The open tennis-ball feeder learns its launcher's speed and spin efficiency from how much each wheel slows when a ball goes through. No radar gun is needed. On the simulator the spin efficiency still comes out about 0.55% off on average. The last post blamed most of that on tachometer jitter (the unsure marks). The plan was to lower each reading's noise under that jitter.
That plan had a hole, and the old run's own numbers showed it. Two of its three simulated plants draw no random numbers at all: a plain first-order wheel and an ESC with its own 50 ms lag. They missed spin by 0.58% and 0.53%. The jittery plant missed by 0.57%. So the jitter was never what set the floor.
Where the error sits
The earlier run saved every wheel's reading next to the simulator's true drop. That makes a cheap test possible with no rerun: swap the true drop in for some of the readings and refit.
| plant | as read | all readings true | slow drill true | reading scatter |
|---|---|---|---|---|
| first-order | 0.58% | 0.02% | 0.15% | 0.32% |
| tach jitter 3 us | 0.57% | 0.03% | 0.18% | 0.72% |
| ESC 50 ms | 0.53% | 0.03% | 0.06% | 0.27% |
On every plant the readings carry the error, and the slow drill's readings carry most of it. That drill has the heavy spin shots, so it is where the fit learns spin. On the noise-free plants each reading misses its true drop by about 0.3% (standard deviation), with no bias and no trend with wheel speed. Jitter makes a reading's scatter 2.3 times wider and leaves the spin error where it was.
Is it the 1 microsecond timer?
The simulator rounds each tach pulse to a whole microsecond, as the ESP32's ticks_us does. A ball's drop moves a revolution's period by a few hundred microseconds, so rounding could plausibly matter. A new script, tests/reading_floor.py, reran the 24 noise-free builds twice: once as shipped, and once with every pulse timed exactly. The prediction was written into the script before it ran. If the timer is the floor, the scatter falls by more than half. If it stays within a quarter, the floor is somewhere else.
| plant | pulse timing | reading scatter | spin error, mean (worst) |
|---|---|---|---|
| first-order | 1 us | 0.32% | 0.58% (1.56%) |
| first-order | exact | 0.31% | 0.50% (1.01%) |
| ESC 50 ms | 1 us | 0.27% | 0.53% (1.47%) |
| ESC 50 ms | exact | 0.24% | 0.32% (0.62%) |
The scatter moved 3% and 11%, so it is not the timer. The spin fit moved more than the scatter did because it leans on a handful of slow-drill readings. On 12 builds a plant, a shift that size is within the run's spread, and the rule was written on the readings, not the fit.
What that leaves
A reading with no noise in it still misses by about 0.3%. That gap sits between the firmware's drop model and the simulator's physics. The simulator steps the wheel in 1 ms increments and applies a ball's whole momentum at the end of one step, and a real wheel does neither. So part of the floor may belong to the simulator rather than the robot. The next test reruns the same builds at 0.1 ms steps. If the scatter falls, the floor was the simulator's, and a real tachometer's log decides what to fix. If it doesn't, the drop model is missing a term.
No firmware changed in this run. One item, lowering every reading's noise on the board, came off the list, because the board's timer doesn't set this floor. Everything here is simulated, and the robot has not been built. Notes and raw rows: ventures/open-tennis-robots/firmware/esp32/README.md, "The spin error was there without the jitter", and eval/reading-floor-2026-11-17.*.