Can AI read a badminton stroke? What one smartwatch can and cannot measure

The question “Can AI understand a badminton stroke?” often collapses several different tasks into one. Identifying the stroke, rating its quality, estimating where the shuttle met the racket, and understanding why the player chose that action are not the same capability.

The 2026 CHI paper BadminSense offers a concrete answer to part of the question. Participants wore a Samsung Galaxy Watch 6 or 6 FE on the racket hand. The system used its inertial sensors and microphone to capture movement, collision, and vibration signals, then segmented strokes, classified stroke types, predicted stroke quality, and estimated shuttle impact location. It is not a camera-based pose-analysis system.

“One smartwatch” means one body-worn sensor, not that every computation happens on the watch itself. The prototype sends data to a Python backend and displays results through a web interface. Its dataset came from 12 right-handed amateur players and retained 848 valid real strokes. Quality labels were produced by 21 advanced players with at least five years of experience; seven independent raters evaluated each stroke from video.

The paper reports 91.43% stroke-classification accuracy, an average quality-rating error of 0.438, and an average impact-location estimation error of 12.9%. These results show that a wrist-worn device contains useful information about a badminton stroke. The watch is doing more than counting steps, and the approach does not require a separate sensor mounted on the racket handle.

Those numbers do not mean that AI can already act as a badminton coach. The paper states several limits.

• The proof of concept covers four overhead strokes—forehand clear, forehand smash, forehand drop, and backhand clear—rather than every action in a match.
• Development and evaluation involved right-handed players only.
• Data were collected with a racket strung at 24 lb; different string tensions and racket properties may change the vibration signal.
• The system does not estimate shuttle speed, even though speed and direction affect how stroke quality is understood.
• Improvement advice is rule based and does not yet use a player’s long-term history or training goals.
• The usability study focuses on short-term experience and perceived reliability. It does not show long-term skill improvement, retention, or transfer into competition.

These limits matter. A model may classify a stroke correctly in its dataset while missing compensation caused by pain. It may estimate the impact area without knowing whether the player chose the shot freely or reacted under pressure. It may produce a quality score without knowing whether success on a predictable feed will still be there next week or survive a real rally.

At Goodminton, AI is better treated as a feedback layer than as a judge or a replacement coach. It can preserve repeated signals that players and coaches would otherwise remember only vaguely. We may ask whether one rear-court action becomes more stable over several weeks, whether impact consistently drifts towards one area, or whether the pattern changes under fatigue. These records can improve the question. The answer still has to return to training.

Our working loop is goal, actual performance, error, adjustment, and verification.

First define what the session is trying to change. Then record the actual performance using video, watch data, shuttle outcome, or coach observation. The difference between the goal and the result is a clue, not an automatic diagnosis. Coach and player choose one small adjustment and test it again in the next block, the next session, or a representative rally. Without that final check, a clean chart is still only a measurement.

For the player, useful feedback does not always mean a more complicated dashboard. Three questions may matter more: what happened on this stroke, can I feel the error myself, and can I repeat the skill when the feed changes and the prompt disappears? If AI makes a player wait for the screen before judging every attempt, it may weaken self-error detection. Feedback should fade over time so the player evaluates first and the system calibrates afterwards.

AI is currently well suited to tasks such as:

• recording stability and change across repeated strokes;
• marking useful moments for post-session review;
• helping a coach find video segments worth inspecting;
• carrying a clear question into the next practice.

It should not decide alone:

• whether a problem comes from technique, timing, tactical choice, fatigue, or pain;
• whether the player needs a movement correction or a simpler task constraint;
• whether a better practice score has transferred into match capability;
• what remains safe under an injury or an individual physical limitation.

BadminSense turns the idea of fine-grained feedback from one smartwatch into a serious research prototype. It shows that the signals are useful. It also makes the boundary visible: measuring a stroke is not the same as understanding the whole player, and better performance today is not proof of learning.

Sources and boundaries:

• Chen et al., BadminSense: Enabling Fine-Grained Badminton Stroke Evaluation on a Single Smartwatch, CHI 2026. The system design, 12-participant dataset, reported metrics, and limitations come from this paper.
• Goodminton’s goal-to-verification loop and its emphasis on retention, transfer, self-error detection, and reduced feedback dependency are coaching methods. They are not experimental findings from the BadminSense paper.
• Goodminton does not claim to have deployed BadminSense, and the paper’s research results are not presented as current Goodminton product capabilities.