The tennis feeder's unsure marks find the noisy wheels, and weighting them down still doesn't help

The feeder firmware marks wheel readings it is unsure of. On 72 simulated builds the marks found the noisy wheels, but down-weighting them gained nothing.

The open tennis-ball feeder works out its launcher's efficiencies from how much each wheel slows when a ball goes through. The ESP32 board fits each wheel's drop from its tachometer pulses. When the fit's own misses leave the drop uncertain (a standard error over 1% of it), it sets an "unsure" bit in the shot record. Until now the Raspberry Pi only counted those marks and warned when more than 10% of a drill's readings carried one.

The last post found where a jittery tachometer hurts most: the fast wheel of a slow spin shot, whose readings scatter 1.0% around the true drop against 0.6% on a fast shot. Two questions followed. Does the board's mark land on those wheels? And would the Pi's fit improve if it counted a marked reading for less?

The run

tests/unsure_weight.py ran a calibration drill and a slow drill on 18 simulated builds (three launchers, two wheel lags, three kinds of plant) on four seeds: 72 builds and 5,616 wheel readings. Before any output existed, I committed a rule with the script. Pick the best down-weight (0.5, 0.25 or 0) on two design seeds. Ship it only if it also beats the shipped weight of 1 on two held-out seeds.

The marks work

On a clean tach and on a laggy ESC the board marked nothing. On the 3 µs jittery tach it marked 41% of the slow shots' fast wheels and 4% of the fast shots' wheels. Of the 23 readings more than 2% off, 16 carried a mark. A marked reading was off by more than 2% 4.5% of the time, an unmarked one 0.55%. The board's standard error tracks the true error too: readings it rated under 0.5% were 0.33% off (rms), and readings rated 1-2% were 0.98% off.

The weighting doesn't

seedsweight 1 (shipped)0.50.250 (dropped)
design0.58%0.57%0.58%0.58%
held out0.54%0.55%0.56%0.59%

That is the mean spin efficiency error over all builds. The design seeds favoured 0.5 by a hundredth of a point, and the held-out seeds reversed it. The rule kept 1.

Why

The simulator knows each wheel's true drop, so the fit can be rerun with some readings swapped for the truth. On the jittery builds the spin error was 0.57% as read and 0.03% with every reading true, so reading noise is nearly all of it. With only the marked readings made true it was still 0.44%, and with only the 23 worst readings made true, 0.52%. The error is a little noise on every reading, not a few bad ones. A filter on the bad ones can't remove it, and down-weighting them also throws away the signal they still carry.

So the mark stays a warning about the build. It is not a filter. The useful next step is lowering every reading's noise under jitter.

What this doesn't show

All of it is the simulator's Gaussian jitter of 3 µs per pulse. A real tachometer may fail in lumps, such as a missed edge or a late interrupt, and there the marked readings could carry more of the error. The details and every number are in the firmware README's "Unsure marks on the noisy wheels" section in ventures/open-tennis-robots/.

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