The open tennis-ball feeder measures each ball's momentum from how much its two launcher wheels slow down. That reading depends on a model of how the wheel follows the speed loop's commands. Some ESCs run a speed loop of their own, which adds a second lag the model leaves out. When that happens, the drops read low and the spin estimate from a calibration drill can be 13-18% off on the simulator.
An earlier run showed that the build's own balls can pick that lag. You refit each ball at several candidate lags and keep the one with the least summed residual. The catch was that the data needed for this lives on the ESP32, and the board never sent it. So the setting (ESC_TAU_US) could only be guessed, and a wrong guess was worse than leaving it at zero.
The board now runs the survey itself. Once a ball's record has gone to the Pi, the board refits each wheel at six lags (0 to 400 ms), one small step per control tick, and keeps a running sum of the residuals. It then fits a parabola through the best grid point and its two neighbours. A new serial frame carries the result to the Pi, and the drill runner prints one line, such as "set ESC_TAU_US = 77000". On the simulator, with a rally and a calibration drill as input (54 wheel readings):
- builds with no ESC lag picked 0 every time;
- ESC lags of 30, 60, 80 and 150 ms were read as 27-28, 54-57, 75-79 and 137-141 ms;
- with the picked lag set, the calibration drill's worst spin error fell from 13-18% to 0.6-1.0%.
One build missed: a slow wheel with 3 microseconds of tach jitter and a 100 ms ESC lag picked 0.
The other idea didn't work. Tach jitter moves each pulse's time, so two neighbouring revolution periods share the same error with opposite signs. Fitting the drop to the pulse times instead of the periods is the textbook-correct weighting for that. It was built and measured on 192 wheels per jitter level, and at 3 to 20 microseconds it spread the drop more, not less (0.86% against 1.03% at 5 us). The nonlinear search over the ball's moment and the wheel's lag seems to rank candidates worse that way. The option stays in the code, off by default.
All of this is simulated. Nobody has built the robot yet, and how long six extra fits per ball take on a real ESP32 has not been measured. The details and tables are in the venture's firmware README.