Author(s): LIN, K., CHEN, Y.M., HUNG, H.C., Institution: NATIONAL SUN YAT-SEN UNIVERSITY, Country: TAIWAN, Abstract-ID: 376

INTRODUCTION:
In response to the increasing demand for badminton-related college courses and the rapid advancement of large language models (LLMs), this study proposes a Badminton AI system to serve as a teacher’s agent. The system is developed in four stages. First, a badminton skill checkpoint angle analysis system is implemented to provide objective metrics for movement evaluation. Second, an LLM for badminton instruction (Badminton GPT) is established based on GPT 5.2. Third, analytical outputs from the angle analysis system are transmitted to the Badminton GPT to generate movement suggestions. In the fourth stage, a LINE-Bot-based teaching platform is developed for learners to interact with the system.
METHODS:
The LINE Bot developed in this study integrates multiple instructional modules, including expert demonstration videos, movement analysis, learning portfolios, reflection, and class preview. Learners can use this LINE Bot to access these modules, upload videos for angle analysis, and receive real-time feedback from the Badminton GPT. To examine the reliability of the skill checkpoint angle analysis system, two human experts independently evaluated the same set of 100 videos alongside the system. Inter-rater reliability was assessed using the intraclass correlation coefficient (ICC), adopting a two-way random-effects model with absolute agreement for single measures. Reliability was first examined between the two human expert raters and subsequently between the AI system and the human experts.
RESULTS:
The findings indicated that the ICC (2,1) for single measures was .97 (95% CI [.95, .98], p < .001), demonstrating near-perfect agreement between the expert raters. When the angle analysis system was included, inter-rater reliability between the AI system and the two human experts remained at a good level (ICC (2,1) = .81, 95% CI [.72, .87], p < .001). Based on these results, the experts further reviewed and confirmed the validity and usefulness of the feedback generated by the Badminton GPT.
CONCLUSION:
Two key findings emerged from this study. First, with near-perfect agreement between the two human experts, the skill checkpoint angle analysis system produced stable and consistent quantitative scores, demonstrating its reliability as a foundation for AI-assisted feedback. Second, the integration of the angle analysis system with Badminton GPT was feasible. Expert review confirmed that the generated feedback was valid and instructionally meaningful, indicating that exact replication of expert scoring is not necessary for effective AI-generated feedback. Future research should explore improved integration strategies between pose estimation and large language models, such as retrieval-augmented generation to provide richer contextual information or explainable machine learning approaches to better align the system’s scoring with expert judgments.