The open tennis-ball feeder learns how well its launcher wheels grip a ball from how much each wheel slows down as a ball goes through. Its firmware fits that drop from the tachometer's revolution times, and until now both the fit and the simulator it is tested on treated the ball's take as one instant. A real ball is not: 150 mm wheels grip it over about 50 mm while it speeds up to 18-27 m/s, so it is in contact for roughly 4 to 6 ms. At high speed a revolution takes only 7 ms.
The prediction that failed
The simulator can now spread a take over a contact time, with a symmetric (half-sine) force or one that peaks at first contact. Before running it, the prediction went into the script: a short take should look to the revolution times like a step at its centre, so the readings should barely move. The rule: if a take of 6 ms or less pushed the readings' scatter above 0.3%, the fit needed a contact-time term.
| take | reading scatter (sd) | spin efficiency error |
|---|---|---|
| instant | 0.12-0.14% | 0.05-0.15% |
| 3 ms, symmetric | 0.27-0.30% | 0.24-0.36% |
| 4 ms, front-loaded | 0.48-0.50% | 0.58-0.96% |
| 6 ms, symmetric | 0.62-0.82% | 1.25-1.47% |
It was wrong past 3 ms, and worst at high speed, where the take is a large share of a revolution (0.91% scatter at 4,500 rpm and up for the 6 ms take).
A term for the contact time
The fit now takes an optional contact time: the drop is split into four equal parts placed at the force profile's quantiles. The solve stays the same size. Refitting the saved inputs of 575 wheel readings offline:
| fit | 6 ms | 4 ms front | 3 ms |
|---|---|---|---|
| step (as shipped) | 0.98% | 0.41% | 0.21% |
| told the true contact time | 0.19% | 0.21% | 0.17% |
| a fixed 4 ms guess | 0.75% | 0.26% | 0.19% |
| picks its own (least misfit of 7 times) | 0.17% | 0.11% | 0.09% |
The first refit didn't fix the 6 ms case (0.88%). The fit searched for the ball's first contact only from the revolution where the slowdown shows, but a long take starts earlier. Widening the search by the take's length fixed that.
The last row is the useful one. A build doesn't know its ball's contact time, and a wrong fixed guess costs almost as much as ignoring it. But refitting at a few candidate times and keeping the one that misses least did as well as knowing the answer. That is how the board already finds its ESC's lag, so it can survey its contact time the same way on a calibration drill.
Rerunning the whole drills with the board told the 6 ms contact time, the readings went from 0.82% to 0.20% scatter (0.62% to 0.18% with a lagging ESC). The spin efficiency error fell from 1.47% to 0.39% (1.25% to 0.35%). That is still above the instant take's 0.05-0.15%, on one seed.
What this does not show
All of it is simulated. The force profiles are assumed shapes, and a real contact time likely shortens as the wheels speed up. The term ships off by default. A real wheel's tach log around a ball is what settles it.
Code and numbers: ventures/open-tennis-robots/tests/take_spread.py, tests/take_fit.py, and "A take that lasts a few milliseconds" in firmware/esp32/README.md.