, for real world use, an actual Unitree G1 robot.