Automated Real-Time Detection and Correction of Children’s Kinesthetic Learning Using Expert-User Performance and Smartphones as Wearables

Abstract

More than 80% of young people (11–17 years) do not meet recommended levels of physical activity, while excessive sedentary smartphone use increases rapidly, highlighting the need for accessible tools that promote active and kinesthetic learning. This study investigates whether smartphones can function as wearable devices capable of tracking movement, detecting biomechanical errors, and providing real-time corrective feedback. Using a user-centered design approach, we developed a gamified Exertion Trainer in which children practiced a straight punch (boxing jab) while wearing a smartphone on their wrist. Embedded accelerometer data were processed on board to deliver immediate, task-specific feedback on arm orientation, using gravity as a fixed reference frame. A randomized crossover trial was conducted with 40 children, comparing a feedback condition with a no-feedback control across two test orders. Quantitative results showed that real-time feedback produced a statistically significant improvement in punch accuracy (p < 0.001) and reduced performance variability, with the strongest effects observed after initial practice and partial retention following feedback removal. Qualitative findings indicated higher engagement and stronger perceptions of kinesthetic learning when feedback was available. These results demonstrate that smartphones can serve as practical wearable devices for delivering biomechanical guidance and supporting movement skill acquisition in children.

1. Introduction

Many children worldwide are overweight and experience delays in motor development because they do not meet daily physical activity recommendations [1,2,3,4,5,6]. These problems intensified during the COVID-19 pandemic, when opportunities for physical activity decreased while sedentary screen use increased sharply [7,8,9,10,11,12,13]. Even after the pandemic, children and adolescents continue to engage in sedentary behaviors and excessive screen exposure [14,15]. Together, these trends underscore the need to repurpose mobile technologies: from predominantly sedentary media devices to tools that actively promote physical interaction and support kinesthetic learning.

A persistent challenge is that automated systems capable of real-time error detection and correction remain prohibitively expensive and often require specialized hardware or technical expertise [16,17]. In contrast, human coaches provide immediate feedback, detecting errors and correcting them as they occur; a capability that automated systems still struggle to offer in accessible and scalable ways [18,19,20]. Wearable and smartphone-based systems offer a potential solution due to their ubiquity, low cost, portability, and integration of high-quality inertial sensors that capture fine-grained movement data [21,22,23].

To our knowledge, few reported applications in the literature or on the market intentionally configure smartphones to function as body-worn devices that both sense (track) movement and provide real-time feedback (correct) to children during their kinesthetic learning. Although smartphones are already pervasive, low-cost, and equipped with high-quality inertial sensors, they remain underexplored as body-worn devices for motion tracking and real-time corrective feedback in children’s motor skill development. These characteristics position smartphones as uniquely scalable wearable platforms, particularly when compared to camera-based or multi-device systems that require additional infrastructure [24,25].

We hypothesize that smartphones can be configured as body-worn sensing devices to track movements, compare them against expert baselines, and deliver real-time corrective feedback. Building on this hypothesis, and following a user-centered design approach, we implemented and tested the Exertion Trainer app, which expanded the purpose of mobile devices beyond conventional cognitive learning apps into the domain of physical education and motor skill acquisition. The app provides a proof-of-concept in which children practice a straight punch (the jab in boxing) while wearing a smartphone on their wrist, allowing the device’s embedded accelerometer to capture motion data and deliver immediate kinesthetic feedback.

Accordingly, this paper presents experimental evidence that smartphones, when used as wearable devices, can support automated, real-time detection and correction of movement errors during children’s motor skill learning. We developed the Exertion Trainer app to support learning of the straight punch technique when using a smartphone as a wearable device attached to the children’s wrist, which senses their movements and provides real-time kinesthetic feedback. We evaluated the Exertion Trainer in a user study involving forty children aged 8 to 16 years.

The study investigates the effect of real-time smartphone-based feedback on children’s biomechanical performance using a controlled crossover design. The findings indicate that feedback enhances posture attainment, engagement, and short-term retention, with timing moderating performance gains.

Our contribution lies in validating smartphones as accessible wearable platforms for real-time motion error correction and kinesthetic learning, rather than in developing the Exertion Trainer app itself. This positions everyday mobile devices as stepping stones toward extensible systems that can integrate smartwatches, multi-point wearables, and emerging sensing technologies such as smart fabrics.

This paper is organized as follows. Section 2, Related Work, provides a detailed review of existing mobile and wearable devices for human movement tracking, human movement error-correction systems, and smartphone applications for physical activity and technique training. It situates the present study within the broader wearable-technology landscape and identifies the research gap addressed by this study. Section 3, Methods and Materials, describes the design, theoretical foundations, and technical implementation of the Exertion Trainer, as well as the user-study methodology and evaluation measures. Section 4, Results, presents the quantitative and qualitative findings. Section 5, Discussion, interprets the results, discusses limitations, and practical implications. Finally, Section 6, Conclusion, summarizes the main contributions and outlines directions for future work.

2. Related Work

Smartphones and smartwatches are widely used to capture human movement in research and clinical contexts [26,27]. In 2008, early commercial applications, such as Runkeeper and Strava [28,29], introduced activity tracking based on the Global Positioning System (GPS), followed by broader health platforms like Google Fit and Apple Health in 2014 [30,31]. Foundational work has demonstrated that built-in inertial sensors support human-activity recognition [32,33], and subsequent studies have summarized smartphone sensor capabilities for movement monitoring [21] and validated smartwatch measurements [34]. More recent systems combine smartphone inertial data with computer vision and machine learning for sport-specific analysis, as seen in HomeCourt (basketball) and SwingVision (tennis) [35,36,37].

2.1. Mobile Devices as Wearables

Smartphones are the most used body-worn devices, and their accuracy in recognizing human activities is influenced by device placement and orientation [21]. Their accelerometer data have been shown to align closely with gold-standard motion-capture systems [16,24]. Research has investigated the placement of smartphones in pockets, backpacks, and jackets [38,39,40]. Some studies have enhanced tracking by pairing phones with additional Inertial Measurement Units (IMUs) [41,42]. Multi-sensor approaches further expand capabilities, such as integrating phones, smartwatches, and accelerometers for gait recognition [43] or using a smartphone-equipped vest for rehabilitation [17].

2.2. Human-Movement Error-Correction Systems

Wearable feedback systems provide either post hoc or real-time error correction, with mixed effectiveness outside controlled settings [44]. Haptic cues generally support motor learning more effectively than video review [45,46], and accurate body tracking is essential to avoid user frustration [47]. Post hoc analysis of smartphones has been used in swimming technique [25] and balance assessment [48]; real-time applications include navigation for blind users [49,50], sedentary-behavior reduction [51], and rehabilitation [17].

2.3. Smartphone Applications for Physical Activity and Technique Training

In 2025, many commercial apps promote activity but do not use the smartphone’s onboard sensors for corrective feedback. Apps such as GoNoodle, Sworkit Kids, Cosmic Kids Yoga, Super Stretch Yoga, and Moovlee rely on instructional videos rather than analyzing users’ movement execution. Pokémon GO encourages walking behavior but does not assess movement technique [52]. High-end tools like SwingVision and HomeCourt use smartphones’ cameras and computer vision for kinematic feedback, require specific environmental conditions (e.g., camera positioning), and are sport-specific. Services such as Nike Training Club provide basic form guidance, but do not perform real-time sensor-based biomechanical assessment. Apps including Fitness Boxing (Nintendo Switch), Strava, Runkeeper, MapMyRun, and Nike Run Club primarily track quantitative metrics (pace, distance) but not movement technique. Similarly, popular home-workout apps (Freeletics, Seven, 30-Day Fitness, Everlast Gym) either operate without sensor-based feedback or use additional proprietary punch trackers (accelerometers), like FightCamp, to count the user’s punches and the speed. This reveals a gap or an opportunity: few solutions leverage only the smartphone’s built-in sensors and actuators to provide detailed, kinematic-aware corrective feedback without external hardware or camera-based setups, especially for complex movements such as boxing.

2.4. Gap and Contribution

Existing work establishes the feasibility of smartphone-based movement tracking and feedback, yet real-time error detection and correction for children’s motor learning remains underexplored. This gap or opportunity motivates our research, which demonstrates that a single smartphone worn on the wrist can provide accurate, real-time corrective feedback grounded in expert baselines.

3. Materials and Methods

This section describes the design, implementation, and evaluation of the Exertion Trainer, a proof-of-concept mobile application that transforms a commercially available smartphone (without any additional custom hardware) into a wrist-worn biomechanical feedback system. The system supports children’s kinesthetic learning of a straight punch (a jab in boxing), a movement selected for its clearly defined biomechanical phases and suitability as an introductory motor skill [46,53,54].

Our methodology integrates biomechanical control theory, the expert-performance model, and user-centered design principles [55]. We evaluated the intervention through a controlled, randomized crossover experiment with 40 participants aged 8–16.

3.1. System Design and Theoretical Foundations

3.1.1. Biomechanical Control Theory

The Exertion Trainer operationalizes a biomechanical control model in which expert-defined trajectories serve as reference states for real-time comparison with user motion. Within this biomechanical control framework, the expert benchmark functions as the desired trajectory, and user performance is continuously compared to it through a user-in-the-loop system.

This system, the Exertion Trainer, is powered by a standard Xiaomi Redmi Note 11 smartphone, worn on the wrist via adjustable elastic holders (Figure 1). The smartphone’s built-in Inertia Measuring Unit (IMU) captures triaxial acceleration at 10 samples per second (10 hertz) during practice, fulfilling Nyquist requirements while reducing data volume and power consumption [56,57,58,59,60,61]. The data from this IMU serves as the foundation for control, enabling the app to infer wrist orientation relative to expert-defined acceleration patterns, approximate punch speed, and detect specific biomechanical states. Based on this real-time analysis, each punch is classified on the smartphone as correct or incorrect against predefined, expert-derived criteria, providing immediate feedback to users. All motion data is stored locally for post hoc analysis. The Exertion Trainer’s biomechanical sensitivity and real-time analysis were validated through prior comparisons with glove-mounted accelerometers, post hoc video methods (e.g., [46,53]), and a standalone MPU6050 sensor, demonstrating highly similar acceleration curves and sufficient fidelity for straight-punch analysis [62].

Crucially, the feedback provided to the user is designed for effective kinesthetic learning. The feedback loop is grounded in Fitts and Posner’s [63] cognitive stage of motor learning and in boxing pedagogy [64]. To reduce cognitive load and maintain instructional clarity, the system issues only one correction at a time. These prompts are positively phrased, delivered via text-to-speech, and rate-limited to one every two seconds [62,65].

For sustained postures, a movement is considered achieved only when the user enters the correct biomechanical range and maintains it for six continuous seconds. Upon completion, the app provides audible countdowns and time-to-target feedback. Ultimately, as illustrated in Figure 2, this closed-loop calibration–tracking–analysis–feedback cycle applies biomechanical control theory to a practical, real-world wearable system designed for children.

3.1.2. User-Centered Design Framework (ISO 9241-210)

We develop the Exertion Trainer guided by a User-Centered Design framework, adhering to the principles and four-stage process (understanding context, specifying requirements, developing solutions, and evaluation) outlined in ISO 9241-210 [55]. This commitment to the user-centered design ensured that the interaction design was consistently shaped by core principles, including suitability for the task, self-descriptiveness, learnability, and conformity with user expectations [55].

A key aspect of this approach was the active involvement of children aged 8–16 throughout the design process. Their feedback directly contributed to critical usability refinements, clearer language, and the integration of engaging, lightweight game-like elements. To further align with the goals of accessibility and real-world impact, the smartphone platform was strategically selected for its ubiquity, affordability, and potential for widespread deployment [66].

3.1.3. Instructional and Interface Design

We grounded the instructional design of the Exertion Trainer in pedagogical guidelines for novice boxers [64] and principles for exertion game design [67]. These foundations directly shaped the sequencing of instruction, the demonstration of the straight punch, and the presentation of feedback. Furthermore, the sequence of body motions was designed to promote kinesthetic learning while deliberately minimizing fatigue, in line with the recommendations reported by Elmanowski et al. [68].

This pedagogical approach is brought to life through a graphical user interface that we built around three core elements designed to support instruction and feedback, sustain engagement, and promote skill acquisition.

First, to ensure accessibility and reinforce comprehension, all instructions and feedback are presented in Spanish, the language of the children who participated in the experiment, in both written and spoken formats. Second, the interface is designed as an interactive, game-based experience tailored to children, a strategy aimed at fostering sustained engagement and an emotional connection, aligning with Ferguson et al. [66]. The app’s journey, outlined in Figure 3, seamlessly guides the user from initial engagement (animated logo, player registration) through setup (calibration) and into deliberate practice with instructional videos.

The third and central element supports the cognitive stage of motor learning [63] by focusing on deliberate practice [69,70]. The interface includes an animated demonstration of the target movement (the straight punch) based on standardized expert biomechanics and validated by amateur boxers. As shown in Figure 4, this animation allows users to mentally and visually grasp the movement before physical execution, creating a clear link between the model and the child’s own replication of the task.

The system’s real-time feedback is crucial for translating this visual understanding into physical performance. The Exertion Trainer app provides specific guidance on achieving biomechanical states, with success defined by the user’s ability to enter and maintain a correct position for a continuous six-second count. Figure 5 illustrates this kinesthetic learning process, contrasting scenarios where a user immediately finds the correct stance, takes time to achieve it, or is unable to maintain it.

Collectively, these three interface elements (accessible language, an engaging game-like shell, and a core of animated deliberate practice) scaffold the cognitive phase of motor learning by ensuring a clear understanding of the movement before its physical execution.

3.2. Technical Implementation

The proposed approach comprises two primary design solution domains: (1) learning from the expert performer and (2) training the user. As a proof of concept, we implemented this approach by coding two complementary applications: one for expert performers to establish movement benchmarks and classify errors, and another for children to receive real-time feedback during training. Both applications utilized biomechanical data captured from a smartphone worn on the wrist during punching motions, with the training application additionally incorporating gamified elements and feedback mechanisms. Sensor data were stored locally on the device to ensure data integrity and minimize transmission delays.

3.2.1. Creating Reference Data: Collecting Expert-Performance

We applied the expert-performance methodology [69,70] to establish the biomechanical benchmarks for the Exertion Trainer. The initial phase of this process involved determining the optimal tracking configuration, which entailed testing multiple body placements of the smartphone as a wearable device on the expert user, including the wrist (palmar and dorsal sides) and the arm above the elbow. Each configuration was evaluated during punch execution, alternating between both right and left hands, and switching foot stances (e.g., left-hand punches with the left foot forward, then the same punching hand with the right foot forward, and vice versa).

Following identification of the dorsal wrist as the optimal location, we conducted 30-minute sessions with expert boxers to establish precise biomechanical benchmarks. During these sessions, experts performed shadowboxing and bag drills while wearing the smartphone on the wrist. Beyond data collection, their expertise was also integral to refining the instructional system itself. The expert performers reviewed the graphical interface and instructional videos our team created, helping define what constitutes a “correct position” and providing recommendations for coaching cues to address common technique errors in learning the straight punch.

Working with the expert user to determine the optimal smartphone placement was a key step in developing the Exertion Trainer app, directly informing its final implementation as a wrist-worn wearable device.

All biomechanical benchmarks were derived from this single wrist-worn smartphone configuration. Although alternative sensing systems (e.g., multi-sensor arrays or camera-based tracking) were explored during early prototyping, they were not adopted due to increased hardware demands and reduced scalability [62]. Accordingly, expert-defined reference trajectories were established using only the embedded inertial sensors of a commercially available smartphone, consistent with the study’s objective of validating smartphones as standalone wearable coaching devices.

3.2.2. Error Classification and Feedback Model

The Exertion Trainer app evaluated two phases of the punch: On Guard and On Extension. For each phase, five common errors were identified and paired with discipline-specific corrective cues as depicted in

Table 1. Error classification for the “On Guard” stance using triaxial accelerometer data.

Baseline *	Error	Error	Error	Error	Error
On Guard	Hand down	Guard too extended	Guard over-closed	Chicken wing	Inverted elbow

* The normative stance is illustrated at left. The five subsequent panels depict common errors, each paired with its corresponding discipline-specific feedback for correction.

and

Table 2. Error classification for the “On Extension” stance using triaxial accelerometer data.

Baseline *	Error	Error	Error	Error	Error
On Extension	Flared fist	Overshot punch	Underthrown punch	Forearm pronation	Forearm supination

* The normative stance is illustrated at left. The five subsequent panels depict common errors, each paired with its corresponding discipline-specific feedback for correction.

, respectively. These formed a hierarchical correction sequence, prioritizing gross deviations before finer adjustments.

Error classifications are defined relative to the expert-established biomechanical model of the straight punch and do not imply that similar movement patterns are universally incorrect across other punches, styles, or sports disciplines.

The system relies exclusively on wrist-mounted triaxial accelerometer data. While additional sensors could expand biomechanical coverage (e.g., force estimation or full-body kinematics), the present implementation was intentionally constrained to a single embedded sensor to evaluate whether real-time error detection can be achieved without external instrumentation.

3.2.3. Accelerometer Function

The IMU records triaxial acceleration along the X, Y, and Z axes. Figure 6 illustrates both the accelerometer axis configuration and the hierarchical correction process. Preliminary testing confirmed that the embedded accelerometer accurately captures wrist kinematics relevant to the straight punch, supporting the results mentioned in Section 3.1.1 [16,24].

3.3. User Study

3.3.1. Participants and Recruitment

Forty female and male participants (ages 8–16) were recruited via digital flyers distributed to local schools, sports centers, and community organizations in Cuernavaca, Mexico. Eligibility required possession of the right upper limb, ability to follow audio instructions, absence of injury, and medical clearance for physical activity.

The Research and Academic Ethics Committee of the School of Engineering at the Universidad Nacional Autónoma de México (UNAM) approved the user-study protocol. Written parental consent was obtained in accordance with the informed-consent procedures, and children provided their assent before participating.

3.3.2. Experimental Design and Procedure

We used a randomized crossover design as summarized in

Table 3. Experimental Design Procedure.

Number of Users	Group (Test Order)	Condition (Feedback or Not)	Stance 1	Stance 2	Repetitions
20	Group 1: Control test first	Control (No Feedback)	On Guard	On an extended punch	×3
		Experimental (Feedback)
20	Group 2: Feedback test first	Experimental (Feedback)
		Control (No Feedback)

. We assigned participant children to one of two groups (20 participants per group); each group interacted with two versions of the Exertion Trainer app (one with no feedback and the other with real-time feedback). Participants were divided into two groups to account for first-interaction effects. Therefore, all participants experienced a control version of the Exertion Trainer app (without real-time feedback) and an experimental version (with real-time feedback).

The only difference between Groups 1 and 2 was the order in which they used the control and experimental versions of the app:

• Group 1: Control (no feedback) → Experimental (feedback)
• Group 2: Experimental (feedback) → Control (no feedback)

The first step in participation was to sign the informed consent and assent forms. Then, participants completed an anthropometric and sports background questionnaire. Then, the calibration of the Exertion Trainer app was done in two stages: (1) device alignment with the phone placed flat, and (2) user alignment through three standard positions (“On Attention,” “On Guard,” and “On Extension”) as shown in Figure 7. The application required correct device placement and enforced a flat-arm calibration stage before allowing participants to proceed, thereby minimizing reference-frame misalignment. These established the initial orientation vectors used throughout the session. The three standard positions were used to define individualized orientation baselines and were not evaluated for biomechanical correctness. Because both experimental conditions relied on the same calibrated reference frame within each participant, any residual alignment variability would have affected both conditions equally. Then, each child completed both versions of the Exertion Trainer app in a single 30-minute session.

The instructional sequence, stance order (“On Guard” and “On Extension”), and timing were identical on both the control and experimental versions of the Exertion Trainer app. Participants were required to follow the app-guided instructions exactly; no free-form or improvised movements were permitted during data collection.

The wrist-worn smartphone’s weight (178 grams) is comparable to standard youth boxing gloves (6–12 ounces; approximately 170–340 grams) and did not restrict movement or cause discomfort during sessions. Each participant was observed by both a parent or legal guardian and a researcher. The physical setup of the user study is illustrated in Figure 8.

3.3.3. Quantitative Measures

The Exertion Trainer application continuously logged and evaluated triaxial accelerometer data (X, Y, Z) at a rate of 10 samples per second. Participant performance was quantified using two complementary outcome measures derived from these signals:

• Posture Achievement Score: The count of six target postures that the participant successfully reached and intentionally stabilized for six consecutive seconds, reflecting discrete task-level motor attainment.
• Time-in-Posture: The total duration for which the child’s limb was within the correct biomechanical range, regardless of the child holding such posture or only passing by. This measure reflects the efficiency with which the correct posture is found and maintained.

To ensure numerical precision and consistency on a smartphone, we converted the raw hexadecimal accelerometer readings into a fixed-point format by multiplying by 1000 and rounding, following standard fixed-point quantization used in embedded inertial sensing [71]. This approach preserved the sensor’s sub-g resolution and standardized heterogeneous integer representations (16-bit and 32-bit outputs). While fixed-point arithmetic can reduce computational effort in real-time systems, in the present study, the conversion was implemented primarily for numerical standardization rather than for measurable CPU or battery optimization.

Each trial automatically generated a log file labeled with the participant’s name and test type (control or feedback). To keep the application lean and suitable for public download, only the essential variables were stored: username, trial phase (instruction or feedback), timestamp, and the X, Y, and Z accelerometer values. Although real-time calculations provided immediate feedback, the derived metrics were not saved because they could be reproduced during post hoc analysis. All log files were extracted from the smartphone after the trials for statistical processing and analysis.

The smartphone’s inertial measurement unit measures both linear acceleration (motion) and constant acceleration due to gravity (approximately 9.81 m/s²). When a smartphone is at rest with the screen facing upward, the Z-axis registers an acceleration of approximately +9.81 m/s²; when the smartphone is upside down, it registers an acceleration of approximately −9.81 m/s². This gravitational component is key for determining device (and thus limb) orientation during straight-punch execution, without explicitly separating gravity from dynamic acceleration, consistent with established posture-estimation approaches [72,73].

Based on these accelerometer signals and the expert-defined biomechanical model described in Section 3.2, stance-specific thresholds were implemented to classify correct postures versus common errors during the On Guard and On Extension phases. The corresponding acceleration value boundaries are presented in

Table 4. X, Y, and Z accelerometer values for the On Guard stance.

Order Within the Algorithm	Accelerometer Axis	What is Being Measured	Under the Correct Range (m/s2)	Correct Range (m/s2)	Above the Correct Range (m/s2)
2	X	The in and out inclination (Coronal plane) of the wrist provides the position of the fist in relation to the head.	X < −3 Guard too extended	X ≥ −3 & X ≤ 5 Correct range	X > 5 Guard over closed
1	Y	Fist up	Y < 0 Fist parallel to the floor or hand down	Y ≥ 0 Correct range	Y ≥ 0 Correct range
3	Z	The inward or outward inclination (Sagittal plane) of the wrist provides the position of the elbow in relation to the ribs.	Z < 0 Inverted elbow	Z ≥ 0 & Z ≤ 5 Correct range	Z > 5 Chicken wing

Note: All acceleration values are expressed in meters per second squared (m/s²).

and

Table 5. X, Y, and Z accelerometer values for the On Extension stance.

Order Within the Algorithm	Accelerometer Axis	What is Being Measured	Under the Correct Range (m/s2)	Correct Range (m/s2)	Above the Correct Range (m/s2)
3	X	Wrist orientation	X < −5 Flared fist	X ≥ −5 & X ≤ 5 Correct range	X > 5 Over-rotated
2	Y	Height of the thrown punch	Y < 0 Underthrown punch	Y ≥ 0 & Y ≤ 2.5 Correct range	Y > 2.5 Overshot punch
1	Z	The straightness of the arm must be forward	Z < 0 Dropped arm	Z ≥ 0 Correct range	Z ≥ 0 Correct range

Note: All acceleration values are expressed in meters per second squared (m/s²).

, respectively.

3.3.4. Qualitative Measures

After each app version, participants completed a 5-point Likert-scale questionnaire based on Laugwitz et al.’s [74] usability framework, assessing: Motivation, Ease of use, Excitement, Friendliness, and Interest. Also, at the end of the session, participants selected their preferred version (Feedback or Control).

3.3.5. Statistical Analysis

Analyses included paired t-tests for within-group comparisons, two-way mixed ANOVA for order and condition effects, and Welch’s ANOVA to detect heterogeneity of variance. Nonparametric Wilcoxon and Mann–Whitney U tests were used to assess robustness to non-normality in some subsets. Data visualizations included bar plots, box plots, and raincloud distributions.

4. Results

This section comprehensively evaluates the Exertion Trainer app across two dimensions: quantitative biomechanical performance and qualitative user satisfaction. Based on 80 user trials involving 40 participants divided into two groups, we examine performance under control and feedback conditions, assess motor improvement retention, and compare subjective experience with objective motion data. Each of the 80 tests yielded an average of 1082 biomechanical samples, resulting in a robust dataset of 86,560 sensor-based data points obtained from the participant children.

For the quantitative evaluation, we utilized the user files stored on the smartphone during the trials. For the qualitative assessment, we entered the questionnaire responses into a custom-designed form. We programmed and executed all descriptive statistics, inferential tests, dataset processing steps, and created every graph and table presented in this section using Python 3.9.6. This Python-based workflow allowed us to systematically compare performance within participants (control versus feedback), between groups (Group 1 versus Group 2), and across test order (Group 1 Control as the first test versus Group 2 Control as the second test, and likewise for the Feedback test).

4.1. Quantitative Evaluation

The Posture Achievement Score was treated as the primary outcome measure because it captures discrete, task-completion motor success aligned with instructional goals. Time-in-Posture was analyzed as a secondary outcome to characterize movement efficiency, execution consistency, and learning dynamics beyond binary posture attainment. These measures operationalize expert-defined boxing technique benchmarks established through the expert-performance methodology described in Section 3.2.1.

4.1.1. Within-Group Evaluation

Posture Achievement Score (Within-Group)

Each participant attempted six target postures (three On Guard and three On Extension). A position was considered fully achieved only if it was held within the target range for at least six seconds (60 samples). This six-second criterion was established based on empirical pilot testing with children in the target age range (8–16 years). The duration was selected to balance measurement rigor with practical feasibility, distinguishing intentional stabilization from brief or incidental alignment while avoiding unnecessary fatigue during repeated trials. Positions that were reached but not maintained for the whole interval were not counted as achieved.

The range of achieved target postures revealed a clear order effect across conditions (

Table 6. Descriptive statistics of achieved target postures.

Test Order	Condition	N	Mean	SD	Min	Max	95% CI
Group 1: Control then Feedback	Control (No-feedback)	20	1.30	1.84	0	5	[0.44, 2.16]
	Experimental (Feedback)	20	4.50	1.64	0	6	[3.73, 5.27]
Group 2: Feedback then Control	Experimental (Feedback)	20	4.75	1.33	3	6	[4.13, 5.37]
	Control (No-feedback)	20	3.60	2.26	0	6	[2.54, 4.66]

Note: N = sample size; SD = standard deviation; Min = minimum value; Max = maximum value; 95% CI = 95% confidence interval of the mean.

). When averaged across test-order groups, the Feedback condition more than doubled the mean number of fully achieved postures, increasing from approximately 2.45 in the Control condition to 4.63 with feedback. In the Control-then-Feedback group, performance in the Control condition was highly variable (range: 0–5), whereas the subsequent Feedback condition showed more consistent attainment (range: 0–6). In contrast, in the Feedback-then-Control group, performance in the Feedback condition was already constrained to higher values (range: 3–6), while variability increased when feedback was removed (range: 0–6). This pattern suggests that early exposure to feedback establishes a performance floor that partially persists after feedback withdrawal, whereas beginning without feedback leads to greater variability throughout the session.

As shown in

Table 7. Paired Samples t-test by Test Order Group for achieved target postures.

Group	t	df	p	Mean Difference	Cohen’s d	95% CI of Difference
Group 1: Control then Feedback	5.49	19	<0.001	3.20	1.23	[1.98, 4.42]
Group 2: Feedback then Control	3.09	19	0.006	1.15	0.69	[0.37, 1.93]

Note: t = t-statistic; df = degrees of freedom; p = two-tailed probability value; Cohen’s d = standardized effect size; CI = confidence interval.

, within-group comparisons revealed that both experimental sequences produced significant advantages for the Feedback condition over the Control condition, though with notable differences in effect size. Group 1 (Control then Feedback) exhibited a substantial effect (d = 1.23) with highly significant results (p < 0.001), while Group 2 (Feedback then Control) showed a moderate effect (d = 0.69) with slightly less statistical significance (p = 0.006). This pattern indicates that the sequence of condition presentation substantially moderates the observed benefit of feedback, with the transition from no-feedback to feedback conditions producing more dramatic performance enhancements than the reverse sequence.

Time-in-Posture (Seconds): Descriptive Outcomes

Twenty-eight participants performed better in the Feedback app than in the Control app test. As illustrated in Figure 9, Group 1, which started with the Control condition, consisted of 12 participants who were unable to fully achieve any target position in the Control app test. Only one participant performed equally in both conditions, and two performed better in the Control app test. Notably, 19 participants achieved at least 2 of the 6 target positions with the Feedback app.

Figure 10 depicts results from Group 2, which began with the Feedback app. Six participants performed equally in both tests, while two improved in their subsequent Control app test. Only two participants failed to achieve any position in the Control app test. In contrast, all 20 participants completed at least 3 of the 6 target positions using the Feedback app.

Both Figure 9 and Figure 10 highlight that the Feedback condition consistently yielded a higher number of fully achieved positions with reduced dispersion across participants, illustrating both improved accuracy and greater execution stability.

Time-in-Posture (Seconds): Statistical Analysis (Within-Group)

This metric captures the time a participant spent correctly aligned in a target posture, regardless of continuity. Every 0.1-s interval where the user’s body was within the target range counted as one sample, providing a granular measure of their overall efficiency and ability to find and correct their posture throughout each 20-s trial.

Descriptive statistics for both test order groups are presented in

Table 8. Descriptive Statistics by Test Order Group for all the postures within the range.

Test Order	Condition	N	Mean	SD	SEM	Min	Max	Median	95% CI	CV (%)	V
Group 1: Control then Feedback	Control (No-feedback)	20	13.46	11.47	2.57	0.1	39.5	12.30	[8.09, 18.82]	85.25%	131.58
	Experimental (Feedback)	20	34.64	9.77	2.18	4.3	50.4	36.35	[30.07, 39.21]	28.20%	95.43
Group 2: Feedback then Control	Experimental (Feedback)	20	34.52	6.00	3.27	25.2	48.1	35.80	[31.71, 37.33]	17.38%	36.00
	Control (No-feedback)	20	29.55	14.64	1.34	3.2	45.9	34.95	[22.70, 36.40]	49.53%	214.24

Note: N = sample size; SD = standard deviation; SEM = standard error of the mean; Min = minimum value; Max = maximum value; 95% CI = 95% confidence interval of the mean; CV (%) = coefficient of variation expressed as a percentage; V = variance.

, indicating that the Feedback app tests yielded higher and more consistent performance compared to the Control app tests. Specifically, Group 2’s Feedback app test exhibited the lowest variability (SD = 6.00, CV = 17.38%), as indicated by standard deviation (SD) and coefficient of variation (CV). In contrast, Group 1’s Control app test showed the lowest mean and the highest variability (Mean = 13.46, CV = 85.25%), highlighting inconsistencies when feedback was not initially provided.

However, when analyzing how many samples each participant achieved in the requested posture, we evaluated whether participants differed when the test provided feedback or not, and whether there was a significant difference in the order the tests were applied.

The order in which participants received the feedback significantly influenced its effectiveness. All tests were two-tailed. When feedback followed a control task (Group 1), participants showed significantly better performance with feedback (t (19) = 5.72, p < 0.001), with a large and statistically significant performance jump of 21.19 points (95% CI [13.44, 28.93]) and a large effect size (Cohen’s d = 1.28). In contrast, when feedback was given first (Group 2), its impact was minimal and not statistically significant (t (19) = 1.57, p = 0.132), showing a 4.97-point improvement (95% CI [−1.64, 11.58]) and a small effect size (Cohen’s d = 0.35). The detailed statistical results, including paired-samples t-tests, presented in

Table 9. Pair T-Test Results by Test Order Group for all the postures within the range.

Test Order	t	df	p	Mean Control (No-Feedback)	Mean Feedback	Mean Difference	Cohen’s d	95% CI of Difference
Group 1: Control then Feedback	5.72	19	<0.001	13.46	34.64	21.19	1.28	[13.44, 28.93]
Group 2: Feedback then Control	1.57	19	0.132	29.55	34.52	4.97	0.35	[−1.64, 11.58]

Note: t = t-statistic; df = degrees of freedom; p = two-tailed probability value; Cohen’s d = standardized effect size; CI = confidence interval.

show that the sequence of administration was a key factor in determining the intervention’s effectiveness.

4.1.2. Between-Group Evaluation

Posture Achievement Score (Between-Group)

We compared the test-order effect, specifically the number of participants in each group who achieved each of the six postures in each condition. As shown in

Table 10. Between-Groups Simple Main Effects of Test Order Within Each Condition on the number of fully achieved positions.

Condition	t	df	p	Mean Group 1	Mean Group 2	Mean Difference	Cohen’s d	95% CI of Difference
Control (No feedback)	−3.534	38	0.001	1.30	3.60	−2.30	−1.117	[−3.618, −0.982]
Experimental (Feedback)	−0.529	38	0.600	4.50	4.75	−0.25	−0.167	[−1.206, 0.706]

Note: t = t-statistic; df = degrees of freedom; p = two-tailed probability value; Cohen’s d = standardized effect size; CI = confidence interval.

, the test order significantly influenced performance only in the Control (no-feedback) condition (p = 0.001), with Group 2 achieving 0.98 to 3.62 more positions than Group 1 (95% CI [−3.618, −0.982]). The presence of feedback eliminated test-order effects (p = 0.6), with the data consistent with either group performing better by up to 1.2 positions (95% CI [−1.206, 0.706]).

Time-in-Posture (Seconds): Statistical Analysis (Between-Group)

We employed a comprehensive analytical approach to examine the effects of feedback condition and test order on performance. Our sample consisted of 40 participants, equally divided between two testing sequences: Group 1 (Control then Feedback) and Group 2 (Feedback then Control), with each participant completing both conditions in a within-subjects design.

Before conducting our primary analysis, we thoroughly checked our assumptions to ensure their validity. Shapiro–Wilk tests indicated normality violations in three of four datasets (Group 1 Control: W = 0.900, p = 0.042; Group 1 Feedback: W = 0.870, p = 0.012; Group 2 Control: W = 0.903, p = 0.047), though the overall model residuals were normally distributed (W = 0.980, p = 0.240). More critically, Levene’s test revealed significant heterogeneity of variances between groups (W = 8.27, p = 0.005).

To address these violations, we implemented a dual approach: (1) Welch’s ANOVA for between-subjects comparisons, which is robust to unequal variances; and (2) Mann–Whitney U and Wilcoxon signed-rank tests as nonparametric alternatives to confirm findings. The Welch’s ANOVA confirmed significant between-group differences in baseline performance (F (1, 72.50) = 7.21, p = 0.009, η²_p = 0.085). Our two-way mixed ANOVA revealed three key effects, as shown in

Table 11. Two-Way Mixed ANOVA.

Source	SS	df1	df2	MS	F	p	η2p
Test_order	1276.00	1	38	1276.00	10.62	0.0024	0.22
Condition	3420.42	1	38	3420.42	28.88	0.0000	0.43
Interaction	1314.63	1	38	1314.63	11.10	0.0019	0.23

Note: SS = sum of squares; df₁ = numerator degrees of freedom; df₂ = denominator degrees of freedom; MS = mean square; F = F-statistic; p = two-tailed probability value; η²_p = partial eta squared effect size.

:

The analysis underscored: (1) a strong main effect of condition (F = 28.88, p < 0.001, η²_p = 0.43), indicating feedback substantially improved performance; (2) a significant test order effect (22% variance explained), revealing sequence-dependent performance differences; and (3) a crucial Test Order–Condition interaction (η²_p = 0.23), showing that feedback’s effectiveness depended on testing order.

Nonparametric analyses corroborated these findings. Wilcoxon tests showed significant within-subjects improvement only when control preceded feedback (Control_First: W = 12.5, p < 0.001; Feedback_First: W = 76.5, p = 0.294), while Mann–Whitney U confirmed baseline group differences (U = 569.5, p = 0.027).

The convergent evidence from parametric and robust analyses indicates that while feedback consistently enhances performance, its benefit is modulated by testing sequence. Participants who began without feedback showed dramatic improvement after receiving feedback, whereas those who started with feedback maintained relatively stable performance across conditions. This pattern suggests that initial feedback exposure creates carryover effects that benefit subsequent unaided performance.

Together with the visualization in Figure 11, it compares how many tenths of a second each of the six target positions was within range for each group and condition. Group 1’s Control app test had the lowest performance, while their Feedback app test showed the highest. Group 2’s Feedback results were comparable to Group 1’s, but their Control results were also considerably better. These results suggest that while beginning without feedback is disadvantageous, prior experience with feedback can improve subsequent unaided performance. Most importantly, the feedback condition itself consistently produced superior results, confirming its fundamental effectiveness.

4.2. Visualization of Order Effects

Figure 12 and Figure 13 analyze the total duration (in tenths of a second) during which users maintained a correct posture within range for each instruction. These plots count all correct entries in the target position, regardless of whether the users maintained the correct posture for 6 s.

In Figure 12, Group 1 participants (who completed the Control condition first) show improved performance in their second test with the Feedback condition. The paired plot reveals upward trends, while the box plot illustrates reduced variability. The distribution plot shows that the Control data is moderately skewed toward lower values, reflecting participants’ initial exposure without corrective guidance, while the Feedback data is more centered and unimodal.

In Figure 13, Group 2 participants (who completed the Feedback condition first) also performed better in the Feedback condition, though the difference is less pronounced. The Control condition shows greater variability, but its distribution is less skewed toward lower values than in Group 1, and the distribution plot highlights overlapping curves, with the Feedback distribution being taller and more peaked. This pattern suggests that prior exposure to feedback influenced subsequent unaided performance.

4.3. Qualitative Evaluation

4.3.1. Likert Questionnaire

Figure 14 presents the results of the five-point Likert-scale questionnaire administered after each test. Group 2’s Control condition (administered after Feedback) received higher Motivation, Excitement, and Interest scores than Group 1’s initial Control condition.

4.3.2. Preferred App

A clear consensus emerged when children chose their preferred app, as shown in Figure 15. The Feedback version was the overwhelming favorite, selected by 17 out of 20 children in each group. Combined, this means that 34 out of 40 young participants (85%) chose the Feedback app as their preferred training method.

5. Discussion

We evaluated the Exertion Trainer app as a proof of concept for a broader approach: a real-time human motion correction system using smart mobile devices as wearable technology. Our results demonstrate the system’s effectiveness in delivering biomechanical feedback that improves motor performance in children learning a complex task. The app facilitates skill acquisition by integrating principles of design, motor learning, and interaction, even in short, single-session trials. In the discussion below, we address the theoretical foundations, key findings, and broader implications of this work.

5.1. Guiding Principles

We grounded the development of the Exertion Trainer in ISO 9241-210:2019, emphasizing the iterative involvement of target users and domain experts [55]. Children, coaches, and amateur boxers actively contributed to each design stage, supported by a multidisciplinary research team. These collaborations shaped sensor placement, instructional flow, and user interface design decisions. To evaluate the technology’s affordances and understand users’ behavior while designing with the body, as framed by Höök [75], both expert performers and children wore the smartphone on their wrists during the iterative development process. This embodied approach allowed our team to capture localized biomechanical data and refine the system based on users’ bodily engagement with the technology. We derived the biomechanical benchmarks from standardized movements validated by trained athletes, consistent with the expert-performance methodology [69].

We adhered to the seven principles of ISO 9241-110:2020 [76] to design the interaction. These were theoretical guidelines, substantiated by Likert-scale results and in situ observations. Children demonstrated autonomous navigation, a clear understanding of prompts, and minimal need for clarification. The interface was learnable, controllable, goal-oriented, and resilient to input errors under real-world conditions, including low lighting and environmental noise.

We emphasized the cognitive stage, where correct technique must be understood, thus shaping the Exertion Trainer app’s instructional content and feedback based on Fitts & Posner’s [63] motor learning theory. The system delivered real-time feedback on one correction at a time to avoid cognitive overload. These choices aligned with England Boxing’s emphasis on clarity, progressive correction, and motivational phrasing [64]. Our observations confirmed that participants responded to corrections intuitively, without needing to be told which limb or task the feedback referred to, highlighting the value of smartphones as wearables in reducing ambiguity.

User experience data reinforced these findings. Children engaged easily with the app, and 85% preferred the feedback version. These outcomes align with Mueller et al.’s [67] principles for exertion games, where engagement arises from embodied interaction and actionable feedback rather than complex game mechanics, supporting easy entry into play, achievable short-term goals, skill-matched challenges, actionable real-time feedback, and social acceptability.

The interplay of qualitative and quantitative evaluations supports our design philosophy: feedback must be timely, localized, and limited to a single actionable correction. This minimizes confusion and enhances the relevance of each prompt. The app demonstrated how Control Theory can inform movement coaching by treating the child as a dynamic system within a closed feedback loop, guiding learners through consistent, real-time adjustments and reinforcement.

5.2. Quantitative Findings

The quantitative results demonstrate a clear and robust benefit of real-time feedback on children’s biomechanical performance. Across all analyses, the Feedback condition consistently outperformed the Control condition. A two-way mixed ANOVA revealed a strong main effect of condition, with feedback accounting for a substantial proportion of the variance in performance (η²_p = 0.43). This finding indicates that real-time feedback produces a meaningful and generalizable improvement in motor performance.

Beyond increasing mean performance, feedback also stabilized movement execution. In the absence of feedback, performance was highly variable, with coefficients of variation reaching up to 85%. In contrast, feedback-enabled trials reduced variability to as low as 17–28%. This reduction suggests that feedback supported more consistent and repeatable movement patterns, guiding participants toward biomechanically valid solutions rather than isolated or chance successes.

The effectiveness of feedback was strongly influenced by test order. When participants began without feedback, performance in the Control condition was lower and more variable than when the Control condition followed a feedback-enabled trial. Within-group analyses showed that 12 participants in Group 1, who were initially unable to fully achieve any target, exhibited the greatest performance gains once feedback was introduced. In contrast, participants in Group 2, who received feedback initially, maintained relatively high performance during the subsequent Control condition. This pattern indicates that feedback is most effective when introduced after an initial period of unguided exploration, and that early guided practice supports short-term retention when feedback is removed.

Between-group comparisons, together with a significant interaction between condition and test order, further supported these order effects. When the Control test was performed first, the performance gap between Control and Feedback was larger (Group 1). When the Control test followed Feedback, this gap was reduced (Group 2), with Control performance exceeding that of participants who had not yet received feedback. This pattern indicates that early exposure to feedback supports later unaided performance.

Finally, feedback substantially improved task success. When averaged across test-order groups, the Feedback condition more than doubled the mean number of fully achieved postures compared to the Control condition. This result indicates that feedback not only improves overall performance but also increases the likelihood of correctly reaching and holding complex target positions, a direct measure of technical skill acquisition.

5.3. Qualitative Findings

Likert-scale data do not directly reflect the improvements in quantitative performance, but they offer complementary insights into user perception. Interestingly, the Control app test of Group 2 (conducted after exposure to the Feedback app) received consistently higher ratings than Group 1’s Control app test across most categories, suggesting that initial feedback may enhance later task understanding and confidence. Group 2’s Control condition even approached the Feedback condition of the same group in terms of perceived ease and interest, indicating possible retention of instructional clarity. In contrast, Group 1’s Control app test (performed without prior feedback) received the lowest ratings overall, except for ease. These patterns reinforce the idea that feedback improves movement performance and positively influences perceived usability and engagement, even when it is no longer present.

Preferred App results further validated the approach. Most (85%) children across both groups preferred the Feedback app.

5.4. Design Implications for Smartphone-Based Kinesthetic Feedback

The findings suggest several design implications for smartphone-based kinesthetic learning systems. First, effective feedback does not require complex external wearables; commodity smartphones, when carefully positioned and calibrated, can provide sufficient sensing fidelity to support meaningful motor learning. Second, feedback mechanisms should prioritize reducing execution variability rather than solely optimizing peak performance, as consistency appears to be a key marker of skill acquisition. Third, brief, task-specific feedback delivered in real time can scaffold learning without overwhelming the user, making such systems particularly suitable for children and novice learners. Together, these implications support the design of accessible, scalable exertion games and learning tools that leverage existing mobile technologies to promote active, embodied learning in everyday contexts.

5.5. From Proof of Concept to General Approach

Our findings establish a replicable approach for designing mobile-based systems that correct human motion. The architecture includes a wrist-worn smartphone, movement benchmarks derived from expert data, real-time IMU-driven motion analysis, and a correction algorithm that prioritizes a single prompt per cycle. Our approach is adaptable to other sports or rehabilitative tasks by modifying movement goals and feedback language. While our focus was on skill acquisition in healthy children, similar task-specific training approaches have been explored in rehabilitation. For example, Agrawal et al. [77] showed how robot-enhanced mobility training can engage children with cerebral palsy through goal-directed interaction. Elmanowski et al. [68] employed haptic feedback for arm rehabilitation in stroke patients. Although these systems use different feedback forms, they share with our approach the emphasis on task precision and localized interaction.

Our Exertion Trainer demonstrated robustness across environmental conditions (e.g., low light) and demographic contexts, underscoring its suitability for broader applications, especially in resource-constrained settings.

5.6. Limitations

This evaluation captured only immediate improvements and short-term retention; longitudinal research is needed to assess sustained learning. The study focused specifically on wrist orientation and punch trajectory, which are directly measurable from a wrist-worn sensor; full-body kinematics were not assessed. Because orientation is inferred from wrist acceleration patterns rather than from a complete inertial fusion model, the system is optimized for straight punch detection and may require additional sensing for more complex full-body movements. The relatively small, region-specific sample limits generalizability, even though statistical significance was achieved. Using a single smartphone model controls hardware variability, but it may not accurately reflect performance across devices with differing sensor capabilities.

Despite these constraints, the results validate the Exertion Trainer as an effective and accessible tool for providing real-time biomechanical feedback. This work lays the groundwork for future systems integrating learning theory, control theory, and ubiquitous computing to support embodied skill development.

5.7. Practical Implications and Cost Accessibility

Compared to professional optical motion-capture systems such as Vicon, which typically require specialized laboratory environments and represent capital investments conservatively estimated in the tens of thousands of U.S. dollars (with entry-level configurations commonly reported above US$15,000). Depth-sensing systems such as Microsoft Kinect also require dedicated external hardware and fixed spatial setups, limiting their portability and scalability.

In contrast, modern smartphones equipped with embedded inertial sensors are widely available at consumer-level price points in the low hundreds of U.S. dollars. The Exertion Trainer leverages hardware users already own: their smartphones. The software can be distributed at a cost of US$5 per download and operates solely through the device’s integrated sensors, without requiring laboratory space, external cameras, or additional instrumentation. It can therefore be deployed in everyday environments such as homes, schools, or parks.

This difference in infrastructure and cost structure, tens of thousands of dollars for laboratory-grade motion capture versus a low-cost downloadable application operating on existing consumer hardware, substantially lowers barriers to scalability and positions smartphone-based biomechanical feedback as a cost-accessible alternative for large-scale educational and community settings.

6. Conclusions

This research confirms that an application, developed for a smartphone used as a wearable, can accurately track and provide real-time corrective feedback on children’s punching biomechanics. The validated Exertion Trainer approach demonstrates the feasibility of using smart mobile devices as wearable technologies for delivering task-specific biomechanical guidance. The controlled crossover evaluation demonstrated that feedback-enabled interaction supports superior motor performance and posture attainment, particularly when feedback follows an initial phase without feedback. Participants also exhibited partial retention after feedback was removed. Importantly, these gains were accompanied by a marked reduction in performance variability, indicating that feedback promoted more stable and consistent movement patterns, rather than merely increasing momentary accuracy. Together, these results confirm that this approach effectively supports motor-skill acquisition even in short sessions with minimal supervision or instructional scaffolding. Qualitative findings further indicate higher engagement and stronger perceptions of kinesthetic learning when feedback was available.

Future work will focus on longitudinal studies, expanding to other sports, and integrating AI-driven adaptive feedback to accommodate varying user abilities. Our results suggest that by transforming widely available mobile devices into personal coaching tools, children can develop correct biomechanics in an engaging, cost-effective manner, potentially reshaping physical education globally.