Content area
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
While the pursuit of educational equity has been propelled by the global expansion of online learning, its promise remains largely unfulfilled in domains requiring embodied competence, such as Physical Education (PE). Unlike disciplines amenable to knowledge transmission, PE is centered on the acquisition of complex motor skills—a process that hinges on iterative practice and personalized, corrective feedback (Hodges & Franks, 2004). Conventional Online Physical Education (OLPE), typically reliant on one-way video demonstrations (e.g., Massive Open Online Courses (MOOCs)), fails to replicate this essential interactive loop. This leaves learners in a state of "practice blindness," unable to accurately self-assess or rectify performance errors. This deficit impedes skill development, undermines motivation, and fosters passive imitation over reflective practice (Lin et al., 2022, 2023a, 2023b, 2023c; Xie, 2021).
Efforts to innovate OLPE, including blended learning and flipped classrooms, have shown some promise but have not fully resolved the core challenge of providing scalable, real-time motor skill guidance (Daum, 2020; Hinojo Lucena et al., 2020). The recent surge in AI applications in education offers a more direct solution. AI-powered systems, leveraging pose recognition technology, can now emulate a core function of expert instruction: providing automated assessment and individualized feedback (Huang et al., 2023; Lin et al., 2023a, 2023b, 2023c). However, existing research has focused on technological implementation, largely neglecting the pedagogical and psychological mechanisms of learning. This study bridges that gap by investigating an integrated theoretical framework. We contend that the efficacy of AI-based feedback is best understood through the complementary lenses of motor learning theory (Schmidt et al., 2018; Wulf et al., 2010), cognitive load theory (CLT) (Sweller, 1988), and self-determination theory (SDT) (Ryan & Deci, 2000).
This theoretical framing is particularly salient in the context of traditional practices like Baduanjin, a health-preserving exercise increasingly adopted in Chinese school curricula. The gentle yet precise movements of Baduanjin demand high fidelity in execution, which is difficult to achieve through simple video imitation. From a motor learning perspective, an AI system can provide the immediate and specific Knowledge of Performance (KP) necessary for learners to refine their postures (Schmidt et al., 2018). From a cognitive science viewpoint, this automated guidance is hypothesized to reduce the extraneous cognitive load associated with self-diagnosis, allowing learners to allocate mental resources to skill internalization (Sweller, 1988). Finally, from a motivational standpoint, by enabling successful task execution and offering flexible practice opportunities, the system is expected to bolster learners' sense of competence and autonomy, thus fostering intrinsic interest in line with SDT (Ryan & Deci, 2000).
Grounded in this multi-faceted framework, the present study developed and evaluated an AI-powered real-time feedback system based on MediaPipe for an online Baduanjin course. Employing a randomized controlled trial, this research was guided by the following research questions (RQs):
RQ1: Does the AI-based feedback system lead to greater improvements in students' motor skill performance compared to a traditional MOOC-based approach?
RQ2: Does the AI-based feedback system enhance students' learning interest and self-directed learning behaviors more effectively than a traditional MOOC?
RQ3: What are the mediating mechanisms through which the AI system facilitates motor skill acquisition?
RQ4: How do students perceive the learning experience within the AI-assisted OLPE environment?
By systematically addressing these questions, this study aims to provide not only robust empirical evidence for the utility of pose recognition technology in OLPE but also a theoretically grounded explanation for its effectiveness, offering valuable insights for the design of future intelligent learning environments.
Literature review
Research on online physical education
The rapid global adoption of Online Physical Education (OLPE), accelerated by the COVID-19 pandemic, has yielded a complex and dualistic body of research (Tegero, 2021). On one hand, studies have validated its effectiveness, demonstrating positive impacts on student attitudes, physical activity levels, fitness, and motivation under various models, including synchronous and blended formats (e.g., Beserra et al., 2022; Horita et al., 2025; Killian, 2023; Lee, 2021; Qi, 2024; Vorlíček et al., 2024; Suherman, 2021; Zheng et al., 2021). On the other hand, a stronger consensus has emerged regarding OLPE's profound pedagogical limitations. The prevalent reliance on pre-recorded videos fosters a non-interactive and socially isolating learning environment, hindering deep engagement and authentic skill practice (e.g., Bowles et al., 2024; Lambert & Luguetti, 2022; Varea & Gonzalez-Calvo, 2022).
A systematic review by Ovens et al. (2022) underscores that the central challenge lies in the need for a fundamental pedagogical transformation, moving beyond mere content delivery (see also Baker et al., 2022). The crux of this issue is the feedback dilemma: the inherent inability of standard OLPE models to provide the timely, specific, and corrective guidance essential for motor skill acquisition (Daum & Buschner, 2012). This failure to close the "perception–action-feedback" loop constitutes the modality's primary pedagogical obstacle. Therefore, addressing this feedback gap is a critical prerequisite for the sustainable development of high-quality OLPE.
AI-powered feedback systems for motor skill learning
To address the feedback dilemma in OLPE, researchers have increasingly turned to Artificial Intelligence. AI's potential to enhance personalized learning and instructional effectiveness has been well-documented across various disciplines (Eltahir & Babiker, 2024; Lin, 2022; Liu & Yushchik, 2024). More specifically, within the burgeoning field of motor skill learning, AI-powered systems leveraging pose recognition are now capable of providing the real-time, individualized feedback that was previously absent. A growing body of evidence suggests these systems effectively improve movement precision, physical fitness, learning motivation, and self-directedness in activities ranging from yoga to university PE courses (e.g., Hsia et al., 2024; Li, 2021; Liu et al., 2024; Ompoc & Aguinaldo, 2025; Hsia & Hwang, 2020).
However, despite these promising outcomes, a critical analysis of the literature reveals a significant theoretical and methodological gap. The predominant focus has been on technological implementation and performance outcomes, often treating the learning process as a "black box" (Hu et al., 2024; Wang & Wang, 2024). Consequently, fundamental questions remain largely unanswered: Why is this feedback effective? How does it precisely influence a learner's cognitive and motivational states? This lack of an explicit theoretical grounding not only hinders the systematic optimization of these systems but also limits the generalizability of their findings. Therefore, a shift from a purely technology-centric evaluation to a theoretically-grounded investigation is imperative to truly understand and harness the pedagogical power of AI in physical education.
Theoretical framework
To address the aforementioned research gap and move beyond a "black box" analysis, this study is guided by an integrated theoretical framework. We synthesize three complementary theories to explain how and why AI-based feedback influences skill acquisition, positing a causal chain that operates sequentially across behavioral, cognitive, and motivational levels.
The framework is first grounded in motor learning theory, which establishes the behavioral foundation. For complex, form-dependent skills like Baduanjin, effective learning hinges on receiving Knowledge of Performance (KP)—specific feedback about the quality of the movement pattern itself—which is demonstrably more critical than mere Knowledge of Results (KR) (Schmidt et al., 2018; Wulf et al., 2010). The AI system in this study is designed precisely to provide this essential KP, functioning as an immediate and high-fidelity source of corrective guidance.
However, the delivery of KP must be cognitively manageable. This introduces our second pillar, Cognitive Load Theory. Traditional OLPE imposes a high extraneous cognitive load, as learners must struggle to self-diagnose errors. By automating this diagnostic process, the AI system acts as a cognitive offloading tool, significantly reducing this unproductive load. This frees learners' limited working memory to engage in germane cognitive load—the deep processing required for skill internalization.
Finally, a cognitively efficient process must also be motivationally compelling. Our framework thus culminates with SDT, which posits that intrinsic motivation is fueled by the satisfaction of basic psychological needs, particularly competence and autonomy (Ryan & Deci, 2000). The AI system is designed to nurture these needs directly: it bolsters competence by providing clear pathways to mastery and supports autonomy by offering learners full control over their practice. By satisfying these needs, the system is hypothesized to foster the high-quality motivation necessary for sustained engagement and persistence.
In synthesis, this framework portrays a synergistic, multi-layered process. The AI system initiates a virtuous cycle: effective behavioral guidance (KP) makes the learning process less cognitively demanding (CLT), which in turn fosters a sense of mastery and control that fuels intrinsic motivation (SDT). This heightened motivation is then expected to translate into increased practice volume—the ultimate driver of skill acquisition. To operationalize these theoretical principles, the following section details the design and architecture of the AI-powered feedback system developed for this study, which serves as the concrete embodiment of this framework.
Real-time motion feedback system based on pose recognition
Mediapipe
To operationalize our theoretical framework, we developed a real-time motion feedback system for Baduanjin instruction. The system is built upon MediaPipe, an open-source framework from Google designed for processing time-series data like video (Google Developers, 2023). We specifically selected MediaPipe's BlazePose module for its superior performance compared to alternatives like OpenPose. BlazePose offers a compelling combination of higher accuracy on benchmark datasets, significantly faster processing speeds, and robust cross-platform support, making it well-suited for a real-time, web-based learning application (Bazarevsky et al., 2020; Zheng et al., 2020). Its efficacy in academic research for tasks such as 3D posture detection and movement classification is also well-established (e.g., Garg et al., 2023; Kim et al., 2023; Lin et al., 2023a, 2023b, 2023c). As illustrated in Fig. 1, the full-body pose model used in this study extracts a rich set of keypoints to enable detailed kinematic analysis.
[See PDF for image]
Fig. 1
Mediapipe human skeleton
Construction and validation of the motion database
The core of the feedback system is a motion database that defines the "correct" execution of Baduanjin movements. Its construction and validation followed a three-step process.
The system first utilized MediaPipe to extract raw skeletal keypoint coordinates from a corpus of 22 videos featuring proficient Baduanjin practitioners. From these coordinates, key kinematic features—specifically, critical joint angles and inter-joint distances relevant to each movement's quality—were calculated using vector dot product and Euclidean distance formulas. Following previous research, we selectively analyzed only the most salient features for each movement to optimize computational efficiency (Hsia et al., 2024). Figure 2 illustrates the key features used for classifying movement quality.
[See PDF for image]
Fig. 2
Analysis of the indicators
Three certified senior experts in Fitness Qigong were then invited to independently annotate these extracted kinematic features. Based on the official Fitness Qigong Competition Rules, they evaluated and classified the practitioners' movement quality for each video, establishing a "ground truth" dataset. This expert-annotated database serves as the standard for the system's automated assessment.
To validate the system's accuracy against this standard, a Bland–Altman analysis was conducted to validate the consistency between the expert-annotated data and the system-annotated data, as shown in Table 1. The analysis revealed a high degree of agreement between the two methods. The mean bias was consistently small across all evaluated movements (ranging from − 0.22° to 7.90°), and the 95% limits of agreement were deemed acceptable for pedagogical purposes. For example, the 95% LoA for TiaoLiPiWei ranged from − 19.49 to 25.77. Critically, over 93% of the differences for each movement fell within these calculated limits, confirming the system's reliability for use in the intervention.
Table 1. Bland–Altman verification of the four movements
Baduanjin movements | Bias | LoA upper | LoA lower | In LoA (%) |
|---|---|---|---|---|
Yubei | 2.85 | 42.65 | − 36.96 | 97.37 |
ShuangShouTuoTian | 7.90 | 32.78 | − 16.98 | 96.02 |
TiaoLiPiWei | 3.14 | 25.77 | − 19.49 | 93.75 |
WuLaoQiShang | − 0.22 | 18.29 | − 18.74 | 94.64 |
Interface and functionality
The system was designed as a web-based application to ensure cross-platform compatibility. Its architecture comprises two main components:
Back-End: The server-side infrastructure houses the core logic of the system. This includes the expert-validated motion database, the MediaPipe integration module for processing the video stream, and, most critically, the feedback algorithm that compares learner performance against the database criteria.
Front-End: The client-side user interface, developed in JavaScript, provides the learning environment. It integrates the webcam feed, a video player for instructional content, and the real-time motion feedback display.
As illustrated in Fig. 3, the learner interface is designed for intuitive use. The left panel features a video player displaying the instructor's demonstration of a Baduanjin movement. The right panel displays the learner's live webcam feed, overlaid with a real-time skeletal model generated by MediaPipe. At the bottom of the screen, a list of key performance criteria for the current movement is displayed. As the student practices, the system provides immediate visual feedback: when a performance criterion is successfully met and maintained, its corresponding text turns green. Once all criteria for a movement are satisfied, the system unlocks the next exercise, creating a structured, mastery-oriented learning progression.
[See PDF for image]
Fig. 3
Program interface
Materials and methods
Participants
A total of 70 s-year undergraduate students (non-physical education majors, aged 18–20) were recruited from a general physical education course pool at a university in Guangzhou. Recruitment was conducted via an open invitation on the university's course enrollment platform for a 10-week online Baduanjin study. The primary inclusion criterion was having no prior experience in Baduanjin or other forms of Qigong. All participants provided written informed consent before the study commenced.
Following the baseline assessment, the 70 participants were individually randomized in a 1:1 ratio to either the AI-Feedback Group (experimental) or the MOOC-Only Group (control). The randomization sequence was generated using a computerized random number generator by a research assistant independent of the study's implementation and data analysis, ensuring allocation concealment. The study protocol was approved by the Academic Committee of the School of Physical Education and Sports Science (SCNU-SPT-2023-0082). As shown in Table 2, the randomization process successfully resulted in two groups that were well-balanced on baseline demographic characteristics.
Table 2. Participant characteristics
Group | N | Male | Female | Age | Major |
|---|---|---|---|---|---|
Experimental group | 35 | 15 | 20 | 18.61 ± 0.85 | Management, Economics, Engineering |
Control group | 35 | 18 | 17 | 18.84 ± 0.83 | Science, Education, Engineering, Economics |
Throughout the 10-week intervention, 10 participants withdrew from the study (seven citing scheduling conflicts and three failing to complete the required assessments), leading to a final analysis sample of 60 participants (n = 28 in the AI-Feedback group; n = 32 in the MOOC-Only group). A flow diagram illustrating participant progression through the trial is provided in Fig. 4.
[See PDF for image]
Fig. 4
Flow diagram of participant progression through the study
Experimental procedure and intervention
The study followed a 10-week, pre-test/post-test randomized controlled trial design. Week 1 (Baseline): All participants, regardless of their future group assignment, attended an initial online orientation session. They were introduced to the study's objectives and the basics of Baduanjin movements. Following this, they provided informed consent and completed a baseline assessment (pre-test) measuring their interest in physical education, self-directed learning ability, and initial Baduanjin proficiency. Weeks 2–9 (Intervention Period): After the baseline assessment, participants were informed of their group allocation. All participants were required to engage in two hours of self-directed learning per week for eight weeks, using materials hosted on the iCourse163 MOOC platform, which included instructor-led demonstration videos.
The intervention differed as follows:
The AI-Feedback Group (Experimental): In addition to the standard MOOC videos, participants in this group utilized the AI-powered real-time feedback system during their practice sessions. The system provided immediate, corrective feedback on their Baduanjin movements.
The MOOC-Only Group (Control): Participants in this group engaged in self-study using only the MOOC platform's video resources. As is standard practice in such courses, they submitted videos of their practice periodically for instructor review, receiving delayed, general feedback.
Week 10 (Post-test and Interviews): Both groups completed a post-test assessment identical to the pre-test. Following the post-test, a subset of participants from the experimental group was invited to participate in semi-structured interviews to share their learning experiences.
To minimize scoring bias, the two instructors responsible for evaluating the Baduanjin performance videos at both pre-test and post-test were blinded to the participants' group allocation.
Measurement tools
Self-directed learning and learning interest are considered key factors influencing learning outcomes (Ardito & Czerkawski, 2021; Owoc et al., 2019). Therefore, this study evaluates students' learning outcomes using three indicators: interest in physical education, self-directed learning ability in physical education, and Baduanjin movement performance scores. Additionally, learning duration in the online course is used as a process indicator. According to the Fitness Qigong Competition Rules, key criteria for evaluating movement performance include joint angles (movement quality), execution speed, and relaxation level (fluency of execution).
Physical Education Learning Interest.
Learning interest was assessed using the College Students' Physical Education Learning Interest Scale (Gu & Xie, 2012). This scale includes five dimensions and has demonstrated good reliability (test–retest Cronbach’s α = 0.849) and construct validity (KMO = 0.922).
Self-Directed Learning in Physical Education.
Self-directed learning ability was measured with the College Students' Self-Directed Learning in Physical Education Scale (Wu, 2010). This instrument assesses four dimensions and has established reliability (test–retest Cronbach’s α = 0.886) and validity (KMO = 0.782).
Baduanjin Performance Scoring.
Students' Baduanjin performance was evaluated based on the Fitness Qigong Baduanjin Competition Rules and Judging Criteria. Scores comprised two components: movement quality (5 points) and fluency of execution (5 points), for a total of 10 points. Three certified national-level Fitness Qigong judges independently scored video recordings of the routines. Inter-rater reliability was high, with an Intraclass Correlation Coefficient (ICC) of 0.921 at pre-test and 0.884 at post-test. A student's final score was the average of the three judges' ratings.
Semi-Structured Interviews
The interviews consisted of five questions and were conducted exclusively with students from the experimental group. Each interview lasted 10 min, with the number of participants and timestamps recorded sequentially. The interview content was analyzed and coded by two researchers, who categorized responses based on relevant themes. In cases where discrepancies arose in coding, the researchers re-evaluated the content and reached a consensus.
Interview Questions:
What was your previous experience with OLPE courses? What challenges or difficulties did you encounter?
Did you experience any technical difficulties when using the pose recognition technology? If so, could you describe them?
Compared to traditional learning methods (e.g., video-based learning or in-person classes), what are the advantages and disadvantages of pose recognition-assisted learning?
After using the pose recognition technology, do you feel that your learning interest and autonomy have improved? Would you be willing to continue using this technology for future movement skill learning?
What improvements would you like to see in this technology in the future? How do you think it will impact OLPE?
Statistical analysis
Data analysis was conducted using SPSS 27.0 and Python 3.10. An independent sample t-test was performed to examine differences between groups and the relationship between teachers and students' data. A paired sample t-test was used to analyze pre- and post-test differences within the same group. To investigate whether the effect of pose recognition technology on Baduanjin skill learning was mediated by learning interest and self-directed learning, a parallel mediation analysis was conducted. This analysis examined the mediating effects of learning interest, self-directed learning, and learning duration on skill learning outcomes.
Results
Baseline equivalence
Prior to the intervention, independent samples t-tests confirmed that there were no significant baseline differences between the AI-Feedback group and the MOOC-Only group across all sub-dimensions of sports learning interest, self-directed learning, and Baduanjin skill performance (all ps > .05, see Table 3, p0 column). The two groups were thus comparable at the outset.
Table 3. Comparison within and between groups
Outcomes | G | W0 | W8 | t0 | p0 | t1 | p1 | t2 | p2 |
|---|---|---|---|---|---|---|---|---|---|
1. Sports learning interest | |||||||||
negative sports learning | E | 2.43 ± 0.67 | 2.02 ± 0.96 | 0.64 | 0.52 | 1.94 | 0.06 | 1.35 | 0.180 |
C | 2.33 ± 0.57 | 2.31 ± 0.69 | 0.14 | 0.89 | |||||
positive sports learning | E | 3.18 ± 0.42 | 4.01 ± 0.66 | 0.25 | 0.80 | − 6.04 | < 0.001 | − 5.46 | < 0.001 |
C | 3.14 ± 0.53 | 3.18 ± 0.52 | − 0.30 | 0.76 | |||||
sports skills learning | E | 3.50 ± 0.47 | 4.00 ± 0.60 | 1.21 | 0.23 | − 3.67 | < 0.001 | − 3.78 | < 0.001 |
C | 3.35 ± 0.53 | 3.41 ± 0.60 | − 0.46 | 0.65 | |||||
extracurricular sports activities | E | 3.18 ± 0.62 | 3.76 ± 0.55 | − 1.12 | 0.27 | − 3.90 | < 0.001 | − 2.29 | 0.025 |
C | 3.37 ± 0.81 | 3.46 ± 0.47 | − 0.56 | 0.58 | |||||
sports attention | E | 2.99 ± 0.68 | 3.84 ± 0.61 | − 0.66 | 0.51 | − 5.05 | < 0.001 | − 3.64 | < 0.001 |
C | 3.11 ± 0.75 | 3.19 ± 0.74 | − 0.47 | 0.64 | |||||
2. Sports self-directed learning | |||||||||
learning motivation | E | 3.28 ± 0.55 | 3.92 ± 0.64 | 0.53 | 0.60 | − 4.14 | < 0.001 | − 4.52 | < 0.001 |
C | 3.22 ± 0.51 | 3.30 ± 0.41 | − 0.72 | 0.48 | |||||
learning process | E | 3.58 ± 0.52 | 3.82 ± 0.54 | 1.87 | 0.07 | − 1.77 | 0.08 | − 2.33 | 0.024 |
C | 3.36 ± 0.45 | 3.50 ± 0.54 | − 1.11 | 0.27 | |||||
learning outcomes | E | 3.41 ± 0.49 | 3.27 ± 0.69 | 0.97 | 0.34 | 0.93 | 0.36 | − 0.07 | 0.940 |
C | 3.30 ± 0.45 | 3.26 ± 0.33 | 0.43 | 0.67 | |||||
learning environment | E | 3.97 ± 0.52 | 4.17 ± 0.55 | 1.52 | 0.13 | − 1.44 | 0.15 | − 2.02 | 0.048 |
C | 3.78 ± 0.50 | 3.88 ± 0.57 | − 0.72 | 0.48 | |||||
3. Baduanjin score | |||||||||
movement specifications | E | 2.70 ± 0.40 | 3.47 ± 0.35 | 1.28 | 0.21 | − 8.41 | < 0.001 | 2.25 | 0.028 |
C | 2.57 ± 0.35 | 3.27 ± 0.35 | − 10.96 | < 0.001 | |||||
demonstration level | E | 2.59 ± 0.37 | 3.45 ± 0.39 | 0.84 | 0.41 | − 12.84 | < 0.001 | 2.44 | 0.018 |
C | 2.55 ± 0.37 | 3.21 ± 0.37 | − 10.68 | < 0.001 | |||||
total score | E | 5.31 ± 0.81 | 6.92 ± 0.71 | 1.07 | 0.29 | − 12.62 | < 0.001 | 2.43 | 0.018 |
C | 5.10 ± 0.73 | 6.48 ± 0.70 | − 11.11 | < 0.001 | |||||
4. Duration of participation in learning | |||||||||
duration | E | – | 23.87 ± 3.48 | – | – | – | – | 2.49 | 0.016 |
C | – | 21.25 ± 4.63 | – | – | – | – | |||
(G: Groups, E: Experimental group, C: Control group, W0: Pre-experiment, W8: Post-experiment,
t0, p0: t-values and p-values for baseline, t1, p1: t-values and p-values for pre- and post-intervention comparisons, t2, p2: t-values and p-values for between-group comparisons)
Intervention effects
Within-group changes
Paired samples t-tests revealed significant improvements from pre-test (W0) to post-test (W8) for the AI-Feedback group in most dimensions of sports learning interest (p < .001, except for 'negative sport learning'), learning motivation (p < .001), and Baduanjin performance scores (p < .001). In contrast, the MOOC-Only group only showed a significant improvement in Baduanjin performance (p < .001), with no significant changes in learning interest or self-directed learning dimensions (p > .05) (see Table 3, p1 column).
Between-group comparisons at post-test:
Independent samples t-tests (df = 58) were used to compare post-test outcomes. All descriptive statistics are detailed in Table 3.
Psychological Outcomes: The AI-Feedback group reported significantly higher scores than the MOOC-Only group on four dimensions of sports learning interest (positive learning: t = − 5.46, p < .001; skills learning: t = − 3.78, p < .001; extracurricular sports: t = − 2.29, p = .025; attention: t = − 3.64, p < .001) and three dimensions of self-directed learning (motivation: t = − 4.52, p < .001; process: t = − 2.33, p = .024; environment: t = − 2.02, p = .048).
Baduanjin Performance: Most importantly, the AI-Feedback group (M = 6.92, SD = 0.71) achieved significantly higher total performance scores than the MOOC-Only group (M = 6.48, SD = 0.70), t(58) = 2.43, p = .018. This superiority was consistent for both movement quality (t = 2.25, p = .028) and execution fluency (t = 2.44, p = .018).
Learning Duration: Behaviorally, the AI-Feedback group (M = 23.87, SD = 3.48) dedicated significantly more time to practice than the MOOC-Only group (M = 21.25, SD = 4.63), t(58) = 2.49, p = .016.
Mediation analysis
The analysis revealed that the intervention had significant positive effects on all three proposed mediators(see Table 4): self-directed learning (B = 0.31, p = .004), learning interest (B = 0.53, p < .001), and learning duration (B = 2.62, p = .016).
Table 4. Analysis of the mediation effect of pose recognition on skill acquisition
Path | Regression coefficient | SE | t | p | 95% Confidence interval (LLCI, ULCI) | Standardized regression coefficient |
|---|---|---|---|---|---|---|
X → M1 | 0.3115 | 0.1024 | 3.0421 | 0.0035 | (0.1065, 0.5164) | 0.7373 |
X → M2 | 0.5310 | 0.1317 | 4.0322 | 0.0002 | (0.2674, 0.7946) | 0.9301 |
X → M3 | 2.6181 | 1.0500 | 2.4933 | 0.0155 | (0.5162, 4.7200) | 0.6185 |
M1 → Y | 0.1653 | 0.2753 | 0.6005 | 0.5506 | (− 0.3864, 0.7170) | 0.0952 |
M2 → Y | 0.1065 | 0.2144 | 0.4968 | 0.6213 | (− 0.3231, 0.5361) | 0.0829 |
M3 → Y | 0.1336 | 0.0148 | 9.0367 | < 0.001 | (0.1040, 0.1632) | 0.7707 |
X → Y | − 0.0145 | 0.1405 | − 0.1029 | 0.9184 | (− 0.2961, 0.2672) | − 0.0197 |
X: pose recognition, Y: skill score, M1: self-directed learning in physical education, M2: learning interest in physical education, M3: learning duration
The total effect of the intervention on skill scores was significant. However, the direct effect of the intervention on skill scores, after accounting for the mediators, was not significant (B = − 0.01, p = .918), suggesting that the intervention's impact was fully mediated.
To test the significance of the indirect effects, we examined the 95% bootstrap confidence intervals (see Table 5 and Fig. 5). The analysis revealed only one significant indirect path: The indirect effect through learning duration (M3) was significant (Point Estimate = 0.35, 95% CI [0.067, 0.673]). The indirect effects through self-directed learning (M1) (95% CI [− 0.110, 0.227]) and learning interest (M2) (95% CI [− 0.154, 0.323]) were not significant, as their confidence intervals contained zero.
Table 5. Analysis of indirect effects
Indirect Path | Indirect Effect | BootSE | BootLLCI | BootULCI | Standardized Indirect Effect |
|---|---|---|---|---|---|
X → M1 → Y | 0.0515 | 0.0825 | − 0.1098 | 0.2272 | 0.0702 |
X → M2 → Y | 0.0566 | 0.1167 | − 0.1536 | 0.3230 | 0.0771 |
X → M3 → Y | 0.3498 | 0.1551 | 0.0671 | 0.6729 | 0.4767 |
Total Indirect Effect | 0.4579 | 0.1818 | 0.1288 | 0.8405 | 0.6239 |
X: pose recognition, Y: skill score, M1: self-directed learning in physical education, M2: learning interest in physical education, M3: learning duration
[See PDF for image]
Fig. 5
Mediator effect pathway
Furthermore, a formal pairwise comparison of these indirect effects (see Table 6) indicated no significant differences between any of the paths. These results indicate that the increased learning duration was the sole significant mediator explaining the positive effect of the AI-based feedback system on motor skill performance in this model.
Table 6. Comparison of the indirect effects
Comparison | Difference | BootSE | BootLLCI | BootULCI |
|---|---|---|---|---|
M1–M2 | − 0.0051 | 0.1840 | − 0.4181 | 0.3427 |
M1–M3 | − 0.2983 | 0.1868 | − 0.6733 | 0.0575 |
M2–M3 | − 0.2933 | 0.1779 | − 0.6392 | 0.0625 |
M1: self-directed learning in physical education, M2: learning interest in physical education, M3: learning duration
Interview results
To complement the quantitative findings, semi-structured interviews were conducted with participants in the experimental group. Thematic analysis of the interview transcripts revealed four key themes that illuminated students' experiences with the AI-assisted learning system.
Theme 1: The "Double-Edged Sword" of Traditional OLPE: Convenience vs. Lack of Feedback
Before discussing the AI system, students reflected on their general experiences with traditional online physical education (OLPE). A strong consensus emerged viewing OLPE as a "double-edged sword." While nearly all students appreciated the convenience and scheduling flexibility, they consistently identified a critical flaw: the absence of real-time guidance and interaction. This lack of feedback led to feelings of uncertainty and inefficiency in their learning. As one student articulated, "When practicing movements, I often feel unsure if I'm doing them correctly, and without immediate feedback… my learning efficiency is low" (S02). This sentiment was echoed by another who highlighted the core challenge: "The biggest challenge is not knowing whether I’m doing the movements correctly. Often, I make mistakes without realizing it" (S21). The lack of interaction also fostered a sense of isolation, with one participant noting, "I miss the interactive atmosphere of practicing with classmates—it feels lonely" (S22).
Theme 2: AI as a "Personal Coach": The Value and Limitations of Real-Time Feedback
When discussing the pose recognition system, students overwhelmingly framed its primary value as providing the real-time, corrective feedback that was missing from traditional OLPE. Many likened the system to a "personal coach" that could instantly identify and correct errors. "I think pose recognition helps me better understand my movements and correct mistakes in real time," stated one participant, adding, "This is crucial for Baduanjin, where precision is essential" (S01). This real-time assessment was seen as a key mechanism for improvement: "The technology made me aware of many mistakes I wasn't previously conscious of, helping me improve my Baduanjin techniques" (S05).
However, students also astutely recognized the technology's current limitations. They frequently mentioned issues with recognition accuracy, particularly "in constrained spaces or when performing complex movements" (S01). Beyond technical glitches, they also noted a lack of human-like qualities. As one student put it, the system "lacks emotional support" (S19), while another pointed out that when "encountering complex issues, the system cannot provide direct explanations or assistance" (S23), highlighting a gap in its interactive and pedagogical capabilities.
Theme 3: From Passive Viewing to Active Engagement: The Motivational Impact of AI Feedback
The shift from passive video-watching to an interactive feedback loop had a profound impact on students' learning motivation and engagement. The quantitative findings of increased learning interest were strongly corroborated by students' qualitative accounts. The real-time feedback loop was described as a powerful motivational driver. "Real-time feedback significantly increased my interest. When the system highlights my mistakes, I become more motivated to correct them," one student explained, "I feel that this technology makes me more proactive in learning" (S01). This process transformed practice into a more engaging, game-like experience, where feedback provided a sense of progress and achievement. As another participant shared, "Seeing my progress gives me a strong sense of achievement" (S23), which in turn "encouraged me to practice outside of class" (S04).
Theme 4: A Promising Future: Suggestions for a Smarter and More Personalized System
Finally, students expressed considerable optimism about the future of this technology, while also offering concrete suggestions for improvement. A recurring theme was the need for greater precision and intelligence. "I hope the system can be more precise in recognizing subtle movements, especially during fast actions," one student suggested (S03). Others envisioned a more interactive and adaptive system, with suggestions to "add a virtual coach feature" (S06) or create a system that could "adapt to different body types and exercise habits" (S14). The overarching sentiment was that with further development, this technology could "transform online physical education by providing personalized feedback for each student" (S03), extending beyond Baduanjin to other activities like martial arts or general fitness.
Discussion
Summary of principal findings
The present study demonstrates that an AI-driven, real-time feedback system significantly enhances motor skill performance in an online Baduanjin course compared to a traditional MOOC. Crucially, mediation analysis revealed that this improvement was primarily driven by an increase in students' total learning duration, rather than by direct psychological factors like interest or self-directedness. These quantitative findings were corroborated by qualitative data from student interviews.
AI-driven feedback as a catalyst for motor skill learning
A central finding of this study is the superior efficacy of the AI-driven feedback system in enhancing motor skill acquisition compared to traditional online instruction. This result strongly supports and extends previous research demonstrating the positive impact of real-time, technology-mediated feedback on learning outcomes in physical activities (Hsia & Hwang, 2020; Lin et al., 2023a, 2023b, 2023c; Yang et al., 2021). The qualitative data from our study provide a rich explanation for this quantitative success. Students consistently described the traditional MOOC experience as fraught with uncertainty, lamenting that they "often feel unsure if I'm doing them correctly" (S02) and have no way to resolve movement errors.
In stark contrast, the AI system functioned as a personalized and persistent "coach," fundamentally altering the learning process. By providing immediate visual and textual feedback, the system translated the abstract standards of "correct form" into concrete, actionable information for each learner. This mechanism addresses a core challenge in motor learning: the development of kinesthetic awareness (proprioception). As students noted, the technology made them "aware of many mistakes I wasn't previously conscious of" (S05), thereby accelerating the crucial cycle of error detection and correction. This aligns with theories of motor learning which emphasize the importance of timely and specific feedback for refining motor programs (Schmidt et al., 2018).
However, the students' feedback also wisely cautions against technological utopianism. While praising its corrective capabilities, they astutely identified the system's limitations, such as occasional inaccuracies and a "lack of emotional support" (S19). This highlights that while AI can effectively scaffold the "what" and "how" of motor performance, it may not yet fully replicate the motivational and affective dimensions of human coaching. Therefore, our findings suggest that the optimal role for such AI systems is not necessarily to replace human instructors, but to act as a powerful, scalable tool that supplements traditional instruction, providing the high-frequency, individualized practice feedback that is often impractical for a single instructor to deliver in a large class setting.
The mediating role of behavior: unpacking the primacy of practice volume
Perhaps the most significant contribution of this study is the elucidation of the causal mechanism driving skill acquisition in this AI-enhanced environment. Our mediation analysis revealed a nuanced and compelling story. While the AI intervention successfully elevated students' learning interest and self-directedness, these psychological states did not function as direct mediators for skill improvement. Instead, total learning duration emerged as the sole, powerful mediator, robustly explaining the link between the AI intervention and superior performance. This finding challenges a simplistic view that equates engagement with mere affective states and invites a deeper, theoretically-integrated interpretation.
We propose a multi-layered causal chain that aligns with our theoretical framework. The process begins at the cognitive level, as explained by CLT. The AI system, by providing immediate and precise Knowledge of Performance (KP), acted as a powerful cognitive scaffold. It drastically reduced the extraneous cognitive load associated with the frustrating process of self-diagnosis, which students described as feeling "unsure if I'm doing them correctly" (S02). This cognitive offloading made the practice sessions less mentally taxing and more efficient. It is this fundamental reduction in cognitive cost that served as the prerequisite for sustained engagement.
This cognitively streamlined experience, in turn, fueled the motivational engine described by SDT. By making error correction achievable and progress visible, the system directly nurtured the learners' fundamental need for competence. As one student noted, "Seeing my progress gives me a strong sense of achievement" (S23). The satisfaction of this need is a powerful driver of intrinsic motivation. While this heightened motivation—manifested as increased interest—did not directly translate into better skills, its crucial role was to convert positive feelings into tangible behavior: students simply chose to practice more. The AI system's success, therefore, lies not just in making learning more enjoyable, but in its ability to effectively channel that enjoyment into productive, effortful practice.
This brings us to the ultimate behavioral outcome, explained by theories of expertise development, such as deliberate practice (Ericsson et al., 1993). The increased volume of practice, which was guided and corrective thanks to the AI's KP, is the cornerstone of skill acquisition. Our findings thus paint a clear picture: the AI system initiated a virtuous cycle where cognitive scaffolding (via CLT) enabled motivational fulfillment (via SDT), which then drove behavioral investment (more practice time), ultimately leading to skill mastery.
Implications for theory and practice
From a theoretical perspective, this research offers several important contributions. First, it provides a nuanced, evidence-based model of how AI feedback influences motor skill acquisition, demonstrating a clear path from cognitive offloading to behavioral engagement. Second, it challenges a simplistic interpretation of SDT in novice learning contexts. Our findings suggest that for foundational skill acquisition, the satisfaction of the need for competence may be the most critical driver, and its primary function is to foster behavioral persistence (i.e., more practice) rather than directly enhancing learning through a purely psychological path. This highlights the importance of measuring behavioral outcomes, like time-on-task, as a primary indicator of engagement and technological efficacy.
From a practical standpoint, our findings provide clear guidance. The goal of such AI technology should not be merely to entertain or "gamify" the experience, but to strategically reduce cognitive barriers and motivate sustained, high-volume practice. Features should be designed with the explicit aim of increasing students' productive time-on-task. Furthermore, our results suggest the most promising future for OLPE lies in blended models. While AI can serve as an excellent "drill sergeant" for high-frequency, technical practice, it currently lacks the nuanced, affective support of a human coach. This frees human instructors to focus on higher-order coaching, holistic motivation, and the uniquely human aspects of teaching.
Limitations and future directions
Despite its valuable insights, this study has several limitations that offer avenues for future research.
First, the study's sample consisted solely of undergraduate students from a single university, which may limit the generalizability of the findings to other populations, such as younger K-12 students or older adults, who may interact with the technology differently. Furthermore, the task itself, Baduanjin, is a relatively slow-paced, form-based exercise. The effectiveness of this AI feedback approach for fast-paced, dynamic, or open-skilled sports (e.g., basketball, soccer) remains an important area for future investigation. Future studies should explore the efficacy of such systems across diverse age groups and cultural contexts.
Second, our 8-week intervention was sufficient to observe short-term skill acquisition, but it did not assess the long-term retention of these skills or their transfer to other related tasks. Longitudinal studies are needed to determine whether the skills learned via AI feedback are durable over time and how they compare to traditional methods in fostering adaptable motor expertise. Additionally, the short duration may not have fully captured the potential impact of a novelty effect, where initial engagement could be attributed to the technology's newness.
Third, while our mediation analysis identified learning duration as the sole significant mediator, this finding should be interpreted with caution due to the study's modest sample size (N = 60). Mediation analyses, particularly for parallel models with multiple mediators, require substantial statistical power to reliably detect smaller indirect effects. It is therefore plausible that the non-significant findings for the indirect paths through learning interest and self-directed learning could be a result of a Type II error, rather than a true absence of an effect. The potential mediating roles of these psychological constructs may be more subtle and require a larger sample to be statistically identified. Future research with larger cohorts is warranted to replicate our model and more definitively assess the relative contributions of behavioral and psychological mediators.
Finally, as noted in the qualitative findings, the technology itself has room for improvement. While effective, the system's accuracy and interactivity are not yet perfect. Future research should focus on advancing the underlying technology, such as improving algorithm precision for complex or occluded movements and enhancing pedagogical intelligence. Exploring the integration of adaptive learning models, which could adjust feedback complexity based on a learner's progress, or incorporating gamification elements beyond simple feedback, represents a particularly promising direction for developing the next generation of intelligent OLPE tools.
Conclusion
In conclusion, this study demonstrates that an AI-driven real-time feedback system is a highly effective tool for enhancing motor skill acquisition and student engagement in online physical education. The system's primary strength lies not merely in creating a more interesting learning experience, but in its capacity to translate that engagement into increased practice volume—the crucial ingredient for performance improvement. While the technology shows immense promise in addressing the long-standing challenges of remote skill instruction, its future lies in a symbiotic relationship with human pedagogy. The ultimate goal is to create blended learning ecosystems where technology provides scalable, data-driven feedback, allowing human instructors to focus on the holistic, motivational, and uniquely human aspects of coaching.
Acknowledgements
We sincerely thank Professor Xiaolei Liu for providing the instructional videos and images for this study. We also extend our gratitude to all the students and judges who participated in this experiment. Additionally, we appreciate the funding support from the Higher Education Teaching Reform Project under the Quality Engineering Initiative of the Guangdong Provincial Department of Education.
Author contributions
JM is responsible for manuscript writing and experimental design. LM is responsible for writing instructions. SQ, BZ and WR were responsible for the implementation of the teaching experiment and data collection. All authors read and approved the final manuscript.
Availability of data and materials
The datasets used and/or analysed during the current study are available from the corresponding author on reasonable request.
Declarations
Competing interests
No potential conflict of interest was reported by the author(s).
Abbreviations
Artificial Intelligence
Cognitive Load Theory
Intraclass Correlation Coefficient
Knowledge of Performance
Knowledge of Results
Limits of Agreement
Massive Open Online Course
Online Physical Education
Physical Education
Randomized Controlled Trial
Research Question
Self-Determination Theory
Publisher's Note
Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
References
Ardito, G; Czerkawski, B. The development of autonomous student learning networks: Patterns of interactions in an open world learning environment for teachers exploring teaching with and through computer science. Sustainability; 2021; 13,
Baker, S; Anderson, J; Burke, R. Equitable teaching for cultural and linguistic diversity: Exploring the possibilities for engaged pedagogy in post-COVID-19 higher education. Educational Review; 2022; 74,
Bazarevsky, V., Grishchenko, I., Raveendran, K., et al. (2020). Blazepose: On-device real-time body pose tracking. arXiv preprint arXiv:2006.10204. https://doi.org/10.48550/arXiv.2006.10204
Beserra, V; Nussbaum, M; Navarrete, M et al. Online physically active academic lessons in COVID-19 times: A pilot study. Teaching and Teacher Education; 2022; 116, [DOI: https://dx.doi.org/10.1016/j.tate.2022.103750] 103750.
Bowles, H; Clift, BC; Wiltshire, G. Joe wicks, lifestyle capitalism and the social construction of PE (with Joe). Sport, Education and Society; 2024; 29,
Calderón, A; Scanlon, D; MacPhail, A; Analy-Long, J. An integrated blended learning approach for physical education teacher education programmes: teacher educators’ and pre-service teachers’ experiences. Physical Education and Sport Pedagogy; 2021; 26,
Chan, WK; Leung, KI; Ho, CC; Qin, Z. Effectiveness of online teaching in physical education during covid-19 school closures: A survey study of frontline physical education teachers in Hong Kong. Journal of Physical Education & Sport; 2021; 21,
Chou, CM; Shen, TC; Hsiao, HC; Shen, TY. Influencing factors on students’ learning effectiveness of AI-based technology application: Mediation variable of the human-computer interaction experience. Education and Information Technologies; 2022; 27,
Daum, DN; Buschner, C. The status of high school online physical education in the United States. Journal of Teaching in Physical Education; 2012; 31,
Daum, DN. Thinking about hybrid or online learning in physical education? Start here!. Journal of Physical Education, Recreation & Dance; 2020; 91,
Eltahir, ME; Babiker, FME. The influence of artificial intelligence tools on student performance in e-learning environments: Case study. Electronic Journal of E-Learning; 2024; 22,
Ericsson, KA; Krampe, RT; Tesch-Römer, C. The role of deliberate practice in the acquisition of expert performance. Psychological Review; 1993; 100,
Garg, S; Saxena, A; Gupta, R. Yoga pose classification: A CNN and MediaPipe inspired deep learning approach for real-world application. Journal of Ambient Intelligence and Humanized Computing; 2023; 14,
Google Developers. (2023). MediaPipe. Retrieved December 1, 2023, from https://developers.google.com/mediapipe
Gu, H; Xie, C. The development of evaluation scale for college students' interest in physical education. Journal of Guangzhou Sport University; 2012; 32,
Ompoc, HMS; Aguinaldo, JC. Evaluating the effectiveness of AI-driven applications in enhancing the physical fitness of tertiary physical education students. International Journal of Human Movement and Sports Sciences; 2025; [DOI: https://dx.doi.org/10.13189/saj.2025.130109]
Hinojo-Lucena, FJ; Lopez-Belmonte, J; Fuentes-Cabrera, A et al. Academic effects of the use of flipped learning in physical education. International Journal of Environmental Research and Public Health; 2020; 17,
Hodges, N. J., & Franks, I. M. (2004). Instructions, demonstrations and the learning process: Creating and constraining movement options. In Skill Acquisition in Sport (pp. 169–198). Routledge.
Horita, SI; Gonzaga, BM; Mélo, RM; Rodrigues, AB; Ferreira, AT; Silva, AS; Horita, SIM; Gonzaga, BMS; Mélo, RMF; Rodrigues, ABM; Silva, ASM; Marcos, ACSB; Junior, GFS; Souza, HAS; Oliveira-Correia, JPS; Pereira, LA; Carvalho, MML; Brandão, NCC; Rosa, RSL; Ribeiro, TM; Souza, YRM; Alberti Gonçalves, M; Ferreira, RR; Riederer, I; Henriques Pons, A; Garzoni, LR; Araujo-Jorge, TC; Mello-Silva, CC. Leveling up engagement: Gamification’s role in creating immersive online events. Smart Learning Environments; 2025; 12,
Hsia, LH; Hwang, GJ. From reflective thinking to learning engagement awareness: A reflective thinking promoting approach to improve students’ dance performance, self-efficacy and task load in flipped learning. British Journal of Educational Technology; 2020; 51,
Hsia, LH; Hwang, GJ; Hwang, JP. AI-facilitated reflective practice in physical education: An auto-assessment and feedback approach. Interactive Learning Environments; 2024; 32,
Hu, Z; Liu, Z; Su, Y. AI-driven smart transformation in physical education: Current trends and future research directions. Applied Sciences; 2024; 14,
Huang, AYQ; Lu, OHT; Yang, SJH. Effects of artificial intelligence–enabled personalized recommendations on learners’ learning engagement, motivation, and outcomes in a flipped classroom. Computers & Education; 2023; 194, [DOI: https://dx.doi.org/10.1016/j.compedu.2022.104684] 104684.
Jeon, J; Lee, S. Large language models in education: A focus on the complementary relationship between human teachers and ChatGPT. Education and Information Technologies; 2023; 28,
Jeong, HC; So, WY. Difficulties of online physical education classes in middle and high school and an efficient operation plan to address them. International Journal of Environmental Research and Public Health; 2020; 17,
Killian, CM. Beyond wires and lights in a box: Online physical education is necessary to enhance program impact. Journal of Physical Education, Recreation & Dance; 2023; 94,
Kim, JW; Choi, JY; Ha, EJ; Park, H. Human pose estimation using mediapipe pose and optimization method based on a humanoid model. Applied Sciences; 2023; 13,
Lambert, K; Luguetti, C. Feminists against fad, fizz ed: A poetic commentary exploring the notion of Joe Wicks as physical education. Physical Education and Sport Pedagogy; 2022; 27,
Lambert, K; Hudson, C; Luguetti, C. What would bell hooks think of the remote teaching and learning in Physical Education during the COVID-19 pandemic? A critical review of the literature. Sport, Education and Society; 2024; 29,
Lee, KJ; Noh, B; An, KO. Impact of synchronous online physical education classes using Tabata training on adolescents during COVID-19: A randomized controlled study. International Journal of Environmental Research and Public Health; 2021; 18,
Li, F. Information teaching platform of college physical education based on artificial intelligence technology. Journal of Physics: Conference Series; 2021; 1852,
Lin, H. Influences of artificial intelligence in education on teaching effectiveness: The mediating effect of teachers’ perceptions of educational technology. International Journal of Emerging Technologies in Learning; 2022; 17,
Lin, KC; Hung, HC; Chen, NS. The effect of wearable technology on badminton learning performance: A multiple feedback WISER model in physical education. Smart Learning Environments; 2023; 10,
Lin, KC; Ko, CW; Hung, HC; Hwang, GJ. The effect of real-time pose recognition on badminton learning performance. Interactive Learning Environments; 2023; 31,
Lin, YN; Hsia, LH; Hwang, GJ. Fostering motor skills in physical education: A mobile technology-supported ICRA flipped learning model. Computers & Education; 2022; 177, [DOI: https://dx.doi.org/10.1016/j.compedu.2021.104380] 104380.
Lin, Y; Jiao, X; Zhao, L. Detection of 3D human posture based on improved mediapipe. Journal of Computer and Communications; 2023; 11,
Liu, D; Sun, X; Zhang, A. A031: Design and practice of AI-powered precision teaching model for public physical education courses in universities. International Journal of Physical Activity and Health; 2024; 2,
Liu, ZY; Yushchik, E. Exploring the prospects of using artificial intelligence in education. Cogent Education; 2024; 11,
López-Fernández, I; Burgueño, R; Gil-Espinosa, FJ. High school physical education teachers’ perceptions of blended learning one year after the onset of the COVID-19 pandemic. International Journal of Environmental Research and Public Health; 2021; 18,
Mahyoob, M. Online learning effectiveness during the COVID-19 pandemic: A case study of Saudi universities. International Journal of Information and Communication Technology Education; 2021; 17,
Örücü, S; Selek, M. Design and validation of rule-based expert system by using kinect V2 for real-time athlete support. Applied Sciences (Basel); 2020; 10,
Østerlie, O; Mehus, I. The impact of flipped learning on cognitive knowledge learning and intrinsic motivation in Norwegian secondary physical education. Education Sciences; 2020; 10,
Ovens, AP; Philpot, R; Bennett, B. Virtually learning: a self-study of evolving pedagogical practices. Movimento; 2022; 28, e28021. [DOI: https://dx.doi.org/10.22456/1982-8918.122595]
Owoc, M. L., Sawicka, A., & Weichbroth, P. (2019, August). Artificial intelligence technologies in education: Benefits, challenges and strategies of implementation. In IFIP international workshop on artificial intelligence for knowledge management (pp. 37–58). Cham: Springer. https://doi.org/10.1007/978-3-030-85001-2_4
Qi, Y. Computer-based physical education platform: Evaluating effectiveness during the COVID-19 pandemic. International Journal of Human-Computer Interaction; 2024; 40,
Ryan, RM; Deci, EL. Self-determination theory and the facilitation of intrinsic motivation, social development, and well-being. American Psychologist; 2000; 55,
Schmidt, R. A., Lee, T. D., Winstein, C., Wulf, G., & Zelaznik, H. N. (2018). Motor control and learning: A behavioral emphasis. Human Kinetics.
Suherman, WS. Physical education online learning during the Covid-19 pandemic: Effectiveness, motivation, and learning outcomes. The International Journal of Social Sciences World (TIJOSSW); 2021; 3,
Sweller, J. Cognitive load during problem solving: Effects on learning. Cognitive Science; 1988; 12,
Tegero, MC. Challenges encountered by physical education teachers in online teaching in times of Covid-19 pandemic. International Journal of Research Publications; 2021; 91,
Varea, V; Gonzalez-Calvo, G. Exploring the changes of physical education in the age of Covid-19. Physical Education and Sport Pedagogy; 2022; 27,
Varea, V; Riccetti, A; González-Calvo, G; Luguetti, C. Physical Education and COVID-19: What have we learned?. Curriculum Studies in Health and Physical Education; 2024; 15,
Vorlíček, M; Prycl, D; Heidler, J; Herrador-Colmenero, M; Nábělková, J; Mitáš, J; Hinckson, E; Mandic, S; Frömel, K. Gameful education: A study of Gamifiter application's role in promoting physical activity and active lifestyle. Smart Learning Environments; 2024; 11,
Wang, Y; Wang, X. Artificial intelligence in physical education: Comprehensive review and future teacher training strategies. Frontiers in Public Health; 2024; 12, 1484848. [DOI: https://dx.doi.org/10.3389/fpubh.2024.1484848]
Wu, B. (2010). An experimental study on the effect of self-directed learning styles on college students' physical education learning. East China Normal University. https://kns.cnki.net/kcms2/article/abstract?v=4OORb77KhuK3zW1ePSfLQdjE72FZ796FQ69ZWpEPSbkdr_shWI0GWYk0I6Z6GEubdP9baVvS1hNCHI9xejqMq6XPTd8Go3FIvFyrs1DlxTXgW4cayqKhMSqihaXwCEnRj7CZMfz520nMFu5Z7LucVq2K9uNib2MaPcmaS9DOGMAGps4CGLPPgoC70ofw81La&uniplatform=NZKPT&language=CHS
Wulf, G; Shea, C; Lewthwaite, R. Motor skill learning and performance: A review of influential factors. Medical Education; 2010; 44,
Xie, M. Design of a physical education training system based on an intelligent vision. Computer Applications in Engineering Education; 2021; 29,
Xu, W; Xing, QW; Zhu, JD; Liu, CY. Effectiveness of an extended-reality interactive learning system in a dance training course. Education and Information Technologies; 2023; 28,
Yang, ACM; Chen, IYL; Flanagan, B; Ogata, H. Automatic generation of cloze items for repeated testing to improve reading comprehension. Educational Technology & Society; 2021; 24,
Zheng, W; Ma, YY; Lin, HL. Research on blended learning in physical education during the COVID-19 pandemic: A case study of Chinese students. SAGE Open; 2021; 11,
Zheng, Y., Huang, D., Liu, S., et al. (2020). Cross-domain object detection through coarse-to-fine feature adaptation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 13766–13775). https://doi.org/10.48550/arXiv.2003.10275
© The Author(s) 2025. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.