.png){kind=icon}
.png){kind=icon}
Return
A ZMP-Aware Task Formulation for Reference-Driven Humanoid Tracking in MuJoCo MPC
.png){kind=icon}
PRE
AI
DOI：10.3390/sym18050768.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Reference-driven humanoid motion tracking aims to reproduce a source motion on a target humanoid while preserving physical executability under actuation limits and changing contact conditions. The problem becomes particularly challenging for dynamic motions involving rapid support transitions, landing impacts, mixed hand-foot contacts, and moderate topology-preserving morphology variation. Existing pipelines often rely heavily on morphology-specific world-frame targets or treat balance and contact quality only indirectly during execution, which limits their reliability under dynamic contact variation. This paper presents a task and cost formulation for reference-driven humanoid tracking within the residual-based MuJoCo model predictive control (MPC) framework. The source motion is decomposed into a pelvis-centered canonical local reference, pelvis height and tilt references, and a pelvis-derived horizontal center-of-mass (CoM) velocity intent, and is tracked online with a zero moment point (ZMP)-aware contact-conditioned residual design including slip, penetration, posture, and control regularization. The formulation is compatible with standard MuJoCo MPC planners, and the evaluation is conducted under a shared iterative linear quadratic Gaussian (iLQG) setting on nominal and morphology-varied humanoids against tracking-only and two-stage inverse-kinematics (IK)-based baselines. The proposed formulation improves success rate, support quality, slip reduction, and progression accuracy, with the clearest gains on contact-sensitive motions; for example, success rate increases from 56.7% to 76.7% on Jump-Turn and from 46.7% to 70.0% on Cartwheel relative to the tracking-only MPC baseline. These results support the use of execution-oriented reference representation and contact-conditioned residual design for physically reliable reference-driven humanoid tracking.
Keywords:
reference-driven humanoid tracking
humanoid motion retargeting
model predictive control
zero moment point
contact-aware control
MuJoCo simulation