NeCR

Human-to-humanoid motion retargeting converts human motion into reference trajectories for humanoid control. Learning-based methods transfer motion semantics efficiently but enforce nothing at inference, while optimization-based methods enforce constraints but can settle on a pose that breaks the supporting contact. What is missing is the signal that tells the optimizer where correction is needed, over which intervals, and with what confidence. We present NeCR, a hybrid neural-optimization framework for contact-consistent motion retargeting. NeCR first predicts a semantically faithful nominal robot trajectory. A surface-conditioned temporal model then predicts continuous whole-body contact modes. Conditioned on them, correction reduces to a sparse, block-banded proximal projection that adjusts only support-relevant degrees of freedom. On public datasets retargeted to a Unitree G1, physical correction becomes nearly free: penetration is nearly eliminated and support-phase sliding falls from 20.6% to 4.4% at a cost of 0.16 cm MPKPE, and NeCR runs at 5.9 ms per frame end to end, over five times faster than real time at 30 Hz. The resulting references are also more trackable, improving policy success across two independently implemented trackers and under cross-reference evaluation. A proprioceptive policy trained on them transfers zero-shot to a real Unitree G1 on whole-body-contact motions such as crawling and fall recovery. Code, retargeted datasets, and trained policies will be released.