Scholarlink

<p>Human motion generation has gained significant attention for its applications in immersive content creation and AI-assisted coaching. However, generating long-term, domain-specific sports motions remains challenging due to their complex kinematic structure and sequential nature. To address this, we introduce TaiChiGPT, a text-to-motion framework that models TaiChi movements as tokenized sequences, enabling coherent long-term motion generation. Our approach integrates a kinematics-based prompt optimization strategy to improve segment-level realism and a two-stage refinement module for global motion quality. Experiments on a dedicated TaiChi motion dataset show that our method results in reductions of 42.7% in FID and 25.4% in DIV compared to existing approaches, demonstrating a substantial quantitative improvement. Additionally, the generated motion data attains a low average relative error of 1.67% in action recognition, as validated by two benchmark models. This work provides a scalable solution for generating complex sports movements and demonstrates the potential of language-inspired motion modeling for intelligent sports applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

TaiChiGPT: complex sports action generation based on large language models

  • 期刊论文

  • 发表日期：

    2026年3月16日

查看全文
  • Jianwei Li,
  • Kehao Ran,
  • Yanwen Ma,
  • Hongwen Xie,
  • Yihong Wu

摘要

Human motion generation has gained significant attention for its applications in immersive content creation and AI-assisted coaching. However, generating long-term, domain-specific sports motions remains challenging due to their complex kinematic structure and sequential nature. To address this, we introduce TaiChiGPT, a text-to-motion framework that models TaiChi movements as tokenized sequences, enabling coherent long-term motion generation. Our approach integrates a kinematics-based prompt optimization strategy to improve segment-level realism and a two-stage refinement module for global motion quality. Experiments on a dedicated TaiChi motion dataset show that our method results in reductions of 42.7% in FID and 25.4% in DIV compared to existing approaches, demonstrating a substantial quantitative improvement. Additionally, the generated motion data attains a low average relative error of 1.67% in action recognition, as validated by two benchmark models. This work provides a scalable solution for generating complex sports movements and demonstrates the potential of language-inspired motion modeling for intelligent sports applications.