Roboparty/UFO

What it solves

UFO provides a framework for training humanoid robots to perform complex movements without requiring manually labeled reward functions. It uses unsupervised reinforcement learning to enable robots like the Unitree G1 and RoboParty RP1 to learn motor control, tracking, and goal-oriented tasks from motion data.

How it works

The framework leverages the mjlab simulation environment to train policies using reinforcement learning. It supports multiple training agents (FB and TeCH) and handles the complexities of humanoid control, such as contact-force dynamics and joint-space scaling. The system includes tools for data preparation, multi-GPU training, and exporting policies to ONNX format for deployment on physical hardware.

Who it’s for

This project is designed for robotics researchers and engineers working on embodied intelligence and humanoid locomotion. It is intended for those who need a scalable, high-performance training pipeline that can adapt to different robot morphologies and motion datasets.

Highlights

• Supports multiple humanoid platforms including Unitree G1 and RoboParty RP1.
• Implements efficient multi-GPU training with optimized features like fused Adam and bucketed gradient synchronization.
• Includes built-in support for tracking, reward inference, and goal-directed movement.
• Provides ready-to-use ONNX exports for deployment on physical robots.
• Offers a modular architecture for adding new robot configurations and retargeted motion data.