Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Motion-R1: Enhancing Motion Generation with Decomposed Chain-of-Thought and RL Binding

Runqi Ouyang, Haoyun Li, Zhenyuan Zhang, Xiaofeng Wang, Zeyu Zhang, Zheng Zhu, Guan Huang, Sirui Han, Xingang Wang

TL;DR

<3-5 sentence high-level summary> Motion-R1 tackles the challenge of producing temporally coherent and semantically faithful motion from text by integrating a Decomposed CoT Data Engine with RL Binding. The approach leverages a motion tokenizer (VQ-VAE) and a reasoning-enabled LLM to generate structured motion plans and tokens, then refines them through GRPO-based reinforcement learning that incorporates multi-modal alignment rewards. Key contributions include (i) a Decomposed CoT Data Engine that creates step-by-step reasoning traces to guide motion planning, (ii) an RL Binding framework that embeds motion and semantic alignment into the reward, and (iii) comprehensive experiments showing state-of-the-art performance on HumanML3D, KIT-ML, and BABEL with improved diversity and realism. This work advances controllable, scalable text-to-motion generation with interpretable reasoning and reduced reliance on expensive human annotations.

Abstract

Text-to-Motion generation has become a fundamental task in human-machine interaction, enabling the synthesis of realistic human motions from natural language descriptions. Although recent advances in large language models and reinforcement learning have contributed to high-quality motion generation, two major challenges remain. Existing approaches often fail to capture the temporal and causal complexities inherent in natural language, leading to oversimplified or incoherent motions. Additionally, RL-based methods are frequently overly complex, hindering their scalability and adaptability across various motion generation tasks. To address these challenges, we propose Motion-R1, a novel framework that combines decomposed Chain-of-Thought reasoning with reinforcement learning to enhance both the quality and interpretability of generated motions. Specifically, we introduce the Decomposed CoT Data Engine, which leverages an automated pipeline to synthesize high-quality reasoning data, allowing the model to better capture the temporal dependencies and causal relationships of human motion. We also propose RL Binding, a reinforcement learning strategy that incorporates multi-modal text-motion alignment into the RL reward function, guiding the model to produce motions that are both semantically accurate and motionally realistic. Extensive experiments across benchmark datasets demonstrate that Motion-R1 achieves state-of-the-art performance, with a 3.5% improvement in MM-Dist on HumanML3D and improvements in R-Precision and FID on KIT-ML and BABEL, surpassing existing methods across key metrics and highlighting its superior capability in handling complex motion generation tasks. Project page: https://motion-r1.github.io/.

Motion-R1: Enhancing Motion Generation with Decomposed Chain-of-Thought and RL Binding

TL;DR

<3-5 sentence high-level summary> Motion-R1 tackles the challenge of producing temporally coherent and semantically faithful motion from text by integrating a Decomposed CoT Data Engine with RL Binding. The approach leverages a motion tokenizer (VQ-VAE) and a reasoning-enabled LLM to generate structured motion plans and tokens, then refines them through GRPO-based reinforcement learning that incorporates multi-modal alignment rewards. Key contributions include (i) a Decomposed CoT Data Engine that creates step-by-step reasoning traces to guide motion planning, (ii) an RL Binding framework that embeds motion and semantic alignment into the reward, and (iii) comprehensive experiments showing state-of-the-art performance on HumanML3D, KIT-ML, and BABEL with improved diversity and realism. This work advances controllable, scalable text-to-motion generation with interpretable reasoning and reduced reliance on expensive human annotations.

Abstract

Text-to-Motion generation has become a fundamental task in human-machine interaction, enabling the synthesis of realistic human motions from natural language descriptions. Although recent advances in large language models and reinforcement learning have contributed to high-quality motion generation, two major challenges remain. Existing approaches often fail to capture the temporal and causal complexities inherent in natural language, leading to oversimplified or incoherent motions. Additionally, RL-based methods are frequently overly complex, hindering their scalability and adaptability across various motion generation tasks. To address these challenges, we propose Motion-R1, a novel framework that combines decomposed Chain-of-Thought reasoning with reinforcement learning to enhance both the quality and interpretability of generated motions. Specifically, we introduce the Decomposed CoT Data Engine, which leverages an automated pipeline to synthesize high-quality reasoning data, allowing the model to better capture the temporal dependencies and causal relationships of human motion. We also propose RL Binding, a reinforcement learning strategy that incorporates multi-modal text-motion alignment into the RL reward function, guiding the model to produce motions that are both semantically accurate and motionally realistic. Extensive experiments across benchmark datasets demonstrate that Motion-R1 achieves state-of-the-art performance, with a 3.5% improvement in MM-Dist on HumanML3D and improvements in R-Precision and FID on KIT-ML and BABEL, surpassing existing methods across key metrics and highlighting its superior capability in handling complex motion generation tasks. Project page: https://motion-r1.github.io/.

Paper Structure

This paper contains 32 sections, 7 equations, 4 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Multimodal Large Language Models
  4. Text-guided 3D Human Motion Generation
  5. Method
  6. Motion-R1 Framework
  7. Motion Tokenizer
  8. Decomposed CoT Data Engine
  9. Training Strategy
  10. Cold-Start Training with Decomposed CoT
  11. GRPO-based Training with RL Binding
  12. Experiments
  13. Experimental Settings
  14. Quantitative Evaluation
  15. Qualitative Analysis and Visualization
  16. ...and 17 more sections

Figures (4)

  • Figure 1: Comparison of traditional approaches and our Motion-R1 framework. (a) Traditional end-to-end models exhibit poor generalization on out-of-distribution motions. (b) Our Decomposed CoT Data Engine enables strong generalization by structuring high-level instructions into intermediate reasoning steps. (c) Existing RL-based methods rely on expensive human annotations to train preference models for reward signals. (d) Our RL Binding mechanism achieves efficient multi-modal alignment without additional annotation cost.
  • Figure 2: Overview of the Motion-R1 framework. Our method introduces two key innovations: (1) a Decomposed CoT Data Engine that generates structured motion planning traces (including <think>, <output>, and <Motion> tokens) via LLM reasoning, enabling fine-grained temporal and causal decomposition; (2) an RL Binding mechanism with GRPO-based training that streamlines optimization via embedded multi-modal alignment, ensuring semantic accuracy and motion realism without human annotations.
  • Figure 3: Visualization comparisons with MotionLLM wang2024motionagent on in-distribution and out-of-distribution prompts.
  • Figure 4: Motion-R1 results on Out-of-Distribution prompts. Left: Complex caption with multi-step reasoning. Right: Abstract caption requiring semantic understanding. Motion-R1 captures structure and intent in both cases.