What is pmt?

PMT, short for Perceptive Motion Tracking, is a research codebase for training a humanoid robot, the Unitree G1, to imitate human movements such as cartwheels and backflips using reinforcement learning, a training method where a program learns by trial and error inside a physics simulator. Training happens inside Isaac Lab, a robotics simulation platform, so this project does not work on its own: you need a working Isaac Lab installation first, and PMT installs into that existing environment. The repository includes scripts for training a policy, meaning the control program that decides how the robot's joints should move, and for playing back a trained policy to watch it perform. It also ships with about 25 pre-built training task setups for the G1 robot, sample motion capture clips for the robot to imitate, and several ready-made checkpoints, which are already trained models you can run immediately without training from scratch. Everything in PMT is controlled through configuration files rather than code changes. A task picks settings for things like which robot, which terrain, which motion, and which reward function to use, and the system automatically works out any settings that depend on those choices, refusing to run if you pick an invalid combination. The project also includes a separate tool called TCRS that can generate optimized versions of raw motion clips adjusted for a specific terrain, and it can additionally run on an alternative physics engine called mjlab for at least the simpler flat ground tasks. Because the pretrained models and demo motion files are large, they are stored separately from the main code, either on Hugging Face or Google Drive, with instructions for downloading and copying them into the right folders after cloning the repository. Getting this project running requires a working Isaac Lab and Isaac Sim setup, a compatible GPU, and Python 3.10 or newer, so it is intended for robotics researchers rather than casual users.