Machine learning and archery skill: Basri 2025 study

One hundred and four biomedical engineering students aged 18 to 22, 42 men and 62 women, with no archery experience at all, shot nine arrows with a plastic toy bow from 3 metres at a 20 cm target with rings scored 0 to 10. Anyone scoring at least 10 out of 90 was classed as able (55 students, 52.88%), the rest as not able (49, 47.12%). The able group averaged 19.50 points, the others 4.02.

Using fourteen variables (sex, weight, height, body mass index, five limb and trunk lengths, an arm strength grade, and trunk, shoulder, elbow and wrist angles at release), the authors trained three types of classifier. In cross-validation the decision tree reached 50%, which is chance level, KNN 68.1% and SVM 69.2%; after optimisation the SVM rose to 72.3% in validation and 80% on the test set. Descriptively, the able group had a wider elbow angle on the drawing arm (103.09 against 83.38 degrees) and a smaller shoulder angle (75.41 against 86.08), with almost the same trunk angle (169.08 against 167.64). The men scored more than the women (16.02 against 9.31).

The test took place in a laboratory, in three sets of three arrows. The bow was a 64 cm plastic toy with 56 cm arrows, and everyone drew the string with the right arm, whatever their dominant hand. A single camera, a tablet recording at 1080p, filmed the right side, with tape markers on the shoulder, elbow, wrist and waist; angles were measured with the free Kinovea software on the frame of release. Limb lengths were taken three times and averaged, and arm strength was graded on the Medical Research Council manual scale from 0 to 5.

The classifiers were built in MATLAB with 10-fold cross-validation and then tried on a separate test set whose size the paper does not give. For the SVM the hyperparameter search ran 30 iterations and settled on a cubic kernel. The angle differences between the able and not able groups are presented only as means, with no statistical test and no analysis of which variables carry most weight in the classifier.

For an archer the limitation is obvious: the task is not archery as anyone who shoots knows it, but a game with a toy bow at 3 metres, and the threshold of 10 out of 90 mostly separates people who hit the target now and then from people who hardly hit it at all. The angles come from a 2D side-on video, which cannot see rotations, and hand dominance was not considered even though everyone drew with the right hand.

The headline result, 80% on the test set, is fragile. All three initial models scored exactly 70% on the test set, and the per-class figures (precision 1.00 and recall 0.60 for not able, 0.71 and 1.00 for able) fit a test set of 10 people; this is our inference from the numbers, but if so, one error more or less moves the result by ten points, and the validation estimate of 72.3% is the more cautious figure. There are also some slips: women are given as 61.5% instead of 59.6%, and the target is at 3 metres in the text but about 5 in the description of the toy. The stated aim is assessing the hands-on skills of engineering students, not archery.

The study tells archers very little: it is an engineering education experiment that uses a toy bow as a motor task. The only observations close to shooting are descriptive: among complete beginners, those who hit the target more often held the drawing elbow more open at release, on average, and the trunk slightly more upright. There is no test of whether these differences go beyond chance, and no link to a real bow, a real draw weight or competition distances.

The idea that an algorithm can tell from a few body measurements whether someone can shoot also calls for caution: with an accuracy between 70 and 80% on such a simple task, and with sex among the variables, the classifier mainly separates groups that are very different from each other. In the registry the study counts as an example of low-cost video analysis applied to beginners, not as evidence on technique.