MoRL: Reinforced Reasoning for Unified Motion Understanding and Generation

Hongpeng Wang1*    Zeyu Zhang2*†    Wenhao Li3    Hao Tang2‡

1The University of Sydney    2Peking University    3Nanyang Technological University

*Equal contribution.   †Project lead.   ‡Corresponding author.

arXiv GitHub Model Datasets

Visualization

A person backflips three times in a row.
A person is practicing karate moves across the floor.
A person looks to the left then kicks something with their right foot.
A person walks along a curved path to the right.
A person walks forward slightly shifting to the right.
A person walks forward with a side-to-side sway.
A person walks up stairs.
Walking slowly along the path shaped like an infinity symbol.

Method

Overview diagram of the MoRL framework
Overview of MoRL. Our framework unifies motion understanding and generation under a reinforcement learning paradigm. Motion and text inputs are tokenized into a shared representation space. A hierarchical post-training pipeline first applies SFT on large-scale synthetic CoT datasets to align motion sequences with reasoning traces and concise descriptions, then employs reinforcement learning with verifiable rewards (RLVR) to refine outputs, enhancing semantic alignment, reasoning coherence, physical plausibility, and text–motion consistency. At inference, the Chain-of-Motion (CoM) decoding strategy enables step-by-step reasoning and reflection, improving both motion understanding and perceptually realistic motion generation.


CoT
Motion CoT data engine. Build based on MotionHubV2 dataset, one branch (MoUnd-CoT-140K) uses motion sequences and captions with Gemini to construct reasoning chains for understanding, while the other (MoGen-CoT-140K) builds reasoning chains for generation.

BibTeX


@article{wang2026morl,
  title={MoRL: Reinforced Reasoning for Unified Motion Understanding and Generation},
  author={Wang, Hongpeng and Zhang, Zeyu and Li, Wenhao and Tang, Hao},
  journal={arXiv preprint arXiv:2602.14534},
  year={2026}
}