nvlabs/protomotions

ProtoMotions is a toolkit that teaches digital characters and humanoid robots how to move like real humans. Instead of manually programming every joint movement, you give the system recordings of human motion, and it trains virtual characters or physical robots to replicate those actions. NVIDIA built this open-source framework to help researchers quickly prototype everything from animated characters to real walking robots. The framework works by placing a simulated character or robot in a virtual physics environment and using reinforcement learning to teach it how to match the movements from your motion data. The system practices over and over until it can reliably perform the motions. A standout feature is that you can train a robot in simulation and then deploy that trained model directly onto real hardware, the project demonstrates this by training a Unitree G1 robot to walk and perform various skills without needing any further real-world adjustment. It also supports generating new motions from text prompts via NVIDIA's Kimodo model. Researchers in animation, robotics, and machine learning are the primary audience. A robotics engineer could use it to teach a humanoid robot to navigate difficult terrain, while a game developer might use it to create realistic character animations driven by physics rather than scripted by hand. The framework can learn from large public motion datasets, with the README noting that it can train a character on over 40 hours of motion data in about 12 hours using four high-end GPUs. The project is designed to be highly modular. You can build custom tasks, like steering or navigating terrain, from standalone components rather than rewriting a whole environment from scratch. It supports multiple physics simulators so you can test how a trained policy behaves under different conditions, and adding a new robot involves providing a configuration file and a physical specification. The deployment pipeline exports a single model file, so frameworks controlling the real robot only need to feed it raw sensor data to get it moving.