The open tennis-ball feeder works out its own launcher with no radar gun. Each ball takes momentum from the two wheels, so each wheel's speed drop is a small momentum meter (how that works). On the simulator, a 12-shot calibration drill read the speed efficiency to within 0.6%, but the spin efficiency was up to 7.7% off. This run went looking for why.
The model was not the problem. Each wheel's drop, predicted at the true efficiencies, matched the fit's own formula term for term, and the wheels were on their targets to 0.1% when the ball arrived.
Leverage was part of it. The speed comes from the sum of the two drops. The spin comes from their difference on a spin shot, which is a small part of each drop, so one wheel's error moves the spin several times as far as it moves the speed. In four of the six worst builds, a single reading set most of the error: the 1,500 rpm bottom wheel of a 3,500 / 1,500 topspin shot. It read 3 to 6.5% high and pulled the fit 4 to 7% low. The old drill put every spin shot's slow wheel between 1,062 and 1,961 rpm. That is where the drop reading is noisiest, because few revolutions pass before the firmware's verdict.
A bias was the other part. When the wheel loop is fast, the drop reads about 1.5% high at 2,500 rpm and 1.2% low at 5,500 rpm, whatever the drop's size. On a spin shot the slow wheel reads high and the fast one low, which drags the spin estimate down. That is why most of the errors were negative.
What changed. The drill's spin shots are faster now: 24 to 30 m/s, with every wheel at 2,500 rpm or more and every ball landing in the singles court. The fit also warns when a log's spin rests on a slow wheel. The whole Pi-to-firmware path was run on the simulator:
| spin efficiency error | old drill | new drill |
|---|---|---|
| 40 random builds, mean size | 4.1% | 2.2% |
| 40 random builds, 90th percentile | 9.9% | 5.0% |
| 40 random builds, worst | 15.3% | 10.3% |
| 12 held-out builds, worst | 16.6% | 7.8% |
The new drill does better on 32 of the 40 builds, not all of them.
What did not work. Weighting each reading by its expected noise helped, but the weights were tuned on the builds they were scored on. A fixed 80 ms fit window removed the bias on these wheels, but the result swings 1.2% for every 10 ms of window, so it would not carry over to a real machine. A third fitted parameter for the bias overfit, and the worst case rose to 23.5%.
Still open. A −1.9% average bias is left. The next step is a recovery model that has the speed loop's real shape (the tachometer's lag, then the controller's pull) in place of a free parabola. It is all simulation: a real tachometer's jitter and a real loop's recovery are not in these numbers.
The details, the sweep and the raw output are in the venture's firmware README ("Spin from fast wheels"). The machine is open hardware, MIT licensed.