A practical study of artificial intelligence-based real-time feedback in online physical education teaching

Jiewei Ma, Lianzhen Ma, Shilong Qi, Bo Zhang, Wenpian RuanSmart 2025 · DOI

Abstract To address poor skill acquisition in online physical education due to a lack of real-time feedback, we developed and evaluated a pose recognition-based system. An 8-week randomized controlled trial study in a university Baduanjin course compared the AI system against a traditional Massive Open Online Course format. Results showed the system significantly enhanced students' movement quality, fluency, learning interest, and self-directed learning. Crucially, mediation analysis identified increased learning duration as the primary significant mechanism driving this skill acquisition, outweighing changes in interest or self-direction within our model. While promising, the technology has limitations in accuracy and interactivity. Future research should focus on optimizing algorithms and integrating adaptive learning to create more effective OLPE strategies.

Introduction

Online Physical Education (OLPE) has long suffered from a "feedback dilemma." While students can watch high-quality demonstrations on platforms like MOOCs, the communication is strictly one-way. This leaves learners in a state of "practice blindness," where they are unable to accurately self-assess their form or rectify errors in real-time.

Unlike academic subjects where knowledge transmission is straightforward, physical education requires the acquisition of complex motor skills through an iterative "perception-action-feedback" loop. Without a coach present, students often fall into passive imitation rather than reflective practice. This study explores a web-based system designed to bridge this gap by using Artificial Intelligence to emulate the role of a persistent, real-time "personal coach."

System Design and Methodology

The researchers developed a feedback system built on MediaPipe’s BlazePose module. BlazePose was selected over alternatives like OpenPose due to its superior processing speed and accuracy on mobile and web platforms, which is critical for real-time interaction.

Technical Workflow

  1. Motion Database Construction: The team recorded 22 videos of expert Baduanjin (a traditional Chinese health exercise) practitioners. They extracted skeletal keypoints and calculated kinematic features such as joint angles and inter-joint distances.
  2. Expert Validation: Three senior Qigong experts annotated the data to establish a "ground truth." A Bland-Altman analysis confirmed a high degree of agreement (over 93%) between the AI’s automated assessments and the experts' scores.
  3. The Feedback Loop: The front-end (JavaScript) overlays a real-time skeletal model on the student’s webcam feed. As students practice, the system checks their posture against the expert database. When a movement criterion is met, the corresponding text on the screen turns green, providing immediate "Knowledge of Performance" (KP).
Skeletal keypoint mapping (left) and real-time pose recognition overlay on a practitioner (right).

Experimental Evaluation

A 10-week Randomized Controlled Trial (RCT) was conducted with 70 university students (non-PE majors). The experimental group used the AI system alongside MOOC videos, while the control group used only the MOOC videos with delayed instructor feedback. After attrition, the final analysis included 60 participants (n=28 experimental, n=32 control).

Key Results

  • Skill Performance: The AI group achieved significantly higher total performance scores ($M=6.92$) compared to the control group ($M=6.48$, $p=.018$). This superiority held for both movement quality and execution fluency.
  • Behavioral Engagement: Students using the AI system practiced significantly longer ($23.87$ mins/week) than the MOOC-only group ($21.25$ mins/week, $p=.016$).
  • Psychological Impact: The experimental group showed significant improvements in learning motivation, sports attention, and positive learning attitudes, whereas the control group showed no significant changes in these areas.

Discussion: The "Practice Volume" Revelation

One of the most important theoretical takeaways is that the AI acts as a "cognitive scaffold." By automating the frustrating process of self-diagnosis, the system reduces extraneous cognitive load. This frees up the student's working memory to focus on "germane load"—the actual internalization of the motor skill.

Crucially, a mediation analysis revealed that psychological factors weren't the direct cause of better skills. Instead, the AI system's primary impact was increasing total learning duration. Because the experience was less taxing and more rewarding (game-like), students simply practiced more. This increased "time-on-task" was the actual engine of skill mastery. Furthermore, students reported that the AI made them aware of mistakes they weren't previously conscious of, suggesting it acts as a catalyst for kinesthetic awareness.

Limitations and Future Work

The study identifies several hurdles for current AI feedback systems:

  • Technical Constraints: Pose recognition can struggle in constrained spaces or during complex movements where limbs occlude one another.
  • The "Human" Gap: Qualitative interviews revealed that the AI lacks emotional support. It can identify what is wrong, but cannot provide the nuanced encouragement or deep explanations a human coach offers.
  • Generalizability: The study focused on Baduanjin, which is slow and form-based. Future research is needed to determine if this approach works for dynamic, open-skilled sports like basketball.

Conclusion

The future of physical education is likely a blended ecosystem where AI handles high-frequency technical feedback, allowing human instructors to focus on holistic motivation and high-level coaching. By reducing the cognitive barriers to practice, AI can transform online learning from passive observation into active, data-driven mastery.

Summary updated

Abstract

Abstract To address poor skill acquisition in online physical education due to a lack of real-time feedback, we developed and evaluated a pose recognition-based system. An 8-week randomized controlled trial study in a university Baduanjin course compared the AI system against a traditional Massive Open Online Course format. Results showed the system significantly enhanced students' movement quality, fluency, learning interest, and self-directed learning. Crucially, mediation analysis identified increased learning duration as the primary significant mechanism driving this skill acquisition, outweighing changes in interest or self-direction within our model. While promising, the technology has limitations in accuracy and interactivity. Future research should focus on optimizing algorithms and integrating adaptive learning to create more effective OLPE strategies.
BibTeX
@inproceedings{ma2025practical,
  author = {Jiewei Ma and Lianzhen Ma and Shilong Qi and Bo Zhang and Wenpian Ruan},
  title = {{A practical study of artificial intelligence-based real-time feedback in online physical education teaching}},
  booktitle = {Smart},
  year = {2025},
  doi = {10.1186/s40561-025-00411-3},
}
Figure