PriMoRetarget is a motion-retargeting framework for transferring human motion to humanoid robots. Rather than balancing competing objectives with manually tuned weights, PriMoRetarget uses an explicit hierarchy between objectives to preserve critical characteristics such as foot placement and interaction contacts while optimizing remaining objectives of lower priority.
Unlike weighted-sum approaches that assign weights to each objective to denote its relative importance,
PriMoRetarget optimizes each objective L^k over the motions S^{k-1} that preserve all higher-priority objectives:
and the remaining set of motions becomes
This is enforced by projecting each gradient update onto the Nullspace of the Jacobians \mathbf{J} of the higher-priority objectives.
Illustration of weighted-sum vs. prioritized optimization for two quadratic objectives. The weighted sum settles on a compromise, while the prioritized update satisfies the higher-priority objective (red line) and improves the other objective only along its free directions.
Illustration of conflicting objectives in motion retargeting. The source pose remains fixed while the box is synthetically moved throughout the motion. PriMoRetarget prioritizes the hand contacts, adapting the hands to follow the box while allowing some pose tracking error. In contrast, GMR compromises between the competing objectives, resulting in neither accurate pose tracking nor consistent hand contacts.
We evaluate PriMoRetarget against existing retargeting methods across different interaction settings. The resulting retargeted motions are used as references for a tracking policy, and selected motions are deployed on a Unitree G1 humanoid robot.
@article{ioannidis2026primoretarget,
title = {PriMoRetarget: Humanoid Motion Transfer Through Prioritized Motion Retargeting},
author = {Ioannidis, Nicholas and Yang, Dun and Smith, Max and Vu, Tue and
Zhang, Ziyu and Omidshafiei, Shayegan and Peng, Xue Bin},
journal = {arXiv preprint arXiv:2610.12642},
year = {2026}
}