Skill Profiling with Attributable Reasoning (SPAR): A Wearable Analysis System for Boxing
Authors: Nibraas Khan, Hanchen David Wang, Enya Bullard, Ritam Ghosh, Ruj Haan, Aarav Agrawal, Meiyi Ma, Nilanjan Sarkar
Organizations: Department of Computer Science, Vanderbilt University, Nashville, TN 37235, USA · Department of Electrical Engineering, Vanderbilt University, Nashville, TN 37235, USA · Independent Researcher · Milton Academy, Milton, MA 02186, USA · Department of Mechanical Engineering, Vanderbilt University, Nashville, TN 37235, USA
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.
Figures & tables
Figure 1: End-to-end overview. An eight-IMU garment and pressure insoles stream to a skill classifier over anatomical joint angles. Three explanation tiers stack on the per-punch skill prediction.
Figure 2: Browser-based data-collection and review application: live 3-D skeleton viewer, WebSocket status, per-sensor calibration display, and punch-type labeling controls.
Figure 3: Two-piece instrumented garment, with sensor placement shown on the body in Fig. 1 . ( a ) Upper-body shirt, carrying the bilateral shoulder, bilateral elbow, thorax, and sacrum IMUs. ( b ) Lower-body pants, carrying the bilateral quadriceps IMUs.
ID
Sex
Ht (m)
Total
Jab
Cross
Lead hook
Rear hook
Lead uppercut
Rear uppercut
E1
M
1.89
313
71 / 978
50 / 984
43 / 1182
47 / 959
59 / 1154
43 / 955
E2
F
1.80
138
23 / 1126
23 / 1056
25 / 1278
20 / 1252
17 / 1305
30 / 1102
E3
F
1.75
174
34 / 1165
21 / 1395
34 / 1198
24 / 1274
32 / 1263
29 / 1233
E4
M
1.75
151
29 / 1299
21 / 1308
26 / 1174
21 / 1202
23 / 1117
31 / 1168
E5
M
1.75
267
31 / 1787
78 / 716
47 / 816
38 / 760
48 / 830
25 / 739
N1
F
1.57
499
82 / 1121
100 / 1037
81 / 1309
66 / 1143
85 / 1077
85 / 909
Table 1: Per-participant breakdown of the Study-1 dataset: identifier, self-reported sex (M/F), height in meters, total labeled punches, and per-type count with mean punch duration in milliseconds (count / ms).
Figure 4: The skill classifier. Top: the eight orientations are solved into 22 joint angles and combined with the six plantar zones to get 28 series. Each series is encoded on its own by the frozen MOMENT encoder and averaged into eight phase bins, giving a 28 × 8 grid of tokens. Bottom: the trained head, averaged over five seeds. Right: CAA swaps one joint’s rows between the participant’s punches and DCS copies one layer’s rows from the matched expert punch, and the modified grid is re-scored by the head.
Figure 5: Coach-evaluation viewer used in Study 2. Coaches replay a recorded punch in the 3-D skeleton view and inspect the three system outputs together, the CAA channel attribution, the DCS counterfactual, and the GNG narrative.
Participant
Punches
LOPO AUC
E1
313
0.503±0.291
E2
138
0.973±0.055
E3
174
1.000±0.002
E4
151
1.000±0.001
E5
267
0.873±0.123
N1
499
0.595±0.136
Table 2: Per-participant LOPO AUC. Each participant’s calibrated scores, averaged over the five seeds, are compared with the opposite-class punches from the other folds. The ± is the standard deviation of that AUC across the five single-seed heads, and the interval is a bootstrap over participants.
CAA — attribution ∣ϕj∣
Joint / zone
novice held-out
expert held-out
Pelvis
0.0207
0.0034
L. hip
0.0191
0.0083
R. hip
0.0057
0.0017
Lumbar
0.0051
0.0044
L. shoulder
0.0042
0.0009
Table 3: Explanation-quality audit under LOPO. CAA: the mean change in expert probability when that joint’s or zone’s series are permuted across punches, split by the held-out participant’s class. DCS: change in expert probability when a layer’s series are substituted with those of the matched expert exemplar.
Figure 6: Mean CAA attribution per unit under LOPO, split by the held-out participant’s class. (a) The eight joints. (b) The six plantar zone forces.
Participant
Foundation
Engine
Delivery
Primary gap
Top CAA unit
N1
+0.388
+0.239
+0.282
Foundation
L. hip
N2
+0.260
+0.191
+0.162
Foundation
L. elbow
N3
+0.228
+0.262
+0.175
Engine
L. hip
N4
+0.203
+0.309
+0.346
Delivery
Pelvis
N5
+0.261
+0.288
+0.334
Delivery
L. hip
N6
+0.220
+0.437
+0.102
Engine
Pelvis
Table 4: DCS and CAA per novice participant. Mean rise in expert probability over the participant’s punches when each layer is replaced with the matched expert’s, the layer with the largest rise (their primary gap), and the joint or zone with the largest CAA attribution on that participant’s fold.
Department of Computer Science, Shantou University, Shantou, China · CSSE, Shenzhen University, Shenzhen, China · Department of Sociology, Hong Kong Shue Yan University, Hong Kong, China