Bow motion during the shot: Hasegawa 2016 study

A single experienced archer shot 109 arrows at 25 metres with a motion sensor fixed to the bow, recording angular velocity about three axes a thousand times a second. After each arrow, before knowing the score, the archer rated the shot on a scale from 1 to 7 and commented on it in words. The authors, researchers at the Tokyo Institute of Technology, wanted to see whether a statistical model (a two-layer hidden Markov model) could recognise the phases of the shot by itself and tell successful shots from unsuccessful ones.

On the whole shot the model separated the large phases, such as set-up, aiming and release, but missed the fine differences in the release: the parts with a lot of movement took up most of the model's states, while aiming, with little movement, stayed a single block. The authors therefore took only the 2 seconds after release in 12 arrows on which the archer had commented on the bow arm, six judged with the bow arm holding its position and six without. In the good shots the bow's angular velocity rose steadily, changing by about 40 degrees per second; in the shots judged poor the change was smaller (20 degrees per second) or larger (60) and died away sooner. One sequence of states appeared only in four good shots, another in two good and two poor ones, and a third only in shots judged poor. In the video at 300 frames per second the difference between the two groups was not clearly visible.

The sensor was a nine-axis module made by Logical Product, mounted on the bow; the target was a 30-centimetre square at 25 metres, and two cameras at 300 frames per second filmed the archer from the front and from behind. The rating ran from very bad to very good in seven steps. Beyond calling the archer experienced, the paper gives neither the bow, the division nor the archer's precise level.

The model splits the signal into windows of 50 samples overlapping by half, builds small Markov models for each and links them into a graph of states, which becomes the representation of the shot (25 states in the second layer for the whole shot). For the release analysis the window starts when the vibration after release begins and covers 2 seconds. The split into good and poor shots comes from the archer's comments on the bow arm; score and rating, the authors write, were used for reference only.

This is a short conference paper, four pages long, presented at a robotics meeting without peer review and based on one archer. The sequences of states are compared over 12 shots with no statistics, and the split into good and poor shots depends on the archer's own judgement: the work shows that in some cases the sensor signal agrees with that judgement, not that the sensor recognises a better release. The score barely differs between the two groups (four 10s, three of them Xs, and two 9s among the good shots; two 10s and four 9s among the poor ones), and two shots judged as having a good bow arm received the same rating, 4, as a shot in which the bow arm had slackened.

The authors themselves note that some shots the data made look successful had been judged poor by the archer, and that the model ignores how long the shot stays in each state. The English summary refers to an angular acceleration sensor, while the text describes an angular velocity sensor. There is no test on other archers or on a second day.

For archers the interest lies in the method more than the result: a light sensor on the bow and a model that learns one archer's movement sequences might one day show at which point in the shot an arrow departed from the usual routine, comparing each shot with that archer's own shots rather than with an external model. The shots this archer judged as having a bow arm that held its position showed a steadier bow movement after release, neither too small nor too large.

With twelve shots from a single archer, and no clear difference in score, this remains a first technical step: it does not say how the bow should move after release, nor that the sensor detects anything the archer does not already feel. The authors present it as groundwork for a future feedback tool for training.