MoRL: Reinforced Reasoning for Unified Motion Understanding and Generation

Hongpeng Wang; Zeyu Zhang; Wenhao Li; Hao Tang

cs.CVFeb 16, 2026

MoRL: Reinforced Reasoning for Unified Motion Understanding and Generation

Authors: Hongpeng Wang, Zeyu Zhang, Wenhao Li, Hao Tang

Organizations: The University of Sydney · Peking University · Nanyang Technological University

Abstract

Human motion understanding and generation are crucial for vision and robotics but remain limited in reasoning capability and test-time planning. We propose MoRL, a unified multimodal motion model trained with supervised fine-tuning and reinforcement learning with verifiable rewards. Our task-specific reward design combines semantic alignment and reasoning coherence for understanding with physical plausibility and text-motion consistency for generation, improving both logical reasoning and perceptual realism. To further enhance inference, we introduce Chain-of-Motion (CoM), a test-time reasoning method that enables step-by-step planning and reflection. We also construct two large-scale CoT datasets, MoUnd-CoT-140K and MoGen-CoT-140K, to align motion sequences with reasoning traces and action descriptions. Experiments on HumanML3D and KIT-ML show that MoRL achieves significant gains over state-of-the-art baselines. Code: https://github.com/AIGeeksGroup/MoRL. Website: https://aigeeksgroup.github.io/MoRL.

Figures & tables

Figure 1: Visualization comparisons with MotionLLM. In the backflip example, MotionLLM fails to maintain a coherent takeoff-rotation-landing trajectory, resulting in unstable body orientation, while MoRL completes a physically plausible flip. In the Wack-style dance, MotionLLM shows inconsistent rotation direction and fragmented poses, whereas MoRL preserves continuous left-to-right rotation and stylistic coherence.

Figure 2: Motion CoT data engine. Build based on MotionHubV2 dataset ( Ling et al., 2024 ) , one branch (MoUnd-CoT) uses motion sequences and captions with Gemini to construct reasoning chains for understanding, while the other (MoGen-CoT) builds reasoning chains for generation.

Figure 3: 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.

Method	Motion Generation							Motion Understanding
	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	FID $\downarrow$	MM Dist $\downarrow$	Div $\rightarrow$	MM $\uparrow$	BLEU@1 $\uparrow$	BLEU@4 $\uparrow$	ROUGE-L $\uparrow$	CIDEr $\uparrow$	BERTScore $\uparrow$
HumanML3D ( Guo et al., 2022a )
GT / Real Motions	0.511	0.703	0.797	0.002	2.974	9.503	-	-	-	-	-	-
SeqGAN ( Goutsu and Inamura, 2021 )	-	-	-	-	-	-	-	47.80	13.50	39.20	50.20	23.40
RAEs ( Yamada et al., 2018 )	-	-	-	-	-	-	-	33.30	10.20	37.50	22.10	10.70
Seq2Seq(Att) ( Plappert et al., 2018 )	-	-	-	-	-	-	-	51.80	17.90	46.40	58.40	29.10

Table 1: Comparison of motion generation and motion understanding on HumanML3D and KIT-ML. Highlights indicate the unified model, bold represent the best results within the unified model. Results marked with ∗ are reproduced by MotionGPT ( Jiang et al., 2023 ) and Lyu et al. ( Lyu et al., 2025 ) , and are computed with unprocessed ground truth texts for linguistic metrics.

Variant	Understanding			Generation
	BERT $\uparrow$	CIDEr $\uparrow$	R-L $\uparrow$	R@1 $\uparrow$	FID $\downarrow$
SFT only	42.65	33.88	48.78	0.420	0.212
w/o $R_{\text{sem}}$	44.10	34.05	50.01	0.488	0.209
w/o $R_{\text{coh}}$	44.32	35.12	49.10	0.512	0.206
w/o $R_{\text{phys}}$	46.18	35.50	51.19	0.518	0.285
w/o $R_{\text{align}}$	45.00	34.62	50.48	0.492	0.225

Table 2: Ablation study of MoRL on HumanML3D.

Generation Reward	R@1 $\uparrow$	R@2 $\uparrow$	R@3 $\uparrow$	FID $\downarrow$	MM Dist $\downarrow$	MM $\uparrow$
Motion-R1 Reward ( Ouyang et al., 2025 )	0.472	0.651	0.742	0.185	3.021	2.31
MotionRL Reward ( Liu et al., 2024 )	0.491	0.676	0.768	0.172	2.914	2.28
Process-aware Reward ( Wang et al., 2024a )	0.486	0.682	0.781	0.181	2.873	2.45
MoRL Reward (Ours)	0.489	0.703	0.814	0.179	2.756	2.42

Table 3: Comparison of different reward designs on the CMS of HumanML3D. All methods share the same backbone and training setup, differing only in the reward used during motion generation.

Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

NLI Model	Params	Var. $\downarrow$	BLEU $\uparrow$	FID $\downarrow$
DeBERTa-v3-XL	1.2B	Low–Med.	+3.3	-2.0
DeBERTa-v3-L	435M	Lowest	+3.4	-2.1
RoBERTa-L	355M	Low	+2.8	-1.5
DeBERTa-B	110M	Med.	+1.9	-0.8
BERT-B	110M	High	+0.7	-0.2

Appendix

Table 4: Comparison of NLI models used as

f_{\text{NLI}}

. Lower reward variance indicates more stable RL updates. BLEU↑ and FID↓ reflect performance gains relative to the SFT baseline.

Method	Lat. (ms) $\downarrow$	Thru. (sps) $\uparrow$	Cost
Single-pass	8.7	115	baseline
CoM (K=8, T=2)	18.4	55	2.1 $\times$

Appendix

Table 5: End-to-end inference efficiency of single-pass decoding vs. CoM. Latency (Lat.) is measured per sample. Throughput (Thru.) is computed at batch size 32.

Method	BERT $\uparrow$	CIDEr $\uparrow$	R@1 $\uparrow$	FID $\downarrow$	Var $\downarrow$	Time $\downarrow$
SFT only	42.65	33.88	0.420	0.232	–	–
PPO	45.12	34.95	0.492	0.228	1.00	1.35 $\times$
DPO	44.31	34.10	0.468	0.241	0.87	1.20 $\times$
GRPO	46.80	35.80	0.527	0.203	0.63	1.00 $\times$

Appendix

Table 6: Comparison of different optimization strategies under identical settings. GRPO provides the best overall performance and training stability.

Panel of Table 7 — Table 7: Qualitative comparison (Part I).

Panel of Table 8 — Table 8: Qualitative comparison (Part II).

Human Motion Understanding

See all 16 papers

IRG-MotionLLM: Interleaving Motion Generation, Assessment and Refinement for Text-to-Motion Generation
Dec 11, 2025Yuan-Ming Li, Qize Yang, Nan Lei +5Text-to-Motion Generation
Sun Yat-sen University · Tongyi Lab, Alibaba Group · Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China +1
Re $^2$ MoGen: Open-Vocabulary Motion Generation via LLM Reasoning and Physics-Aware Refinement
Apr 20, 2026Jiakun Zheng, Ting Xiao, Shiqin Cao +3Text-to-Motion Generation Human Motion Generation
1East China University of Science and Technology · 2Hong Kong University of Science and Technology · Institute of Artificial Intelligence (TeleAI), China Telecom
Open-UniMo: Towards Unified Motion-Language Understanding and Generation in the Open World
Sep 13, 2026Guocun Wang, Kenkun Liu, Guorui Song +7Text-to-Motion Generation
Tsinghua University, Beijing 100190, China · SSE, The Chinese University of Hong Kong, Shenzhen 518172, China · MMLab, Nanyang Technological University, Singapore 639798

Dec 11, 2025J/K move · Enter open · S save

IRG-MotionLLM: Interleaving Motion Generation, Assessment and Refinement for Text-to-Motion Generation

Yuan-Ming Li, Qize Yang, Nan Lei, Shenghao Fu+4

Sun Yat-sen University · Tongyi Lab, Alibaba Group · Key Laboratory of Machine Intelligence and Advanced Computing, Ministry of Education, China+1

Text-to-Motion Generation

Open paper AskPDF arXiv

Recent advances in motion-aware large language models have shown remarkable promise for jointly learning motion understanding and generation knowledge. However, these models typically treat understanding and generation separately, limiting the mutual benefits that could arise from interactive feedback between tasks. In this work, we reveal that motion assessment and refinement tasks can act as crucial bridges to enable knowledge flow from motion understanding to generation. Specifically, we propose Interleaved Reasoning for Motion Generation (IRMoGen), a novel paradigm that tightly couples motion generation with assessment and refinement through iterative text-motion dialogue. To realize this, we introduce IRG-MotionLLM, the first model that seamlessly interleaves motion generation, assessment, and refinement to improve the alignment between generated motion and goal text. IRG-MotionLLM is developed progressively with a novel three-stage training scheme, initializing and subsequently enhancing native IRMoGen capabilities. To facilitate this development, we construct an automated data engine to synthesize interleaved reasoning annotations from existing text-motion datasets. Extensive experiments demonstrate the properties brought by IRMoGen training, and the advanced cross-benchmark and cross-evaluator performance of IRG-MotionLLM. Code and models are available at https://github.com/HumanMLLM/IRG-MotionLLM.

Abstract

Human Motion Understanding

Explore similar work

IRG-MotionLLM: Interleaving Motion Generation, Assessment and Refinement for Text-to-Motion Generation

Re2^22MoGen: Open-Vocabulary Motion Generation via LLM Reasoning and Physics-Aware Refinement

Open-UniMo: Towards Unified Motion-Language Understanding and Generation in the Open World

Re $^2$ MoGen: Open-Vocabulary Motion Generation via LLM Reasoning and Physics-Aware Refinement