acceptodds
Under review as a conference paper at ICLR 2027

Explainable Badminton Coaching via Knowledge-guided Multimodal Anomaly Detection

Abstract

Badminton strokes require highly coordinated movements across body parts, spanning foot loading, trunk rotation, and racket-arm muscle control. This makes coaching difficult with the naked eye and videos alone, as most of these signals are invisible. Recent badminton AI coaches have begun to incorporate wearable sensors. However, without a biomechanically organized modeling process, the errors they flag cannot be verifiably tied to the body part and stroke phase where they occur. Prior works also either stop at stroke- and skill-level classification, or issue coaching feedback that is neither explainable nor traceable to evidence. To address these limitations, we propose (), a multimodal framework for explainable badminton-stroke coaching from wearable sensors. We treat coaching as anomaly detection, with a conditional latent diffusion model operating over structured representations of skeletal motion, muscle activity, plantar pressure, and eye gaze, organized by a body-part-sensor graph that links biomechanically related signals. Detected deviations are then verified and transformed into coaching errors via a two-level knowledge graph and a reasoning multimodal large language model. Experiments on backhand drive and forehand clear strokes from the MultiSenseBadminton dataset show that BIC outperforms existing methods in both skill-level discrimination and anomaly detection. In blind evaluations, human badminton experts also rate BIC as more comprehensive and professional.

Primary area:
learning on time series and dynamical systems
Keywords:
Time-Series Anomaly Detection, Multimodal Learning, Latent Diffusion Models, Knowledge Graphs, AI Coach, Sports Analytics, Wearable Sensors

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_NZt2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

The paper proposes BIC, a three-stage pipeline for explainable badminton-stroke coaching from multimodal wearable sensors (skeleton IMUs, EMG, plantar pressure, and eye gaze). Experiments on backhand drive and forehand clear strokes from MultiSenseBadminton cover three evaluations: skill-level discrimination (expert vs. beginner AUROC), phase-level anomaly localization, and an expert-adjudicated assessment of coaching findings and report quality. BIC reports large improvements over TSAD baselines, ChainBMD, and a generalist MLLM. Overall, this paper is well written but also raises some concerns.

Critical strengths

Framing coaching as “detect deviations first, then judge” is intuitive. It aligns better with how human coaches give feedback than the classify-then-attribute pipelines used in prior work. Structuring cross-modal fusion around a biomechanical body-part–sensor graph is a sensible way to inject domain priors, and the ablations support it: graph versus flattened skeleton, and modality-specific versus global denoiser. The pipeline is complete and well engineered. It connects low-level signal anomalies to rule-verified, human-readable coaching output, and grounding findings in deterministic rule checks is a reasonable attempt to limit LLM hallucination. The paper includes a human expert study with per-finding adjudication rather than only aggregate Likert ratings, and the appendix gives extensive implementation details, including full prompts.

Critical weaknesses
  1. The introduction frames AI coaching as a scalable alternative to human coaches reviewing every repetition. However, the proposed system relies on a lab-grade setup: 21 IMUs, three EMG sites, bilateral pressure insoles, and an eye tracker. This is far from deployable in routine training, so the scalability argument does not hold. The paper also stresses that the decisive signals (muscle activation, plantar pressure) are invisible to the naked eye. Yet the ground truth for coaching correctness comes from experts judging footage, skeletal video, and reports (Sec. 4.3, App. K.2). It is unclear how experts could reliably adjudicate findings such as elevated forearm activation or heel-weighted loading. If they cannot, the evaluation does not test the signals the paper claims are essential. I feel like the authors should elaborate more on the "motivation" part or the logic of the research gap.

  2. The differentiation from ChainBMD (modeling the non-racket side, avoiding template outputs) is an engineering difference rather than a conceptual gap. Methodologically, the pipeline combines established components (ST-GCN, TCN, latent DDPM, an LLM-drafted KG, and an MLLM). The paper does not offer an insight that transfers beyond this application. The title and abstract emphasize explainable, traceable coaching, but no experiment measures traceability. The Level-2 causal and correlational edges are drafted by an LLM. Their correctness is never validated, for example, by expert agreement on the edges or on the generated explanation chains.

  3. The prompt in App. F.2 describes it as a corroboration mechanism that raises or lowers confidence. Sec. 3.4, however, states that traversing it triggers no additional evaluation. Please clarify which is the case.

  4. Each fold trains on two experts and tests one expert against eight beginners. The model may learn subject identity (anthropometrics, sensor placement) rather than skill. Strokes from the same subject are not independent, and no subject-level results or statistical tests are given. Excluding intermediate players because of class overlap simplifies the task. The results only show separation of the two extremes, not skill grading.

  5. Sampling rate, rationale for the 3+3-frame window, number of diffusion samples used for scoring, score variance, and the phase segmentation procedure are not given.

Questions for authors
  1. λ3 appears twice in L_total, could you briefly explain the reason in the main text?

  2. How did experts verify findings that depend on EMG or plantar pressure using footage, skeletal video, and reports? What fraction of BIC’s findings involve these modalities, and how were they rated?

  3. What are the results for a baseline of MLLM + matched expert reference + TSAD hints, without the KG and rules?

  4. Can you report subject-level AUROC with statistical tests, and show that the model is not separating subjects by identity?

  5. Did the baselines use the same windowing, normalization, and pretrained encoders as BIC? What explains the below-chance scores?

  6. Could you please explain the 1/(|M_sens|−1) normalization in the main text?

10 comments
AuthorsOpenReview ↗

Response to Reviewer NZt2 (Part 1)

Thank you for the detailed assessment and for recognizing the motivation for detecting deviations before assigning coaching interpretations. We address the scope, empirical grounding, and reproducibility questions below.

W1(a, c): Motivation, scalability, and research gap.

We appreciate the reviewer's concern regarding the practical deployment of a system requiring multiple wearable sensors. We acknowledge that the current sensing setup is resource-intensive and limits its immediate applicability to routine training.

However, our primary objective is not to develop a consumer-ready sensing system, but to investigate how heterogeneous biomechanical measurements can be automatically interpreted to provide accurate, explainable coaching feedback.

This sensing setup is motivated by the nature of badminton execution. Video and skeletal motion reveal joint kinematics, but cannot directly measure muscle activation, plantar-pressure distribution, or gaze behavior. These signals provide complementary evidence about muscle recruitment, weight transfer, and visual attention, which are relevant to coordinated stroke execution. Consequently, errors involving these measurements may remain undetected through visual inspection alone, limiting the specificity of corrective feedback.

The central research gap is therefore not simply the lack of sensing modalities, but the absence of a systematic approach to localize multimodal execution deviations, verify their biomechanical significance, and translate them into actionable coaching errors. BIC addresses this through biomechanically structured anomaly detection and knowledge-guided error verification. We agree that reducing sensor requirements is important for practical deployment. The present work establishes the modeling and diagnostic framework using the available synchronized multimodal measurements.

We have also revised the Introduction to explicitly acknowledge the deployment limitations imposed by specialized wearable sensors.

W2(a): Methodological contribution beyond ChainBMD.

We appreciate the reviewer's observation that BIC builds upon established neural architectures. Our contribution lies in how multimodal anomaly detection and biomechanical knowledge are integrated to perform error-level diagnosis, rather than introducing a new standalone neural backbone.

(1) From skill-level attribution to error-level diagnosis.

ChainBMD first predicts a player's skill level and then attributes the prediction to body regions before mapping them to coaching errors. However, a body region associated with a skill prediction does not necessarily identify which biomechanical rule was violated, when the violation occurred, or what measurement supports the diagnosis. BIC instead directly models expert execution and identifies localized deviations. These observations are mapped to candidate biomechanical rules, which are subsequently verified against measurable sensor-derived features and expert-calibrated thresholds. Thus, coaching errors are established through explicit biomechanical evaluation rather than inferred from skill-classification attribution.

(2) Biomechanically conditioned anomaly modeling.

The contribution is not the use of a graph itself, but how biomechanical relationships are incorporated into the definition of abnormal execution. In multimodal sports data, individual sensor signals may appear plausible in isolation while their coordination across body parts is inconsistent with expert execution. BIC addresses this by conditioning future-signal reconstruction on anatomically related modalities, using local-global contexts and cross-modal consistency learning. Consequently, anomaly detection considers both modality-specific deviations and their relationships with other biomechanical measurements.

(3) Knowledge-guided verification and discovery.

BIC combines deterministic rule verification with MLLM-based discovery of errors beyond the predefined rule catalog. Removing KG-based verification reduces coaching F1 from 0.728 to 0.432 on the common ablation reference pool (Section 4.4, and the answer to Q3 below), demonstrating its substantial contribution to reliable error identification.

Therefore, BIC's contribution extends beyond combining existing components. It reformulates coaching from skill-level attribution to error-level diagnosis, incorporates biomechanical dependencies into multimodal anomaly detection, and verifies identified errors against measurable evidence. Together, these mechanisms provide a systematic framework for producing localized, evidence-grounded coaching feedback, with their effectiveness supported by the component ablations.

AuthorsOpenReview ↗

Response to Reviewer NZt2 (Part 2)

W2(b): Transfer beyond badminton coaching.

We thank the reviewer for asking how BIC’s design might apply beyond badminton. BIC illustrates two transferable design principles for explainable multimodal time-series analysis. First, anomaly detection can incorporate domain-defined relationships between heterogeneous measurements instead of treating their channels independently. Second, deterministic rule verification can connect learned anomaly signals to interpretable findings before language-model-based discovery. These principles may be applicable to other sensor-based movement assessment tasks, and their generalization are worth exploring.

W1(b) and Q2: How were findings involving non-visible sensor modalities evaluated?

(1) Evaluation protocol (Appendix I.3).

Before evaluation, all four badminton experts were introduced to the sensing modalities, including what each sensor measures and how to interpret its visualizations. For each finding, they received player-versus-matched-expert sensor plots corresponding to the cited body region, alongside player video and skeleton playback. The plots were aligned with stroke phases, allowing experts to assess whether the reported biomechanical error was supported by the measured signals during the relevant phase.

We added Figure 18 in Appendix I.3 to illustrate this procedure using an insole-related finding. It compares the player's plantar-pressure distributions with those of the matched expert, with stroke phases annotated and the impact frame highlighted. Reports were anonymized and randomized, and TSAD annotations were hidden.

(2) Expert judgments across modalities (Appendix I.8).

We further report the distribution of expert verdicts for BIC findings by supporting modality:

Evidence modalityShareAgreeNot sureDisagree
Skeleton31.0%95.7%0.9%3.4%
EMG21.9%86.6%3.0%10.4%
Plantar pressure38.0%83.7%4.1%12.2%
Gaze9.1%64.4%11.8%23.8%

EMG and plantar-pressure evidence together account for 59.9% of BIC findings. Among expert verdicts, 86.6% for EMG and 83.7% for plantar pressure are positive, demonstrating that experts frequently supported findings based on these non-visible measurements.

AuthorsOpenReview ↗

Response to Reviewer NZt2 (Part 3)

W2(c, d) and W3: Traceability, KG auditing and validation, and the role of Level-2 relationships.

(1) Quantitative evaluation of traceability (Appendix G.5).

We introduce a traceability analysis in Appendix G.5 of the revised manuscript. A finding is considered rule-verified when it can be traced through an identified anomaly location, a matched Level-1 coaching rule, and a deterministic measurement evaluated against an explicit acceptance threshold. This establishes a measurable evidence path from sensor observations to the reported coaching error.

Stroke typeRule-verified (83.4%)Stage 3 (16.6%)
Backhand87.1%78.3%
Forehand92.4%97.8%
All strokes91.4%91.3%

From the above table, across all BIC findings, 83.4% are rule-verified, while the remaining 16.6% originate from Stage-3 MLLM discovery. The two groups achieve comparable expert acceptance rates of 91.4% and 91.3%, respectively. These results quantify the coverage of deterministic verification while demonstrating the complementary role of MLLM-based discovery.

(2) Expert auditing of the knowledge graphs (Appendix G.4).

Both KGs are initially drafted by an LLM but undergo expert review before deployment (Appendix G). Three badminton experts examined the coaching rules and biomechanical relationships, correcting, removing, or adding entries where necessary. These auditors are distinct from the four experts evaluating coaching reports. They reviewed and edited a shared KG individually without direct communication. All revisions were visible to the other auditors, who could retain or further revise previously edited entries. Thus, the auditing process involved iterative expert refinement rather than independent blinded annotation or formal consensus voting. Consequently, inter-auditor agreement was not measured. We added the following audit statistics in Appendix G.4:

Graph elementDraftedEditedRemovedAddedFinal
Level-1 rules18239202164
Level-2 edges29669141283

As such, the deployed KG is not used directly from LLM generation but is refined through domain-expert auditing.

(3) Expert evaluation of instantiated Level-2 explanations (Appendix G.6).

We also conducted an additional blinded evaluation of Level-2 relationships instantiated across 20 randomly selected strokes (10 backhand and 10 forehand) from the set that contains at least two expert-accepted findings and one relationship from both systems. In total there are 79 relationships. For each relationship, two badminton experts assess its biomechanical correctness using the connected findings and player-versus-expert multimodal visualizations.

Experts confirmed the biomechanical correctness of 69/79 relationships (87.3%), with 10 uncertain and none disagree. Agreement between raters reached 91.1% (72/79).

(4) Clarification of Level-2 inference.

We apologize for the ambiguity in Appendix F.2. The Level-2 KG serves an explanatory rather than a confidence-scoring function. After Level-1 deterministic evaluators verify individual errors, Level 2 connects them through expert-reviewed causal and correlational relationships to construct explanation chains. It does not modify detection confidence or validate additional errors.

The confidence-adjustment part in Appendix F.2 is intended to provide the LLM with the task context and explain the concept of biomechanical corroboration. It guides the LLM in constructing meaningful biomechanical relationships, rather than specifying an implemented confidence-update mechanism. During inference, Level 2 is used solely to organize already verified errors into explanatory chains, consistent with Section 3.4.

Reviewer_drHQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes the Badminton Intelligent Coach (BIC), a system that identifies and explains badminton technique errors from synchronized skeletal motion, gaze, plantar pressure, and electromyography. An expert-trained conditional diffusion model detects deviations from expert movement, using modality-specific denoisers and local/global cross-modal conditioning. Detected anomalies trigger biomechanical rules in an expert-audited knowledge graph, which provides candidate explanations. A multimodal language model then compares sensor visualizations with a matched expert reference to identify additional issues and generate coaching feedback.

Validation uses MultiSenseBadminton data for backhand drives and forehand clears, with four subject-disjoint expert training/calibration/test configurations and eight beginner participants. BIC reports skill-discrimination AUROCs of 0.886 and 0.949. On 60 strokes assessed by four badminton experts, it reports an error-detection F1 of 0.760 versus 0.540 for the strongest listed baseline, alongside higher expert ratings of feedback quality. Ablations support the contributions of cross-modal conditioning, the consistency objective, modality-specific modeling, and the final multimodal reasoning stage.

Critical strengths
  1. A concrete bridge from sensor anomalies to interpretable coaching. BIC combines modality-specific temporal modeling, explicit biomechanical rule checks, an expert-audited knowledge graph, and reference-guided multimodal interpretation. This matters because it bridges the gap between recognizing non-expert movement and explaining a specific, potentially correctable technique issue. The detailed prompts and rule examples make the approach more inspectable than a purely free-form coaching system.

  2. Strong reported detection results with useful component validation. The expert-subject-disjoint experiments show substantial skill-discrimination gains over the listed anomaly-detection baselines. Ablations show meaningful reductions when removing conditioning or consistency terms, and removing the final multimodal stage reduces coaching F1 from 0.760 to 0.652. Together with the four-expert assessment, these results provide encouraging evidence that the integrated system is useful, subject to the evaluation limitations below.

Critical weaknesses
  1. The coaching reference set is partly constructed from the outputs being evaluated. Section 4.3 pools expert-endorsed findings from the compared systems, adds expert-identified omissions, and uses an LLM for semantic deduplication. Expert review and additions are valuable safeguards, and the LLM is not the correctness judge. Nevertheless, recall remains relative to an output-conditioned pool rather than an independently established inventory of errors. The paper does not report inter-rater agreement, the contribution of independently added errors, or sensitivity to this pooling procedure. Furthermore, the connection between these semantic judgments and the three-consecutive-frame criterion in the phase-level evaluation is insufficiently clear. These gaps make it difficult to assess the reliability of the central precision/recall results on the 60-stroke evaluation.

  2. The generalist baselines receive materially different information and output constraints. Appendix F.4 limits their reports to at most five findings and three priority actions, excludes expert references, and prohibits inferred relationships or causal chains. BIC receives a matched expert reference and graph-supported relationships; its example in Appendix H.3 contains eight findings and four actions. These differences can directly affect recall, comprehensiveness, and perceived professionalism or actionability. The comparison shows an advantage over these baseline configurations, but it does not isolate the benefit of anomaly detection and knowledge guidance from the benefit of additional reference information and a larger reporting budget. This weakens the interpretation of the end-to-end gains.

Questions for authors
  1. What exact agreement threshold and treatment of “not sure” were used to accept findings? Please report inter-rater agreement, the number of reference errors experts added independently, and whether the knowledge-graph auditors overlapped with the evaluators. Was an independent annotation or leave-one-system-out pooling analysis conducted, and, if so, how sensitive are the rankings to the reference-set construction?

  2. How were the ground-truth frame intervals for the “at least three consecutive anomalous frames” criterion established? Were they independently annotated from the footage, or derived from the phase-level error pool? Please explain the mapping used to obtain the labels and scores in Table 2.

  3. Why were the generalist baselines restricted to five findings and three actions without an expert reference? Do you have results using the same reference information and report budget across methods? Such results would help determine how much of the coaching improvement comes from the proposed anomaly-detection and knowledge-guidance stages.

6 comments
AuthorsOpenReview ↗

Response to Reviewer drHQ (Part 1)

We thank the reviewer for the detailed and constructive feedback. We particularly appreciate the concerns regarding the reliability of the expert-adjudicated reference pool, the fairness of the MLLM comparisons, and the phase-level evaluation protocol. To address these concerns, we conducted additional controlled experiments, statistical analyses, and sensitivity studies, and incorporated the key findings and methodological clarifications into the revised manuscript. We respond to each concern below.

W1 and Q1: Reference-pool construction, expert agreement, and statistical reliability.

We acknowledge that the reference pool in the manuscript is built from accepted system findings and expert-reported omissions rather than an independently exhaustive inventory. In the revised version, we clarified the adjudication protocol and reference-pool construction (Section 4.3, Appendix I.3), reported expert agreement and modality-specific judgments (Appendix I.8-I.9), and evaluated sensitivity to reference-pool construction (Appendix I.6).

(1) Expert adjudication and pool construction.

In specific, the four badminton experts independently reviewed each proposed finding. Acceptance required at least two "agree" verdicts and no `"disagree" verdicts. A "not sure" verdict was neutral. We replaced the phrase "majority agreement" in Section 4.3 with this explicit criterion. Experts independently added 31 omitted errors (2.6% of the original pool). Semantically equivalent findings were merged using an LLM, while correctness was determined exclusively by expert judgments. The KG auditors were distinct from the report evaluators. We explicitly report recall relative to this expert-adjudicated pool as pooled recall.

(2) Inter-rater agreement and statistical significance (Appendix I.8-I.9).

For Table 3 of the manuscript, the finding-level ordinal Krippendorff's α\alpha is 0.671, with 83.9% raw pairwise agreement. For coaching-quality ratings (Table 4 of the manuscript), the inter-rater agreements are summarized here:

MetricAcc.Compr.Prof.Act.
Ordinal α0.5770.6390.6090.627
Within one point (%)87.887.387.382.9

Ordinal α\alpha ranges from 0.577 to 0.639, and 82.9--87.8% of rating pairs differ by at most one point on the five-point scale. We further conducted significance test to Table~4, BIC significantly outperforms the strongest baseline across all four quality dimensions under paired Wilcoxon tests (p<10−5p<10^{-5}). On the original submission pool, a paired bootstrap stratified by stroke type (10,000 resamples) yields a mean bootstrap F1 improvement of 0.2140.214 for BIC over the original all-modality MLLM baseline (95% CI [0.134,0.295][0.134, 0.295], one-sided p<10−4p < 10^{-4}).

(3) Reference-pool sensitivity (Appendix I.6).

We did not conduct a separate, independently annotated error inventory prior to reviewing system outputs, because identifying all biomechanical errors across heterogeneous sensor modalities would require extensive stroke-by-stroke expert annotation without a predefined error inventory.

Starting from the original submission pool, we evaluate the sensitivity of coaching correctness to the composition of the expert-adjudicated reference pool. We recompute recall and F1 after removing errors identified exclusively by each system in turn. Expert-adjudicated precision remains unchanged, so the resulting F1 values measure sensitivity to the reference pool rather than newly adjudicated accuracy. Results on the original four-system pool are shown below:

Pool variantChainBMDMLLM (skel)MLLM (all)BIC
Original0.4080.5110.5400.760
Without BIC-only errors0.4930.6480.6400.657
Without ChainBMD-only errors0.3360.5420.5750.803
Without MLLM (skel)-only errors0.4280.4640.5600.789
Without MLLM (all)-only errors0.4120.5180.5180.772

Across all five reference-pool variants, BIC retains the highest F1, including when BIC-exclusive errors are removed.

AuthorsOpenReview ↗

Response to Reviewer drHQ (Part 2)

W2(a, b) and Q3: Output constraints, expert references, prohibited causal chain.

We tested three changes to the MLLM baselines: (1) the removal of output limits, (2) the provision of matched-expert references, and (3) their combined effect. Appendix L.2-L.4 of the revised manuscript describes the configurations, and Table 27 reports their results.

We first compare the number of findings and actions reported by the capped and uncapped MLLMs. We count findings and actions directly from the generated reports, independent on the reference pool used in Table 3 manuscript.

MethodNum. of Findings (Mean ± SD)Num. of Actions (Mean ± SD)
MLLM (skel), capped3.44 ± 0.612.62 ± 0.70
MLLM (skel), uncapped2.91 ± 0.382.65 ± 0.44
MLLM (all), capped4.09 ± 0.622.85 ± 0.44
MLLM (all), uncapped3.03 ± 0.522.74 ± 0.51

Removing the caps reduced the mean number of findings for both MLLMs. Neither uncapped configuration reported more than four findings per stroke. The mean number of actions changed little, rising slightly for MLLM (skel) and falling slightly for MLLM (all).

(1) MLLM fair-comparison pool (Appendix L.1).

The original submission pool used in Table 3 combines expert-accepted findings from ChainBMD, the original MLLM (skel), the original MLLM (all), and full BIC, plus expert-reported omissions. For the MLLM fair comparison here, the same four experts adjudicated new findings from the MLLM variants using the original blinded protocol (Section 4.3 and Appendix I.3). The accepted findings were then semantically deduplicated and merged with the original adjudicated findings, as described in Appendix L.1. All rows below, including the original capped MLLM outputs and BIC, are scored against this same MLLM fair-comparison pool. The added findings include errors missed by BIC, which changes its recall and F1 relative to Table 3.

We first report all the results in the table below and explain each findings subsequently. The best score for each configuration is shown in bold.

TestMethodRef.CapP ↑R ↑Excl. R ↑F1 ↑
--MLLM (skel), capped--50.7950.3070.1290.432
(1)MLLM (skel), uncapped----0.7750.2340.0740.356
(2)MLLM (skel) + ref.✓50.8810.2930.1030.458
(3)MLLM (skel) + ref., uncapped✓--0.7660.2480.0540.370
--MLLM (all), capped--50.7550.3440.0540.463
(1)MLLM (all), uncapped----0.7060.2460.0350.358
(2)MLLM (all) + ref.✓50.8610.3690.0760.505
(3)MLLM (all) + ref., uncapped✓--0.7860.2330.0260.354
--BIC (ours)✓--0.8940.5430.4230.660

Table R1. Finding-level results on the MLLM fair-comparison pool.

(2) Removing output limits (Appendix L.2).

The original five-finding and three-action limits aim to follow a precision-first coaching strategy, consistent with BIC's Stage-3 prompting. From Table R1 ("MLLM (skel/all), uncapped"), removing these limits does not improve the MLLM baselines. Instead, F1 decreases from 0.4320.432 to 0.3560.356 for MLLM (skel) and from 0.4630.463 to 0.3580.358 for MLLM (all).

(3) Providing matched-expert references (Appendix L.3).

The original MLLM baselines excluded expert references to benchmark standard zero-shot MLLM analysis of player streams alone, treating FastDTW expert retrieval as part of BIC's proposed Stage-3 design. To isolate whether matched-expert visuals explain the gap, we now provide the MLLM baselines with the same phase-aligned player-versus-expert visualizations as BIC's Stage 3 (using FastDTW). Stage-1 anomaly annotations and Stage-2 KG verification outputs were omitted. Their prompts are also updated to instruct them to compare the executions when identifying coaching errors.

As shown in Table R1 ("MLLM (skel/all) + ref"), this improves their precision and F1, which supports our motivation for incorporating matched-expert comparisons into BIC's Stage 3. Nevertheless, BIC retains substantially higher recall and F1 compared to the strongest reference-augmented baseline in the same reference pool. These results highlight the additional contribution of BIC's localized multimodal anomaly detection and deterministic biomechanical verification beyond reference-guided MLLM reasoning alone.

AuthorsOpenReview ↗

Response to Reviewer drHQ (Part 3)

(4) Removing output limits AND providing matched-expert reference (Appendix L.4).

We additionally test both changes simultaneously. From Table R1 ("MLLM (skel/all) + ref, uncapped"), the uncapped reference-augmented MLLMs achieve F1 of 0.370 (skel) and 0.354 (all), both below their corresponding capped, reference-augmented configurations (0.458 and 0.505, respectively). This confirms that combining these conditions does not close the performance gap.

(5) Prohibited relationships or causal chains inference.

The original MLLM prompts restricted inferred causal relationships to keep individual coaching findings grounded in observed evidence and independently assessable under our finding-level adjudication protocol. This constraint was intended to reduce unsupported causal interpretations rather than limit the number of detectable errors. Nevertheless, we acknowledge that it may restrict the baseline's ability to explain relationships between findings.

To examine this concern, we additionally evaluate the strongest MLLM baseline (MLLM (all) + matched-expert reference as shown in Table R1.) with inter-finding relationships permitted. Specifically, we removed the instruction "Do not infer or output relationships, causal chains, knowledge-graph links, or connections between findings" and introduce a separate relationships field in the JSON output. Each relationship identifies two reported findings and specifies their relationship type.

We conduct an additional blinded evaluation by two badminton experts on 20 randomly selected strokes (10 backhand and 10 forehand) from the set that contains at least two expert-accepted findings and one relationship from both systems. These give 110 generated relationships (79 from BIC and 31 from the MLLM). Experts assess the correctness of each relationship from the corresponding stroke evidence.

Experts confirmed the biomechanical correctness of 87.3% of BIC's relationships (69/79), compared with 54.8% for the relationship-permitted MLLM (17/31). Inter-rater agreement is 91.1% for BIC and 71.0% for the MLLM baseline, respectively. BIC therefore had a higher proportion of expert-confirmed relationships in this evaluation (Appendix G.6).

Although permitting relationships increases the MLLM's explanatory flexibility, only 54.8% of the proposed relationships are rated as correct. Thus, allowing causal inference does not necessarily yield more reliable coaching explanations.

(6) Additional report-quality evaluation (Appendix L.5).

To assess whether the baseline constraints affect overall coaching quality, we conducted a follow-up blinded evaluation on all 60 strokes with the same four badminton experts. We evaluated three MLLM (all) configurations: (i) with matched-expert references, (ii) with matched-expert references and both output limits removed, and (iii) with matched-expert references and inter-finding relationships permitted. BIC was also re-evaluated in the same round to ensure a direct comparison. Experts rated each report on the original four 5-point Likert dimensions. The original MLLM (all) scores from Table 4 are included for reference and were not re-evaluated.

MethodAcc.Compr.Prof.Act.
MLLM (all)3.583.693.633.66
MLLM (all) + ref., uncapped3.833.104.054.05
MLLM (all) + ref.3.833.364.103.91
MLLM (all) + ref. + relationships3.893.524.194.08
BIC (ours)4.524.424.544.48

The relationship-enabled MLLM achieves the highest mean ratings among the three revised configurations on all four dimensions, but remains below BIC. Inter-rater agreement, measured using ordinal Krippendorff's α\alpha, was 0.59, 0.59, 0.67, and 0.53 for accuracy, comprehensiveness, professionalism, and actionability, respectively, with corresponding within-one-point agreement rates of 78%, 71%, 87%, and 82%. These results indicate that providing matched-expert references, removing output limits, and permitting inter-finding relationships do not eliminate BIC's advantage in expert-rated coaching quality. The revised manuscript includes this follow-up evaluation in Appendix L.5.

Reviewer_79nHOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper presents BIC, a three-stage pipeline for explainable badminton-stroke coaching from wearable sensors (MultiSenseBadminton: skeleton, EMG, insole pressure, gaze).

Critical strengths
  • Substantial human evaluation and documentation. Four experts adjudicate findings blindly, can add missed errors, and rate report quality, which goes beyond proxy metrics. BIC’s margins are consistent across all of these measures. Prompts, KG statistics, rule definitions and the annotation interface are documented in detail.
  • Traceable detect-then-verify design. Each error can be traced from a flagged location to a KG rule and a deterministic test. The Stage-3 ablation shows that the rule core and the MLLM discovery step are complementary.
Critical weaknesses
  • The coaching comparison is confounded. The MLLM baselines are capped at five findings and receive no expert reference. BIC is uncapped and sees DTW-matched expert plots and expert-calibrated ranges, which alone can inflate recall, exclusive recall and comprehensiveness. The ground-truth pool comes only from the compared systems, and neither inter-rater agreement nor significance tests are reported. The missing control is Stage 3 alone with the same visuals.
  • The TSAD evidence is fragile. Each fold separates one held-out expert from 8 beginners, with intermediates removed.
Questions for authors
  • Do BIC’s margins hold if the MLLM baselines are uncapped and given the same matched-expert visuals (i.e., Stage 3 alone)?
  • How were the TSAD baselines adapted, and why do several score below 0.5 AUROC? How does a simple DTW-to-nearest-expert baseline perform?
  • Are the conditioning contexts built from past embeddings only, or do they include future segments? How many diffusion samples are drawn per anomaly score?

Thank you for the additional experiments, which address my questions. My remaining concerns are that the reference pool is still built from the compared systems' outputs rather than an independent annotation, the subject-level evaluation rests on 4 experts and 8 beginners from one dataset and two stroke types, and the blinded quality ratings for the revised baselines are still pending. I keep my score.

14 comments
AuthorsOpenReview ↗

Response to Reviewer 79nH (Part 1)

Thank you for recognizing the traceable detect-then-verify design and for emphasizing the importance of fair baselines and independent evidence for skill discrimination. We address the evaluation and implementation concerns below.

W1(a): Fair comparison with MLLM baselines.

We tested three changes to the MLLM baselines: (1) the removal of output limits, (2) the provision of matched-expert references, and (3) their combined effect. Appendix L.2-L.4 of the revised manuscript describes the configurations, and Table 27 reports their results.

We first compare the number of findings and actions reported by the capped and uncapped MLLMs. We count findings and actions directly from the generated reports, independent on the reference pool used in Table 3 manuscript.

MethodNum. of Findings (Mean ± SD)Num. of Actions (Mean ± SD)
MLLM (skel), capped3.44 ± 0.612.62 ± 0.70
MLLM (skel), uncapped2.91 ± 0.382.65 ± 0.44
MLLM (all), capped4.09 ± 0.622.85 ± 0.44
MLLM (all), uncapped3.03 ± 0.522.74 ± 0.51

Removing the caps reduced the mean number of findings for both MLLMs. Neither uncapped configuration reported more than four findings per stroke. The mean number of actions changed little, rising slightly for MLLM (skel) and falling slightly for MLLM (all).

(1) MLLM fair-comparison pool (Appendix L.1).

The original submission pool used in Table 3 combines expert-accepted findings from ChainBMD, the original MLLM (skel), the original MLLM (all), and full BIC, plus expert-reported omissions. For the MLLM fair comparison here, the same four experts adjudicated new findings from the MLLM variants using the original blinded protocol (Section 4.3 and Appendix I.3). The accepted findings were then semantically deduplicated and merged with the original adjudicated findings, as described in Appendix L.1. All rows below, including the original capped MLLM outputs and BIC, are scored against this same MLLM fair-comparison pool. The added findings include errors missed by BIC, which changes its recall and F1 relative to Table 3.

We first report all the results in the table below and explain each findings subsequently. The best score for each configuration is shown in bold.

TestMethodRef.CapP ↑R ↑Excl. R ↑F1 ↑
--MLLM (skel), capped--50.7950.3070.1290.432
(1)MLLM (skel), uncapped----0.7750.2340.0740.356
(2)MLLM (skel) + ref.✓50.8810.2930.1030.458
(3)MLLM (skel) + ref., uncapped✓--0.7660.2480.0540.370
--MLLM (all), capped--50.7550.3440.0540.463
(1)MLLM (all), uncapped----0.7060.2460.0350.358
(2)MLLM (all) + ref.✓50.8610.3690.0760.505
(3)MLLM (all) + ref., uncapped✓--0.7860.2330.0260.354
--BIC (ours)✓--0.8940.5430.4230.660

(2) Removing output limits (Appendix L.2).

The original five-finding and three-action limits aim to follow a precision-first coaching strategy, consistent with BIC's Stage-3 prompting. From Table R1 ("MLLM (skel/all), uncapped"), removing these limits does not improve the MLLM baselines. Instead, F1 decreases from 0.4320.432 to 0.3560.356 for MLLM (skel) and from 0.4630.463 to 0.3580.358 for MLLM (all).

(3) Providing matched-expert references (Appendix L.3).

The original MLLM baselines excluded expert references to benchmark standard zero-shot MLLM analysis of player streams alone, treating FastDTW expert retrieval as part of BIC's proposed Stage-3 design. To isolate whether matched-expert visuals explain the gap, we now provide the MLLM baselines with the same phase-aligned player-versus-expert visualizations as BIC's Stage 3 (using FastDTW). Stage-1 anomaly annotations and Stage-2 KG verification outputs were omitted. Their prompts are also updated to instruct them to compare the executions when identifying coaching errors.

As shown in Table R1 ("MLLM (skel/all) + ref"), this improves their precision and F1, which supports our motivation for incorporating matched-expert comparisons into BIC's Stage 3. Nevertheless, BIC retains substantially higher recall and F1 compared to the strongest reference-augmented baseline in the same reference pool. These results highlight the additional contribution of BIC's localized multimodal anomaly detection and deterministic biomechanical verification beyond reference-guided MLLM reasoning alone.

AuthorsOpenReview ↗

Response to Reviewer 79nH (Part 2)

(4) Removing output limits AND providing matched-expert reference (Appendix L.4).

We additionally test both changes simultaneously. From Table R1 ("MLLM (skel/all) + ref, uncapped"), the uncapped reference-augmented MLLMs achieve F1 of 0.370 (skel) and 0.354 (all), both below their corresponding capped, reference-augmented configurations (0.458 and 0.505, respectively). This confirms that combining these conditions does not close the performance gap.

(5) Additional report-quality evaluation (Appendix L.5).

To assess whether the baseline constraints affect overall coaching quality, we conducted a follow-up blinded evaluation on all 60 strokes with the same four badminton experts. We evaluated three MLLM (all) configurations: (i) with matched-expert references, (ii) with matched-expert references and both output limits removed, and (iii) with matched-expert references and inter-finding relationships permitted. BIC was also re-evaluated in the same round to ensure a direct comparison. Experts rated each report on the original four 5-point Likert dimensions. The original MLLM (all) scores from Table 4 are included for reference and were not re-evaluated.

MethodAcc.Compr.Prof.Act.
MLLM (all)3.583.693.633.66
MLLM (all) + ref., uncapped3.833.104.054.05
MLLM (all) + ref.3.833.364.103.91
MLLM (all) + ref. + relationships3.893.524.194.08
BIC (ours)4.524.424.544.48

The relationship-enabled MLLM achieves the highest mean ratings among the three revised configurations on all four dimensions, but remains below BIC. Inter-rater agreement, measured using ordinal Krippendorff's α\alpha, was 0.59, 0.59, 0.67, and 0.53 for accuracy, comprehensiveness, professionalism, and actionability, respectively, with corresponding within-one-point agreement rates of 78%, 71%, 87%, and 82%. These results indicate that providing matched-expert references, removing output limits, and permitting inter-finding relationships do not eliminate BIC's advantage in expert-rated coaching quality. The revised manuscript includes this follow-up evaluation in Appendix L.5.

AuthorsOpenReview ↗

Response to Reviewer 79nH (Part 3)

W1(b): Reference-pool validity, inter-rater agreement, and statistical significance.

The original submission pool (Section 4.3, Appendix I.3, and Table 3) combines expert-accepted findings from ChainBMD, the original MLLM (skel), the original MLLM (all), and full BIC, plus expert-reported omissions. Four badminton experts independently reviewed each proposed finding. Acceptance required at least two "agree" verdicts and no "disagree" verdicts. A "not sure" verdict was neutral. We replaced the phrase "majority agreement" in Section 4.3 with this explicit criterion. Experts independently added 31 omitted errors (2.6% of the original pool). We report recall relative to this expert-adjudicated pool as pooled recall.

(1) Inter-rater agreement and statistical significance (Appendix I.8-I.9).

For Table 3 of the manuscript, the finding-level ordinal Krippendorff's α\alpha is 0.671, with 83.9% raw pairwise agreement. For coaching-quality ratings (Table 4 of the manuscript), the inter-rater agreements are summarized here:

MetricAcc.Compr.Prof.Act.
Ordinal α0.5770.6390.6090.627
Within one point (%)87.887.387.382.9

Ordinal α\alpha ranges from 0.577 to 0.639, and 82.9--87.8% of rating pairs differ by at most one point on the five-point scale. We further conducted significance test to Table 4, BIC significantly outperforms the strongest baseline across all four quality dimensions under paired Wilcoxon tests (p<10−5p<10^{-5}). On the original submission pool, a paired bootstrap stratified by stroke type (10,000 resamples) yields a mean bootstrap F1 improvement of 0.2140.214 for BIC over the original all-modality MLLM baseline (95% CI [0.134,0.295][0.134, 0.295], one-sided p<10−4p < 10^{-4}).

(2) Reference-pool sensitivity (Appendix I.6).

Starting from the original submission pool used in Table 3, we remove errors identified exclusively by each system in turn and recompute recall and F1 with expert-adjudicated precision held fixed.

Pool variantChainBMDMLLM (skel)MLLM (all)BIC
Original0.4080.5110.5400.760
Without BIC-only errors0.4930.6480.6400.657
Without ChainBMD-only errors0.3360.5420.5750.803
Without MLLM (skel)-only errors0.4280.4640.5600.789
Without MLLM (all)-only errors0.4120.5180.5180.772

BIC ranks first in every variant, even after removing BIC-exclusive errors.

W1(c) and Q1: Stage 3 alone with same visuals.

To test how well Stage 3 performs on its own with the same matched-expert visualizations as full BIC, we evaluate four configurations of BIC: (1) full BIC, (2) BIC without Stage 3, retaining only TSAD-based localization and deterministic KG verification, (3) BIC without Stage 2, providing Stage 3 with matched-expert references and TSAD hints but no KG verification, and (4) Stage 3 alone, using the original BIC's Stage 3 prompt (Appendix F.3) and providing matched-expert references without TSAD hints or KG outputs. We start from the original submission pool used in Table 3. The same four experts adjudicated new findings from the three BIC ablations using the original blinded protocol (Section 4.3 and Appendix I.3). The accepted findings were then semantically deduplicated and merged with the original adjudicated findings. All four rows below, including full BIC and the three ablations, are scored against this same BIC ablation pool. The added findings include errors missed by the full BIC, which changes its recall, Excl. R and F1 relative to Table 3.

MethodS1S2S3P ↑R ↑Excl. R ↑F1 ↑
BIC (full)✓✓✓0.8940.6370.3880.728
BIC w/o Stage 3✓✓--0.8700.5130.3460.634
BIC w/o Stage 2✓--✓0.7980.3030.0260.432
Stage 3 alone----✓0.7860.2740.0060.397

With the same FastDTW-matched expert visualizations as full BIC and without TSAD hints or KG outputs, Stage 3 alone achieves recall 0.274 and F1 0.397, compared with 0.637 and 0.728 for full BIC on the same ablation pool. The results are added to Section 4.4 of the revised manuscript.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We sincerely thank all three reviewers (79nH, drHQ, and NZt2) for their constructive suggestions and recognition of the strengths of our work. In particular, we appreciate the following positive assessments:

  • Intuitive and well-motivated coaching formulation. Reviewer NZt2 recognized that formulating coaching as detecting deviations before judging performance is intuitive and better aligned with human coaching than conventional classify-then-attribute approaches. The reviewer also acknowledged the sensible incorporation of biomechanical priors through the body-part--sensor graph.
  • Traceable and interpretable error diagnosis. Reviewer 79nH highlighted BIC's traceable detect-then-verify design and the complementary contributions of rule-based verification and MLLM discovery. Reviewer drHQ recognized the concrete bridge from sensor anomalies to interpretable coaching, emphasizing the identification of specific, potentially correctable technique errors. Reviewer NZt2 also acknowledged the complete pipeline connecting anomaly detection to rule-verified, human-readable feedback.
  • Substantial empirical evaluation and documentation. Reviewer 79nH recognized the extensive human-expert evaluation, consistent performance improvements, and detailed documentation of rules and prompts. Reviewer drHQ highlighted the strong detection results and meaningful component ablations, while Reviewer NZt2 appreciated the per-finding expert adjudication and comprehensive implementation details.

We also appreciate the reviewers' concerns regarding evaluation fairness, reference-pool reliability, generalization, and methodological clarity. We have conducted additional experiments and strengthened the manuscript accordingly. The major revisions are summarized below.

1. Fairer Baseline Comparisons and Stage Ablations [79nH, drHQ, NZt2]

We evaluated MLLM baselines with output limits removed, with matched-expert references, and with both changes jointly (Appendix L). We further evaluated relationship-enabled MLLM reports through blinded human-expert assessments of coaching quality and biomechanical relationship correctness (Appendices L.5 and G.6). We also ablated BIC's three stages (Section 4.4 and Appendix M.1). BIC achieved the highest finding-level F1 in both comparisons.

2. Expert Evaluation Reliability and Statistical Validation [79nH, drHQ, NZt2]

We clarified expert adjudication and reference-pool construction, quantified inter-rater agreement and statistical significance, and evaluated sensitivity to reference-pool composition (Section 4.3, Appendix I.3, I.6, I.8, I.9). The results support the reliability and robustness of BIC's coaching improvements.

3. Subject-Level Generalization and Baseline Controls [79nH, NZt2]

We added subject-level AUROC, one-sided exact Mann--Whitney tests, and a nearest-expert FastDTW baseline (Section 4.2 and Table 1). BIC achieved the highest subject-level AUROC on both stroke types.

4. Biomechanical Grounding and KG Validation [NZt2, drHQ]

We quantified rule-based traceability, documented expert auditing of both KGs, and directly evaluated the correctness of biomechanical relationships reported by BIC and the relationship-enabled MLLM baseline (Appendix G.4-G.6). These analyses strengthen the evidence supporting BIC's explainable coaching feedback.

5. Methodological Clarifications and Sensitivity Analyses [79nH, drHQ, NZt2]

We clarified diffusion conditioning, baseline adaptation, phase annotation, and anomaly-score calibration (Sections 3.3 and 4.2-4.3 and Appendices I.1 and K.4). We tested sensitivity to diffusion sampling, window length, phase boundaries, and the minimum consecutive-frame criterion (Appendices E.1, E.4, I.1, and I.5).

6. Recall Against Independent Expert Annotations [79nH]

We conducted an independent expert annotation on 20 strokes, with system outputs and prior adjudications hidden (Appendix I.7). BIC retained its recall advantage over the baselines when evaluated against these expert annotations.

7. Evaluation Across Different Sensing Settings [79nH]

To assess BIC’s adaptability, we retrained it in two settings: using IMU-derived skeletal motion alone and using all sensor streams except pressure insoles. In both settings, BIC achieved higher precision and recall than MLLM baselines given the same sensor inputs (Appendix M.4).

We sincerely appreciate the reviewers' valuable feedback, which has helped us strengthen the experimental validation and methodological transparency of our work. The revisions are highlighted in blue in the updated manuscript. We hope these additional experiments and clarifications address the reviewers' concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.