Bench notes.
Short writeups of things we measured. Each one is a number, the instrument it came off and what we did about it — mostly from work in the register, sometimes from a bench afternoon that went nowhere.
We moved the ballistic fit off the GPU and the loop got slower: 41 ms to 63 ms
Forty lines of math should not need an Orin, so we ported the trajectory fit to the STM32G4. The math ran in under a millisecond. The extra CAN round trip we added to feed it cost 22, and we reverted the same afternoon.
PETG creeps in direct sun: 0.6 mm of mast lean in four hours, on a dial indicator
Two PLA camera masts had already crept, so we reprinted in PETG and assumed that closed it. Readings every thirty minutes on a south-facing bench say otherwise. The fix was a section change, not a material change.
What a rolling shutter costs you at 8 m/s: 34 mm of centroid error across one frame
We shot a thrown object with a rolling and a global-shutter sensor side by side and measured how far the centroid walked between the first row and the last. That number is why every camera on the bin is global shutter.
Six hub motors on one dynamometer: torque constants ran 4 to 19 percent under datasheet
We bought six candidate BLDC hubs and measured Kt on a bench dyno before committing to a reduction ratio. Not one of them hit its published figure, and the worst was off by enough to change the gearbox.
200 gsm board tears at a 0.8 mm blade radius and folds clean at 1.2 mm
Station four of the folding cell tore roughly one part in nine on heavier stock. We ground three blade radii, ran sixty parts through each, and took the one that stopped tearing without softening the crease.
5 notes so far. They go up when the measurement is taken, which is why the gaps between them are uneven.