General Motion Retargeting

Stanford's open-source tool that retargets human motion onto many different humanoid robots in real time.

GMR is a motion-retargeting method and codebase open-sourced in 2025 by Jiajun Wu and C. Karen Liu's group at Stanford (with Araújo, Yanjie Ze, and others). Motion retargeting converts human motion into robot joint angles, and the hard part is that humans and robots differ in body proportions and joint structure. GMR first specifies which body parts on the human correspond to which parts on the robot, aligns the two bodies' rest poses, scales each body part separately, and then solves joint angles with a two-stage inverse-kinematics optimization, which reduces foot skating, self-penetration, and sudden jumps in joint angle. Supported inputs include SMPL-X-format data such as AMASS, BVH files such as LAFAN1, OptiTrack FBX exports, live Xsens streams, and motion extracted from monocular video via GVHMR. The README lists support for 18 humanoid robots, running at 60–70 frames per second on an ordinary CPU. The paper shows it produces better results for training tracking policies than open-source alternatives like PHC and ProtoMotions, and it also serves as the retargeting module in the TWIST teleoperation system.

ExampleA clip of a human running and jumping from the LAFAN1 motion-capture dataset is retargeted with GMR onto the Unitree G1 humanoid, and the result is used to train a motion-tracking policy in simulation to imitate it.

Also called

GMR, Retargeting Matters: General Motion Retargeting for Humanoid Motion Tracking

Related

Motion Retargeting · Foot Skating / Floating / Penetration · Twist · LAFAN1 · AMASS (Archive of Motion Capture as Surface Shapes) · Motion Tracking

Sources

Retargeting Matters: General Motion Retargeting for Humanoid Motion Tracking (arXiv)
GMR (GitHub)

As of

2026-09