HM2O: Humanoid Motion Retargeting via Monolithic Optimization

Weixiang Gao, Chengzhou Tang

Abstract:

Abstract— Transferring human motion to humanoid robots
remains difficult due to compounded kinematics and dynamics
gaps, and common multi-stage pipelines can accumulate errors
across stages. In this work, We propose HM2O (Humanoid
Motion Retargeting via Monolithic Optimization), a joint opti-
mization framework that minimizes these gaps simultaneously
by (i) optimizing robot joint-link scales together with base
pose and joint angles inside the robot forward kinematics,
(ii) decoupling position and rotation targets so they may use
different human–humanoid joint mappings, and (iii) adding
efficient dynamics regularizations including a ground-offset
term, a stance (anti-slippage) term, and a biomechanics-inspired
Froude-number prior to improve physical feasibility. HM2O is
efficient and runs in real time on a single CPU core. We evaluate
our method on large-scale AMASS dataset, with off-the-shelf
motion tracking policies including MaskedMimic, TWIST and
GMT without training. Both the quantitative and qualitative
results demonstrate that our HM2O generates motion sequences
that are more physically feasible, and boost the performance
of a motion tracking policy