OmniRetarget

A data-generation method that retargets a human's interaction with objects and terrain onto a humanoid robot together, not just body pose.

OmniRetarget is a motion-retargeting and data-generation method proposed in 2025 by Amazon FAR (Frontier AI and Robotics), together with MIT, UC Berkeley, Stanford, and CMU. Retargeting converts human motion into robot joint motion; when only body keypoints are aligned, tasks like carrying a box or climbing onto a platform often end up with the hands missing the box or the feet sinking into the floor. OmniRetarget uses an “interaction mesh” to jointly model the spatial and contact relationships among the human, the object, and the terrain, preserving those relationships while respecting the robot's joint limits, and it can also swap in different robots, terrains, or objects for data augmentation. After generating over 8 hours of trajectories from OMOMO, LAFAN1, and its own motion-capture data, only 5 reward terms were needed to train motions like box-carrying and climbing, up to about 30 seconds long, on a Unitree G1. The project page states it received the Best Conference Paper award at ICRA 2026.

ExampleMotion-capture data of a person carrying a box is retargeted onto a Unitree G1 while keeping both hands pressed against the box and both feet planted on the ground; the resulting trajectories are then used to train a reinforcement-learning policy that can carry boxes on the real robot.

Also called

OmniRetarget: Interaction-Preserving Data Generation for Humanoid Whole-Body Loco-Manipulation and Scene Interaction

Related

Motion Retargeting · General Motion Retargeting · OMOMO · LAFAN1 · Loco-manipulation · Unitree G1

Sources

OmniRetarget (arXiv 2509.26633)
OmniRetarget 项目主页 (Chinese)

As of

2026-06