PaperlayerView exact revision on arXiv

Source-linked AI summary

Skill Profiling with Attributable Reasoning (SPAR): A Wearable Analysis System for Boxing

Nibraas Khan, Hanchen David Wang, Enya Bullard, Ritam Ghosh, Ruj Haan, Aarav Agrawal, Meiyi Ma, Nilanjan Sarkar

arXiv:2609.30753v1cs.HCcs.LG

TL;DR

Wearable boxing systems mostly classify punch type, leaving field-based assessment of how well a punch is performed less addressed. SPAR combines full-chain sensing, expert/novice classification, and audience-specific explanations; on 4,713 punches from 17 participants, its leave-one-participant-out AUC is 0.842.

  • Problem

    Most wearable boxing systems classify which punch was thrown, while technique improvement calls for assessing how well each punch is executed and what to change.

    Source: §1 · p. 4

  • Method

    SPAR converts full-chain wearable signals into anatomical joint angles, classifies each punch as expert or novice, and provides attribution, counterfactual, and narrative explanations for different audiences.

    Source: §1 · p. 2, §5 · p. 17

  • Results

    0.842 leave-one-participant-out AUC on 4,713 punches from 17 participants.

    Source: §5 · p. 17

  • Takeaways & Limitations

    Six practicing coaches said they would use counterfactual explanations in class and narratives as take-home material, supporting differently formatted explanations for different audiences.

    Source: §5 · p. 17

    The classifier may perform poorly on an expert whose movement differs from the experts represented in its training data.

    Source: §4.3 · p. 16

Abstract

from arXiv · show

A punch is a ballistic, full-body action driven by a kinetic chain running from the legs through the trunk to the arm, where a small sequencing error separates a scoring strike from a miss. Wearable sensors can capture this movement in the gym, but most deployable systems only classify which punch was thrown rather than assess how well it was thrown. We present Skill Profiling with Attributable Reasoning (SPAR), an eight-IMU garment and pressure-insole system that classifies each punch as expert or novice and treats an explanation of that prediction as feedback. Feedback is only useful if the person receiving it can act on it, so SPAR explains the prediction at three tiers, a per-joint attribution for the analyst, a counterfactual over kinetic-chain layers for the coach, and a plain-language narrative of the two for the athlete. Across 17 participants and 4,713 punches, SPAR reaches a leave-one-participant-out AUC of 0.842 (95% CI [0.769, 0.907] over participants). A frozen time-series foundation model encodes the joint-angle and plantar-force series, and a small transformer trained on the cohort classifies the encoding. We audit the two quantitative tiers and report six themes from a thematic analysis of interviews with six practicing boxing coaches.

1 Introduction

Most wearable boxing systems classify punch type, while laboratory studies cannot capture technique under ordinary training conditions. SPAR extends sensing across the kinetic chain to assess each punch as expert or novice and explain the prediction for analysts, coaches, and athletes.

Movement measurement and deep learning

  • Prior wearable research spans laboratory-to-field movement measurement, activity recognition, and boxing punch classification, but commonly uses fewer sensors or stops at punch labels.Wearables avoid capture-volume constraints; deep-learning methods learn signal features, while self-supervised pretraining can reduce labeled-data needs.Source: §1 · pp. 2–3

Boxing biomechanics and explainable AI

  • Punch biomechanics links skilled technique to proximal-to-distal timing, while explainable-AI methods offer attribution, counterfactual, and language-based explanations for different audiences.Prior studies report that more skilled boxers drive legs and trunk before the arm; attribution alone identifies where a model looked but not what an athlete should change.Source: §1 · pp. 3–4

Boxing biomechanics and wearable sensing

  • Most boxing wearables classify punch type or report physical outcomes, while technique assessment still needs a cue about what to change.Laboratory studies measure punch mechanics with high fidelity but do not capture ordinary training, where faults can emerge under strain and fatigue.Source: §1 · pp. 2, 4

Positioning

  • SPAR uses an eight-IMU garment, pressure insoles, participant-calibrated anatomical joint angles, and an expert/novice classifier as a proxy for skill.The system instruments the upper- and lower-body kinetic chain and explains the movement difference behind its prediction.Source: §1 · p. 4
  • CAA attributes a prediction to joints and plantar zones, DCS searches for a corrective change across kinetic-chain layers, and GNG renders both outputs in plain language.The three tiers target analysts, coaches, and athletes, respectively.Source: §1 · p. 4
  • SPAR combines full-chain wearable sensing, per-punch expert/novice assessment, and three explanation tiers in a boxing-gym setting.The studies collect a gym dataset, evaluate classification with leave-one-participant-out cross-validation, audit two quantitative tiers, and assess all three tiers with practicing coaches.Source: §1 · pp. 4–5

2.1 Wearable System

The wearable system combines an eight-IMU garment, pressure insoles, and a browser-based collection application to capture boxing movement across the kinetic chain.

Wearable System

  • Eight garment-mounted IMUs observe the trunk, legs, shoulders, and elbows, with placement motivated by punch-force contributors and proximal-to-distal sequencing.Bone-close landmarks are used to minimize soft-tissue artifact.Source: §2.1 · p. 5
  • The IMUs stream orientations to a laptop at a nominal 100 Hz, but achieved rates vary between participants and occasional samples drop.The median inter-sample interval corresponds to 87 Hz, participant means range from 61 to 89 Hz, and the 99th-percentile interval is 21 ms; raw acceleration is not recorded.Source: §2.1 · p. 5
  • Pressure insoles stream three plantar-zone forces per foot at 100 Hz, and a wall-clock anchor aligns their data with the IMU stream.The system uses the six zone forces, not each foot’s reported total force.Source: §2.1 · p. 5
  • A browser application records IMU data, labels six canonical punch types, and displays a live 3D skeleton for session data-quality checks.The insoles record in parallel through the paired iPhone application.Source: §2.1 · p. 5

2.2 Study 1: Data Collection

Study 1 collected labeled boxing punches from experts and novices under a standardized session protocol, using calibrated wearables and a browser-based review application.

2.2.1 Participants and IRB

  • The cohort comprised 5 experts and 12 novices, stratified by competitive experience and training duration, with adult eligibility and recent-injury criteria.Experts had at least three years of competitive experience; novices had one year or less of training.Source: §2.2.1 · p. 5

2.2.2 Session Protocol

  • Participants completed sensor calibration and a five-second T-pose, then threw six punch types in a fixed order while the first author marked punch boundaries.The browser application displayed a live skeleton, sensor calibration, WebSocket status, and punch-type labeling controls.Source: §2.2.2 · pp. 5–6, Figure 2 · p. 6

2.2.3 Resulting Dataset

  • The resulting dataset contains 4,713 segmented punches from 17 participants, covering all six canonical punch types.Segments shorter than 100 ms were discarded as noise.Source: §2.2.3 · p. 6

2.3 Signal Processing and Skill Classifier

SPAR converts wearable signals into anatomical joint-angle and plantar-force series, encodes them, and trains a transformer to classify punches as expert or novice.

2.3 Signal Processing and Skill Classifier

  • The classifier solves sensor orientations into anatomical joint angles, encodes each joint-angle and plantar-force series separately, and classifies the encoded series.Separating the series through encoding supports later attribution to individual joints or groups.Source: §2.3 · p. 6

2.3.1 Preprocessing

  • Punches are resampled to a uniform 100 Hz grid, and their durations are retained as classifier input rather than normalized away.Orientations use RotationSpline interpolation; insole channels are interpolated and z-scored per participant.Source: §2.3.1 · pp. 6–7

2.3.2 Representation

  • The representation uses 22 anatomical coordinates from eight orientations, alongside six plantar-force series, for 28 separate time series per punch.MOMENT encodes each series with frozen weights; series are upsampled threefold, standardized using corpus-level scaling, and pooled into eight phase bins.Source: §2.3.2 · pp. 7–8
  • Table 1 reports participant identifiers, sex, height, total punch counts, and per-type counts with mean durations.The dataset breakdown is organized per participant.Source: Table 1 · p. 8

2.3.3 Classifier

  • A two-layer transformer classifies the 224 tokens, using series and phase tags, a duration token, and a punch-type embedding.The head has 0.8M trainable parameters while the 35M-parameter encoder remains frozen; five seeded heads are averaged.Source: §2.3.3 · p. 8, Figure 4 · p. 9

2.3.4 Training and Evaluation Protocol

  • Evaluation uses leave-one-participant-out folds and AUC because an unseen participant’s single-class fold cannot support threshold-based accuracy.Fold scores are converted to percentiles using nested cross-validation reference scores, and uncertainty is bootstrapped over participants.Source: §2.3.4 · pp. 9–10

2.4 Three-Tier Explanation Framework

SPAR provides explanations at analyst, coach, and athlete levels, using individual-unit attribution, kinetic-chain counterfactuals, and grounded plain-language narratives.

2.4 Three-Tier Explanation Framework

  • The three explanation tiers address different audiences: CAA attributes predictions to units, DCS searches kinetic-chain layers, and GNG expresses both in plain language.The layers are Foundation (hips and insoles), Engine (pelvis and lumbar spine), and Delivery (shoulders and elbows).Source: §2.4 · p. 10

2.4.1 Contrastive Ablation Attribution

  • CAA estimates each joint’s or plantar zone’s contribution by shuffling its tokens across the held-out participant’s punches and re-scoring them.Five random pairings are averaged, and absolute attribution values are reported by class across folds.Source: §2.4.1 · p. 10

2.4.2 Directed Counterfactual Search

  • DCS replaces one layer of a novice punch with tokens from its closest matched training expert, then identifies the layer producing the largest expert-probability rise.The match is chosen by distance on series outside the candidate layer; the selected layer is averaged across the participant’s punches.Source: §2.4.2 · p. 10

2.4.3 Grounded Narrative Generation

  • GNG gives a language model ranked CAA units and DCS layer rises, and checks that its stated primary gap matches the payload.The prompt restricts the narrative to payload-supported claims and prohibits unsupported cues or drills.Source: §2.4.3 · pp. 10–11

2.5 Study 2: Coach Evaluation

Study 2 used semi-structured interviews with six boxing coaches to examine how they interpret and might use SPAR’s explanations.

Study 2: Coach Evaluation

  • Six coaches from Nashville-area boxing and combat-sport gyms, with 4 to 20 years of experience, took part in individual 25-to-30-minute interviews.They reviewed recorded punches alongside the system’s three outputs and discussed interpretation, agreement, and possible gym use.Source: §2.5 · p. 11
  • The interviews asked whether coaches would identify the same issue as SPAR and how they would use the system in their gyms.Participants included group-class coaches and coaches working in technique analysis with competitive athletes.Source: §2.5 · p. 11
  • One analyst used reflexive thematic analysis, producing 36 initial codes grouped into six themes and checked against transcripts over repeated passes..Source: §2.5 · p. 11

3 Results

Results report LOPO skill-classification performance, participant-level variation, and ablations of encoder-side choices.

3.1 Skill Classifier

  • LOPO AUC is 0.842 ± 0.097 (bootstrap 95% CI [0.769, 0.907] over participants), with expert and novice folds averaging 0.869 and 0.830.The classifier's mean is over participants and five seeds.Source: §3.1 · p. 11
  • Thirteen of 17 participants exceed 0.70 AUC and seven exceed 0.90, while the three lowest are E1 at 0.503, N1 at 0.595, and N10 at 0.668.The participant-level results vary despite similar mean AUCs for expert and novice folds.Source: §3.1 · p. 11

3.1.1 Ablations

  • In LOPO evaluation, keeping MOMENT’s per-series normalization reduced AUC to 0.523 ± 0.048; native-rate encoding rather than 3× upsampling yielded 0.724 ± 0.050, while phase-bin mean-and-standard-deviation features with the same head yielded 0.736 ± 0.021.With normalization, the embedding carries only trajectory shape, not posture or range of motion.Source: §3.1.1 · p. 11

3.2 Explanation-Quality Audit

The LOPO audit evaluates joint and plantar-zone attributions alongside counterfactual substitutions across kinetic-chain layers. The tiers align in aggregate but often differ for individual novices.

3.2.2 Counterfactual

  • Replacing any one layer with a matched expert’s raises mean expert probability by about the same amount: +0.274 for Foundation, +0.272 for Delivery, and +0.271 for Engine.The primary gap varies across novices: Delivery for five, Engine for four, and Foundation for three.Source: §3.2.2 · p. 12, Table 4 · p. 14

3.2.3 Agreement

  • The pelvis and hips have the largest novice-fold attributions, while Foundation and Engine are the primary gap for seven of 12 novices.For individual novices, CAA’s top-ranked unit often differs from DCS’s primary-gap layer because the tiers answer different questions.Source: §3.2.3 · pp. 12–13, Table 4 · p. 14

3.2.1 Attribution

  • On novice folds, pelvis and left-hip attributions lead at 0.0207 and 0.0191; expert-fold attributions are smaller, led by the left elbow at 0.0099.Neighboring-joint rankings are unresolved given fold-to-fold spread, and plantar-zone attributions are an order of magnitude smaller than joint attributions.Source: §3.2.1 · pp. 11–12, Table 3 · p. 13, Figure 6 · p. 14

3.3 User Study

Thematic analysis of six coaches’ interviews produced 36 initial codes and six themes.

3.3.1 Theme 1: DCS lets a coach reach more athletes

  • All six coaches said they would use DCS most because its single highest-leverage fix lets them cue more athletes in classes of 15 to 30.They described limited close observation time, rather than lack of coaching knowledge, as the constraint.Source: §3.3.1 · p. 14

3.3.2 Theme 2: What fits depends on the coach

  • Coaches’ roles shaped their preferences: group-class coaches rejected CAA for class use, while technique analysts valued it for reviewing punches objectively.All coaches favored DCS in class and GNG as take-home material.Source: §3.3.2 · pp. 14–15

3.3.3 Theme 3: One voice does not fit every audience

  • Four coaches liked GNG’s slightly formal voice for take-home material, while two wanted a different, blunter voice for face-to-face coaching.Source: §3.3.3 · p. 15

3.3.4 Theme 4: SPAR can train new coaches

  • Four coaches with trainer-certification duties saw SPAR as a potential coach-development tool because its narrative gives new trainers words for technique patterns.They described repeated exposure as helping trainers develop their own vocabulary.Source: §3.3.4 · p. 15

3.3.5 Theme 5: Families as part of youth coaching

  • Both youth-program coaches proposed GNG for involving parents, but said children should not see CAA displays of their own bodies.They envisioned CAA being used by the coach and parent in youth settings.Source: §3.3.5 · p. 15

3.3.6 Theme 6: Traceability earns trust

  • Two coaches said traceability supported trust because the narrative derives from named joints and the region whose correction most raised the prediction.Source: §3.3.6 · p. 15

3.3.7 Other feedback

  • Coaches identified deployment gaps including five-minute setup, laptop-only use, and requests for directional CAA, exportable GNG, adjustable voice, and finer advanced-athlete cues.The coaches wanted setup under a minute and use on a phone or tablet.Source: §3.3.7 · p. 15

4 Discussion

The discussion argues that explanations should be tailored to the people acting on a prediction, while recognizing that audiences and their needs vary across settings. It also qualifies SPAR’s findings by the cohort, binary skill categories, punch-labeling approach, and stance represented in the study.

4.1 A Three-Tier Explanation Framework for Multiple Audiences

  • Study 2 supports giving different audiences distinct explanation tiers that trace back to the same model output.The authors argue that one explanation is unlikely to serve everyone who must act on that output.Source: §4.1 · pp. 15–16

4.2 A Wider Spectrum of Explanations

  • Coaches’ varied roles and requests for different narrative voices and audiences motivate setting the explanation audience at use time.The proposed direction is a more flexible framework than the three fixed tiers designed for this study.Source: §4.2 · p. 16
  • The narrative tier could reason directly over punch encodings, but that would sacrifice the grounding guarantee valued by coaches unless its reasoning were made explicit.The current narrative instead receives a short text payload containing a primary gap and three joint names.Source: §4.2 · p. 16

4.3 An Expert Outside the Training Distribution

  • E1’s fold AUC was 0.503, and its mean token similarity to the other experts was 0.80, unlike the other experts’ 0.91–0.95 range.The authors attribute this outlier to the small expert cohort: the other four experts were more similar to one another than to E1.Source: §4.3 · p. 16

4.4 Beyond Boxing

  • SPAR’s audience-oriented explanation framework is presented as applicable beyond boxing, while organizing existing xAI methods by whom they serve rather than by technique.The authors cite clinical and rehabilitation settings as examples where different people rely on the same prediction.Source: §4.4 · p. 16

4.5.1 Cohort size

  • The 17-person cohort is small for cross-subject IMU work, and each expert-held-out fold learns the expert pattern from only four people.Participant-bootstrap intervals span roughly ±0.07, and the authors note that more experts would address the atypical E1 case.Source: §4.3 · p. 16, §4.5.1 · p. 17

4.5.2 Binary skill label

  • The expert/novice categories separate boxers with three or more competitive years from those with one year or less, leaving intermediate experience unrepresented.Because skill is a continuum, two classes cannot place a boxer between groups or distinguish boxers within a group.Source: §4.5.2 · p. 17

4.5.3 Punch-level labels

  • Punches inherit the thrower’s expert or novice label, so an individual’s poor and strong punches receive the same class label.The recordings have no punch-level outcome measure, and coaches may disagree about how to label a punch.Source: §4.5.3 · p. 17

5 Conclusions

SPAR extends wearable boxing analysis from identifying punch type to assessing expert or novice technique, using full-chain sensing and a learned classifier. On the collected data, it reaches a LOPO AUC of 0.842, and six coaches’ responses support presenting explanations differently to different audiences.

5 Conclusions

  • 0.842 LOPO AUC was achieved on 4,713 punches from 17 participants, extending wearable boxing systems beyond punch-type reporting to expert/novice assessment.The cohort included five experts with at least three years of competitive boxing and 12 novices with one year or less.Source: §5 · p. 17
  • Six practicing coaches said they would use the counterfactual in class and the narrative as take-home material, and identified audiences the system had not designed for.The authors conclude that explanations need different formats and presentations for different audiences.Source: §5 · p. 17
  • The authors state that the explanation framework is not tied to boxing or wearable sensing and applies to systems communicating predictions to different kinds of people.Each person acting on a prediction needs an explanation they can act on.Source: §5 · p. 17
Loading 2609.30753v1…