The gap between “this system recognises what you did” and “this system helps you do it better” is where almost all wearable movement technology falls down. Your watch counts reps. It has nothing to say about the reps.
SPAR — Skill Profiling with Attributable Reasoning, posted 25 September 2026 — is built specifically to close that gap. The authors are Nibraas Khan, Hanchen David Wang, Enya Bullard, Ritam Ghosh and Ruj Haan.
The problem, stated well
The paper opens with an unusually good description of what it’s measuring: 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.
That’s the crux. A punch is not an arm movement. Power comes from the ground up through a timed sequence, and the failure mode is almost always timing between segments rather than any single segment being wrong. Which is exactly what a system that only classifies which punch was thrown cannot tell you.
As they put it: wearable sensors can capture this in the gym, but most deployable systems only classify which punch was thrown rather than assess how well it was thrown.
The system
- An eight-IMU garment plus pressure insoles — so the kinetic chain is instrumented from the floor up, which is the only way to see leg-to-trunk-to-arm sequencing
- Classifies each punch as expert or novice
- Treats an explanation of that prediction as the feedback
And the design principle that makes it more than a classifier: feedback is only useful if the person receiving it can act on it, so SPAR explains its prediction at three tiers, including a per-joint attribution.
Why “the explanation is the product” is the idea worth stealing
This inverts the usual relationship between a model and its interpretability. Normally explainability is a debugging aid, or a compliance requirement, bolted on so a developer can audit a decision.
Here the explanation is the entire deliverable. The classification — expert or novice — is almost worthless on its own; a boxer already knows whether a punch felt right. What they can’t know is which joint, in which order, at which moment let it down. The per-joint attribution is the coaching.
Three tiers of explanation is the other good call, and it reflects something real about feedback: different granularities are actionable at different moments. Mid-round you can absorb “your hips are late.” Reviewing footage you want the per-joint breakdown. A coach planning a training block wants the profile across sessions. One explanation cannot serve all three.
Where this generalises
Boxing is the test case; the architecture applies to anything where skill is sequencing rather than position:
Musical performance. Instrumental technique is full-body and timing-critical in exactly this way — a drummer’s stroke, a violinist’s bow arm, a pianist’s weight transfer. Systems that tell a student they played the wrong note are abundant; systems that tell them their wrist initiated before their forearm are not.
Dance and physical performance. The same kinetic-chain logic, and the same gap between recognising a move and assessing it.
Craft and fabrication. Throwing a pot, planing a board, welding — skills where the expert/novice difference is sequencing and pressure, both of which IMUs and pressure sensors can see.
Rehabilitation, where per-joint attribution is arguably more valuable than in sport.
For anyone building movement-sensing work, the transferable lesson is the design principle rather than the hardware: decide what the person will do with the output before choosing what to detect. A classifier that’s 98% accurate at naming the movement is less useful than a rougher system that can say which part went wrong.