Adaptive Reference Refinement for Physics‐based Motion Retargeting
Abstract Physics‐based character control for high‐dimensional humanoid motion requires reference motion‐capture data to guide stable training of the reinforcement learning (RL) policy. However, the motion dynamics of the reference trajectory and the control target may differ in terms of mass, torque limits, and morphology. Conventional retargeting compensates for offsets based on kinematic structure, without considering real‐world dynamics such as gravity, contact forces, and torque control, and the resulting motion fails to robustly guide policy training for the control target. We propose a motion refinement network that learns to automatically adjust retargeted results to be physically plausible, reflecting the non‐linear relationship arising from the dynamic interplay between different skeleton structures and the environment. As the policy directly learns to track refined references in physical simulation, the resulting motions are free of physical artifacts, such as foot sliding, ground penetration, and implausible joint configurations, which are frequently observed in retargeting results applying inverse kinematics. While concurrent training of motion refinement and the RL policy often diverges, we introduce a warming‐up scheme that initially trains the policy using only a fixed refinement network, then gradually introduces refinement offsets. The offsets are then selectively applied to the retargeted motion only when a substantial deviation is required. Experiments show that our method consistently improves physical plausibility over retargeting baselines using inverse kinematics, and further improves tracking performance particularly when combined with baselines that produce reasonably reliable initial references. Our method also generalizes robustly to retargeting across extremely different morphologies and physical properties, including dramatic changes in limb lengths and carrying of excessively heavy payloads.
Authors
- Seungbin Jeong (ORCID: https://orcid.org/0009-0008-2597-5652)
- Young Min Kim (ORCID: https://orcid.org/0000-0002-6735-8539)
Institutions
- Seoul National University (KR)
Publication Details
- Journal
- Computer Graphics Forum
- Published
- 2026-11-13
- DOI
- https://doi.org/10.1111/cgf.70610
- Primary Topic
- Human Motion and Animation
- Type
- article
- Field-Weighted Citation Impact
- 0.00