Iterative Closed-Loop Motion Synthesis for Scaling the Capabilities of Humanoid Control

CVPR 2026
Weisheng Xu1*, Qiwei Wu1*, Jiaxi Zhang1, Jing Tan1, Yangfan Li1, Yuetong Fang1, Jiaqi Xiong2, Kai Wu1, Rong Ou1, Renjing Xu1†
1Hong Kong University of Science and Technology (Guangzhou), 2University of Oxford
*Equal Contribution, †Corresponding Author
Paper Supp arXiv Code
CLAIMS Pipeline

CLAIMS is a closed-loop automated framework that co-evolves motion data synthesis and controllers, scaling the capabilities of physics-based humanoid control.

Abstract

Physics-based humanoid control relies on training with motion datasets that have diverse data distributions. However, the fixed difficulty distribution of datasets limits the performance ceiling of the trained control policies. Additionally, the method of acquiring high-quality data through professional motion capture systems is constrained by costs, making it difficult to achieve large-scale scalability.

To address these issues, we propose a closed-loop automated motion data generation and iterative framework, CLAIMS. It can generate high-quality motion data with rich action semantics, including martial arts, dance, combat, sports, gymnastics, and more.

Furthermore, our framework enables difficulty iteration of policies and data through physical metrics and objective evaluations, allowing the trained tracker to break through its original difficulty limits. On the PHC single-primitive tracker, using only approximately 1/10 of the AMASS dataset size, the average failure rate on the test set (2201 clips) is reduced by 45% compared to the baseline.

1. Evolution of Tracking Capability on Dynamic Kicks

Comparing tracking performance of different training loops (L0, L3, L6) across escalating difficulty levels.

Difficulty Level: L0 Kick

Reference Motion

L0 Tracker

L3 Tracker

L6 Tracker (Ours)

Difficulty Level: L3 Kick

Reference Motion

L0 Tracker

L3 Tracker

L6 Tracker (Ours)

Difficulty Level: L6 Kick

Reference Motion

L0 Tracker

L3 Tracker

L6 Tracker (Ours)

2. Generalization to Extreme OOD Maneuvers

Comparing our final L6 Tracker with the PHC baseline on out-of-distribution dynamic motions.

OOD Motion 1

Reference

PHC Tracker

Ours (L6)

OOD Motion 2

Reference

PHC Tracker

Ours (L6)

OOD Motion 3

Reference

PHC Tracker

Ours (L6)

OOD Motion 4 (Failure Case)

Reference

PHC Tracker (Fail)

Ours L6 (Fail)

3. Synthesized Motions Across 5 Professional Domains

Showcasing the diversity of our generated high-quality action semantics.

Motion Sample Set 1

Combat

Dance

Gymnastics

Martial Arts

Sports

Motion Sample Set 2

Combat

Dance

Gymnastics

Martial Arts

Sports

BibTeX

@inproceedings{xu2026iterative,
  title={Iterative Closed-Loop Motion Synthesis for Scaling the Capabilities of Humanoid Control},
  author={Xu, Weisheng and Wu, Qiwei and Zhang, Jiaxi and Tan, Jing and Li, Yangfan and Fang, Yuetong and Xiong, Jiaqi and Wu, Kai and Ou, Rong and Xu, Renjing},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={16398--16407},
  year={2026}
}