aerolog-scaling asks how delivery-drone operators sharing a city should be stacked into altitude layers inside the 120 m (400 ft) column. Yesterday's step (11an) measured a drone barometer against RTK-fixed GNSS height in public PX4 logs. The noise it found (about 0.8 m) kept the stack at five two-level bands 13.3 m apart. It also found something the planner did not model: a speed-dependent offset. Air flowing past the airframe changes the static pressure the sensor reads, so the barometer reads high or low by an amount that grows with speed. In the 5 en-route flights where speed varied enough to fit it, the offset at 10 m/s was -4.70, -1.44, +0.81, +1.06 and +1.55 m.
Why this one is different
Every error the planner had seen so far was either noise that differs drone by drone, or a pressure field that nearby drones share and that cancels in their altitude gap. This offset is neither. It is set by the airframe and the cruise speed. Drones of one operator fly the same airframe at the same speed, so it cancels between them. Two operators with different airframes carry different offsets, and that difference is a constant bias in their altitude gap for the whole cruise. Between two operators drawing from those 5 flights, the difference has a standard deviation of 3.3 m, more than three times the noise.
So I expected it to widen the stack. I wrote five predictions and pushed them before running anything: at the measured size the planner would drop to 4 bands or fewer, a plan ignoring the offsets would cost at least 25% more losses of separation, and the offsets would cost at least 20% even when planned for.
What happened
Each of 12 or 16 operators drew an offset from the 5 flights, 4 draws on each of 48 New York layouts, and the planner from 11an re-chose its spacing and band labels with and without knowing the offsets.
- At the measured size, the stack holds. The planner that knows every operator's offset still picks 13.3 m (5 bands) in the median layout, in 80 and 87 of 96 layout-draws. Ignoring the offsets costs a median 3.4% (12 operators) and 1.8% (16), not 25%. The tail is long, though: one draw cost 73%.
- Knowing the spread is enough. A planner that knows only the distribution of offsets, not which operator has which, is within 2% to 3% of one that knows each.
- Calibrating half of it away removes the effect. At a half or a quarter of the measured offset, nothing changes in any layout.
- At twice the offset, the stack thins. The fits stop at about 10 m/s, but the layouts cruise at 12 to 28 m/s, and a speed-squared law would give more there. As a crude stand-in I doubled the offsets: the planner then picks 3 bands at 24 m, and ignoring the offsets costs a median 13% to 16%.
Three of the five predictions failed. The planner absorbed the offset by putting operators with similar offsets in the same or distant bands, and the 5.7 m of margin between a 13.3 m spacing and the 7.6 m (25 ft) separation minimum covers most of the rest.
What it means for a planner
At the speeds the logs cover, the five-band, 13.3 m stack from the earlier steps stands. At delivery cruise speeds the airframe offset, not the barometer's noise, is the term most likely to set the spacing. PX4 already has static-pressure compensation coefficients for this (EKF2_PCOEF_*), and on these layouts calibrating half the offset away is enough. A calibration requirement looks cheaper than widening every layer.
Limits
Five flights chosen by their uploaders, speed measured over the ground with wind not removed, offsets not extrapolated past 10 m/s, so "twice" is a stand-in, not a measurement. Operators draw offsets from those flights at random, so two can share one airframe's offset, which helps the planner. The logs do not say whether compensation was set. Preliminary.
The script, results and the updated paper draft (section 6.5) are in research/aerolog-scaling (experiments/hub_airspace/speed_offset.py).