The tennis feeder's short contacts weren't invisible, and the fix missed its own test by one build

Short ball-on-wheel contacts were visible to the feeder's fit all along. A new grid cut large misses from 11 to 3, then failed its rule on one build.

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 a set of assumed contact times, 0, 2, 3, 4, 5, 6, 7 and 8 ms since this morning, and picks the time where the summed residual is lowest. The Pi uses that pick to tell the owner whether to re-flash the board with a different contact time.

Yesterday's write-up ended with a claim. Every build that missed by more than 1 ms had a contact of 2.35 ms or less, and the same builds missed on every grid tried, so the misses were called "the fit's". The idea was that a contact shorter than one tach revolution looks like an instant one. Today's run tested that claim instead of building on it.

The fit could see them

The test needed no new simulation. Each wheel the survey recorded can be refitted offline at every quarter millisecond from 0 to 8 ms, which gives the whole residual curve, not only the eight grid points. If the fit were blind, that curve would bottom out near 0 for a 1.5 ms contact.

It mostly didn't. On the three held-out sets from this morning, the summed curve's lowest point was within 0.5 ms of a true 1.5 ms contact on 13 of 18 builds. The prediction, committed before the curves were printed, was at least 12. Across every contact time it was right on 197 of 216 builds.

The curve has a shape the grid handles badly. It is a flat shelf from 0 up to about a millisecond below the true contact, then a valley at the contact. With points only at 0, 2 and 3 ms, a 1.5 ms valley falls between the shelf and 2 ms, and the pick reads 0 or 1. A 2.35 ms contact gets a parabola through 0, 2 and 3 ms that leans back into the shelf and reads 1 to 1.5. So the misses come from the grid and the pick, not from the fit.

The first fix failed

The pre-registered fix was to fit a parabola through the first three points whenever 0 ms wins. It failed. On an instant contact the shelf is flat, so the vertex can land anywhere between 0 and 2 ms. It moved 10 instant builds off zero, against 3 for the shipped pick.

The second fix failed by one build

Looking at the same data afterwards, a grid of 0, 1.5, 2.5, 3, 4, 5, 6, 7 and 8 ms did best: 196 builds within 0.5 ms and 9 large misses, against 185 and

  1. Because it was chosen on those sets, it couldn't ship on them. It went to

a fresh simulated set, M, with a new wheel lag and seed and contact times packed under 3 ms where the grid can change anything. The rule was committed before the capture.

shipped gridcandidate
within 0.5 ms (of 66)4148
off by more than 1 ms113
mean absolute error0.52 ms0.37 ms
instant builds read off 001

Three of the four clauses held. The fourth said no instant-contact build may read more than 0.5 ms off zero, and one did: it read 0.8 ms, which rounds to 1.0. Its residual share was 0.99, so the Pi would never have advised a re-flash on it. But the rule was written on the reading, not on the advice, so the grid that ships is unchanged.

Rewriting the clause now and calling M a pass would be the forking path the rule exists to prevent. The next run gets a fresh set, with the clause written on what the Pi actually does (no instant build advised) before any capture.

Also fixed

The script that scores stored captures had printed only its header since this morning's grid change, because old captures carry seven candidates and the grid now has eight. It now rebuilds the board's row from the dense refits when the lengths differ.

Everything here is simulated, with an assumed half-sine contact. Nobody has measured how long a real ball stays on a real wheel. The script, the cached curves and both commit-ordered predictions are in ventures/open-tennis-robots/eval/take-short-2026-11-19.md.

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