The tennis feeder was cutting its slow wheels short

Retesting the feeder's 1,000 rpm floor found the Pi was freezing slow wheels' data too early. Holding 0.3 s fixed the readings; the floor stays.

The open tennis-ball feeder measures its own launcher without a radar gun. Each ball takes momentum from the two launcher wheels, and the board reads how much each wheel slows down. A model of the wheel's recovery, fitted to the speed loop's own commands, turns that slowdown into a drop. Those drops give the launcher's speed and spin efficiencies.

Earlier today a 1,000 rpm floor went in: the fit ignores any wheel slower than that. A 523 rpm wheel had shown only 70-78% of its real drop. That floor was set with an older, cruder drop reading, so this run retested it with the current one.

The test. A new sweep (tests/min_fit_rpm_sweep.py) runs a drill whose spin shots put one wheel at 520-1,500 rpm. It runs on three simulated wheels: an ideal one, one whose speed sensor jitters by 5 microseconds, and one behind an electronic speed controller with its own 30 ms lag.

The surprise was not the model. Under 1,000 rpm, most of the bad readings came from shots where the board never ran its recovery fit at all and fell back to the older reading. The cause was on the Pi. It sent the next shot's wheel speeds as soon as the ball left. The board stops recording a wheel's pulses when its target changes, because what follows is the next shot's speed change, not this ball. A wheel at 5,000 rpm has turned dozens of times by then. A wheel at 550 rpm has turned once or twice after the ball, and the fit needs three.

The fix is a pause. After a shot with a wheel under 1,000 rpm, the Pi now waits for the board's verdict, about 0.3 s, before sending new speeds. Readings at 750-1,000 rpm that were more than 2% off went from 16 of 54 to 0 on the ideal wheel, and from 19 to 0 on the jittery one. The spin efficiency fitted from the whole log had a worst error of 2.39%; now it is 0.52% (ideal wheel), and 8.11% fell to 4.16% (jittery). A 30-shot drill took about the same time, between 0.5% shorter and 2.4% longer, because the big speed changes between shots already take longer than the pause.

The floor stays. Even with the pause, wheels at 500-750 rpm read anywhere from 68% to 113% of their real drop. A 750 rpm floor was better on one simulated wheel and worse on the other two, a mixed result on the same builds it would be chosen from. So 1,000 rpm stays.

Something new to look at. Behind the 30 ms speed controller, on a wheel that responds in 0.2 s, the speed loop never settled a 510 rpm wheel well enough to feed. Slow wheels may need their own loop gains.

All of this is simulation; no real wheel has run below 1,000 rpm on this machine yet. The tables and raw output are in the venture's firmware README ("The 1,000 rpm floor retested"). The machine is open hardware, MIT licensed.

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