The open tennis-ball feeder aims every shot with two numbers: how much of the wheels' surface speed the ball leaves with, and how much of their speed difference turns into spin. Until today the only way to measure them was a radar gun and slow-motion video. Get the speed figure 5% wrong and a deep groundstroke lands about a metre short.
Earlier today the firmware learned to notice an empty hopper by how much the launcher wheels slow when a ball goes through. That drop is the ball's momentum. The two wheels' pushes add up to the ball's linear momentum, and their difference gives its spin. So with the wheel's inertia known, the drops tell you the exit speed and the spin.
What landed. Every feed stroke the firmware judges now leaves a record: each wheel's speed just before the ball, how far each one fell, and whether a ball went. The board sends it to the Pi with every telemetry frame until the next stroke. tennisbots run drill.calibration.json --shot-log shots.jsonl logs a 12-shot drill made for this, and tennisbots fit-launcher shots.jsonl prints the two efficiencies as a block the drill files take.
Two mistakes the simulator caught. The first version measured each drop from the wheel's target speed. But the feeder lets a wheel sit anywhere within 3% of target, and against an 8.8% drop that offset moved readings by up to 30%. The drop now comes from the wheel's own speed in the 2.5 revolutions before the ball. The second mistake was keeping that history as the last three readings. A slowed wheel's next pulse arrives late, and the tach reports a falling half-revolution on every 10 ms tick. Those readings pushed the pre-ball speed out of the history, and at 1,500 rpm a drop read a third of its size.
The tach misses part of each drop. It reports the slowest single revolution, and the speed loop is already pulling the wheel back during that revolution. On the simulator it saw 78% of the drop at 1,500 rpm with a fast loop and 97% at 6,500. The fit models that share from the revolution's length and the loop's recovery time, and fits the recovery time too when the log covers a range of speeds.
How close, on the simulator. Across 18 simulated builds with true efficiencies the planner didn't know, the fit got the speed efficiency within 3.9% and the spin efficiency within 11.7%. Reading the drops naively, without the recovery model, was off by up to 10.9%. Run through the Pi's real drill loop, three builds came within 2.3% to 4.4%.
What it can't do alone. The drops measure efficiency divided by the wheel's inertia, and the inertia is still the CAD's figure, not a weighed wheel. One radar session settles it: fit-launcher --radar-speed-eff turns that one reading into the true inertia. After that, every drill keeps the calibration current for free, as the wheels' felt wears or a softer batch of balls comes in.
None of this has met a real wheel yet. The simulator's wheels are simple first-order lags and each ball's momentum leaves in one instant, and a real ESC may recover differently. The first build's check is a calibration drill logged next to a radar gun. Details and every number: ventures/open-tennis-robots/firmware/esp32/README.md, "Shot log".