YanjieZe/GMR

What it solves

GMR (General Motion Retargeting) converts human motion captured from various sources (like SMPL-X, BVH, FBX, or even monocular video) into joint commands that a humanoid robot can execute. It solves the problem of bridging the gap between human body movement and the different degrees of freedom and physical constraints of diverse robot models, enabling applications like teleoperation and reinforcement learning for robot control.

How it works

GMR takes human motion data (each frame contains per-body-part global translation and rotation) and produces robot motion data (base position, base rotation, and joint angles) via a retargeting process that accounts for the robot's specific structure. It supports multiple input formats (e.g., SMPL-X from AMASS/OMOMO, BVH from LAFAN1/Nokov/Xsens, FBX from OptiTrack, and even PICO streaming) and multiple humanoid robots (17+ supported). The output can be visualized in MuJoCo, recorded as video, or saved as pickle files for downstream use. The code is designed for real-time performance and is tuned to work well with reinforcement-learning tracking policies.

Who it's for

Researchers, roboticists, and developers working on humanoid robots who need to transfer human motion data into robot commands—either for real-time teleoperation (like the TWIST/TWIST2 pipeline) or for generating training data for RL policies. It's also useful for anyone who wants to test a new robot with existing human motion datasets without writing custom conversion code.

Highlights

• Real-time high-quality retargeting, enabling whole-body teleoperation.
• Carefully tuned for good performance of RL tracking policies.
• Supports 17+ humanoid robots and multiple motion data formats (SMPL-X, BVH, FBX, Xsens, etc.).
• Easy to switch between robots by changing one argument.
• Includes tools for batch processing, video recording, and conversion to other formats (e.g., CSV for MimicKit).
• Provides a library and scripts that run on Ubuntu 22.04/20.04 via either uv or conda/pip.