The tennis feeder's big take misses come from the fit, not the grid

A pre-registered test of a finer contact-time grid on three fresh simulated sets: 14 large misses against 12, so the grid stays. Most misses were shared.

The open tennis-ball feeder's board estimates how long each launcher wheel grips a ball (the "take") by fitting the wheel's speed drop at a few candidate contact times and picking the best. The shipped candidates are 0, 2, 3, 4, 5, 6 and 8 ms. An earlier run found that a nine-point grid (0 to 8 ms by 1) halved the builds read off by more than a millisecond, 10 against 20. That number came from tables the earlier rule did not name, so adopting the grid on it would have meant picking the measure after seeing the result.

Named first, then measured

This run wrote the measure and the rule down and committed them before any new data existed (pre-registration). The primary measure was builds off by more than 1 ms. The nine-point grid had to cut that by at least 4 over 162 simulated builds, because it costs two more fits a wheel and a protocol change. Earlier sets had placed every true take on the half-millisecond, exactly between grid points. This time one take was drawn at random from each millisecond, so neither grid got takes placed to suit it.

What came back

pooled over 3 fresh setsshippednine-point
builds off by more than 1 ms1412
mean absolute error0.53 ms0.45 ms
within 0.5 ms93121
instant builds wrongly advised a re-flash00

Two fewer, not four, so the shipped grid stays. The interesting part is why the gap closed. Ten of the large misses are the same builds on both grids, and all of them have takes under 2 ms: a 1.75 ms take read as 0, or an instant build read as 2.5 ms. In those cases the survey's residual curve bottoms out in the wrong place, and adding a candidate at 1 ms cannot fix a curve whose minimum is at 0. The next item is the fit's blind spot at short takes, not the grid.

The finer grid does help one thing. On the shipped grid every 7.45 ms take read as 8.0, its last point. That shows up as builds within 0.5 ms (93 against 121). The previous test scored on that same measure and tied, so it now needs a fresh set of its own, most cheaply as one added point at 7 ms.

Everything here is simulated, with one seed per set. The full tables, the rows and the commands are in the firmware README.

All writing · More in Tools and projects · More on Open Tennis Robots · All tools and research