Abstract
Existing intelligent sports analysis systems mainly focus on "scoring and visualization," often lacking automatic performance diagnosis and interpretable training guidance. Recent advances in Large Language Models (LLMs) and motion analysis techniques provide new opportunities to address the above limitations. In this paper, we propose SportsGPT, an LLM-driven framework for interpretable sports motion assessment and training guidance, which establishes a closed loop from motion time-series input to professional training guidance. First, given a set of high-quality target models, we introduce MotionDTW, a two-stage time series alignment algorithm designed for accurate keyframe extraction from skeleton-based motion sequences. Subsequently, we design a Knowledge-based Interpretable Sports Motion Assessment Model (KISMAM) to obtain a set of interpretable assessment metrics (e.g., insufficient extension) by contrasting the keyframes with the target models. Finally, we propose SportsRAG, a RAG-based training guidance model built upon Qwen3. Leveraging a 6B-token knowledge base, it prompts the LLM to generate professional training guidance by retrieving domain-specific QA pairs. Experimental results demonstrate that MotionDTW significantly outperforms traditional methods with lower temporal error and higher IoU scores. Furthermore, ablation studies validate the KISMAM and SportsRAG, confirming that SportsGPT surpasses general LLMs in diagnostic accuracy and professionalism.
Introduction
Modern sports analysis systems excel at digitizing motion—tracking joints and calculating kinematic variables. However, they often suffer from a significant "interpretability gap." A raw kinematic plot or a 3D joint model does not inherently tell an athlete what they are doing wrong or how to improve. While Large Language Models (LLMs) offer potential for generating natural language guidance, feeding raw biomechanical data directly into generalist models often leads to hallucinations and generic, non-actionable advice. SportsGPT addresses these challenges by creating a semantic bridge between raw motion data and expert-level coaching.
The SportsGPT Framework
SportsGPT employs a three-stage architecture to transform raw video into professional training reports, establishing a closed loop from input to guidance.
1. MotionDTW: Precise Keyframe Extraction
Standard Dynamic Time Warping (DTW) often fails in unconstrained sports videos due to varying action speeds and redundant frames. MotionDTW introduces a two-stage coarse-to-fine strategy. It first identifies the general action interval using a sliding window and then performs fine-grained alignment. By utilizing a multi-modal feature space (angles, velocity, acceleration) and joint-specific weighting, it achieves high temporal precision, reducing average temporal error to just 0.91 frames.
2. KISMAM: The Semantic Bridge
The Knowledge-based Interpretable Sports Motion Assessment Model (KISMAM) compares extracted keyframes against elite reference models. Instead of outputting raw numerical deviations, it maps these differences to semantic labels (e.g., "insufficient start extension"). This step is crucial, as feeding raw data directly to LLMs significantly degrades diagnostic performance; the semantic bridge allows the model to reason effectively about multi-joint coordination.
3. SportsRAG: Actionable Guidance
The final module, SportsRAG, uses the semantic labels from KISMAM to query a massive 6-billion-token knowledge base. This corpus includes authoritative sports science textbooks, expert QA pairs, and historical professional reports. By retrieving relevant domain-specific context, the underlying LLM (Qwen3-8B) can synthesize professional, specific, and feasible training plans, such as detailed load-managed prescriptions.
Evaluation and Results
The framework was evaluated using the new Athletics Motion Analysis Benchmark, which includes 390 annotated video sequences of sprinting and jumping.
Performance and Accuracy
- Alignment Precision: MotionDTW achieved state-of-the-art accuracy with an ultra-low inference latency of 1.42 ms.
- Diagnostic Quality: In double-blind evaluations by senior sports specialists, SportsGPT outperformed state-of-the-art generalist LLMs (including GPT-5 and Claude 4.5 equivalents) across accuracy, professionalism, and feasibility. SportsGPT scored 3.80/5 for accuracy and 3.77/5 for feasibility.
Key Insights
Ablation studies confirmed that the RAG module is essential for turning generic fitness advice into expert-level prescriptions. Without it, the LLM's feasibility scores plummeted as it regressed to operationally vague suggestions.
Conclusion
SportsGPT establishes a robust, end-to-end pipeline for interpretable sports analysis. By combining precise motion alignment with specialized retrieval-augmented generation, it provides a scalable solution for professional-grade athletic coaching. Future work will focus on handling stylistic diversity through multi-reference voting and expanding the framework to more complex sports like gymnastics and ball games.
Summary updated