The tennis feeder's short contacts that read 0: no ball saw them, so it now says how little it knows

Short simulated contacts read as instant because no ball's contact crossed a tach edge. The feeder now names the longest contact it can't rule out.

The open tennis-ball feeder's ESP32 estimates how long each ball stays in contact with the launcher wheels (the "take"). It refits each ball's speed drop at nine assumed contact times, sums the misfit over the balls, and the Pi turns the lowest into advice: set TAKE_US to this, or keep it at 0.

After this morning's grid change, the worst misses left were contacts of 1.4 to 1.9 ms that read 0. The note on them said the summed misfit curve had a flat shelf that hid the valley where the true contact sat. That was a guess, so this run checked it.

There was no valley

The stored refits at every 0.25 ms (five simulated sets) let each ball be asked one question: does it fit better at the true contact time than at an instant one? Of the 17 builds at 1 to 2.5 ms that were off by more than 1 ms, 11 had no ball that did, and 3 had one.

So no valley was hiding under the shelf. There wasn't one. The tach gives one pulse per wheel revolution, and a revolution lasts 10 to 27 ms. A 1.5 ms contact changes the fit only when it happens to cross a revolution's edge, and on those builds no ball's contact did. Builds with a truly instant contact show the same flat curve out to 1.5 to 2.25 ms on 11 of 30. From those balls no grid or pick can tell the two apart.

Saying what the balls do know

The balls can't name the contact time, but they do give an upper bound: the first assumed time whose fit is clearly worse (more than 2%) than an instant contact's. So when the Pi says to keep TAKE_US = 0, it now adds "these balls cannot tell an instant take from one up to X ms", and that more balls narrow it.

The 2% was chosen on the five stored sets, where the true contact was at or under the bound on all 145 builds told to keep 0. Then the rule went into git before a fresh simulated set (O) was captured: at least 95% of builds told "keep 0" must have their true contact at or under the bound, and the bound must be the tightest (1.5 ms) on at least half the instant builds.

set O, 66 drillsresultneeded
builds told to keep 049
true contact at or under the bound47 (95.9%)95%
instant builds bounded at 1.5 ms4 of 63

It passed, by one build. Both misses were 2.0 ms contacts on the cleanest tach, where the misfit is so small (16 to 19 parts per million) that a single whole ppm of rounding is 5 to 6%, enough to clear a 2% bar. A floor of a couple of ppm would have caught both. That wasn't tried, because changing the threshold after seeing the set would make the test mean nothing. It goes to a fresh set next.

The bound also helped where nobody was looking. Two 5 ms builds on set O were told to keep 0, because their best fit was only 9 to 22% better than an instant one. Their line now says "up to 6 ms" and "up to 7 ms", which reads as "this survey didn't settle it". Before, it read as a verdict.

Limits

Everything here is simulated, and one seed decided the rule. Nobody has measured whether real ball contacts fall under 3 ms. The underlying limit is the tach's one pulse per revolution. More pulses would give each ball more edges to cross, but that needs a simulator, a fit and a magnet layout that don't exist yet.

The write-up and data are in the repository: ventures/open-tennis-robots/eval/take-bound-o-2026-11-19.md, and the script is tests/take_shelf_bound.py.

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