Why a crossing-path formula undercounts drone conflicts by 7%, and the one term that fixes it

A crossing-path formula misses drone conflicts that begin at take-off. Steiner's formula gives the missing πR² term; adding it closes a 7% gap to 0.25%.

The aerolog-scaling programme estimates how often delivery drones from several operators come within a separation radius R of each other in a shared layer. Its closed form (milestone 11d) counts pairs of straight legs whose paths cross and gives each crossing a window of departure offsets, 2R|w| / |v_a × v_b|. The previous step built an exact estimator and found the formula 7% low on average (3% to 12% across 24 hub layouts), yet with the right ranking of operator pairs. This step explains why.

Two ways a conflict can begin. Two straight legs come within R of each other at most once, because their separation is convex in time. That episode starts in one of two ways. Either the relative position crosses into the R-disc while both drones fly (flux), or one drone takes off, or leaves a delivery point for home, already inside the disc (a start). Both parts can be computed exactly for every pair of legs.

Steiner's formula settles the average. Fix the two velocities and durations and slide leg b over every position. For a fixed launch offset, the positions that produce a conflict form a segment of length |w|·overlap, thickened by R. Its area is 2R|w|·overlap + πR². Integrating over offsets gives 2R|w|D_aD_b + πR²(D_a + D_b). The crossing-window formula integrates to exactly the first term. So the closed form is the flux term, and it leaves out the πR² start term entirely. Truncation at leg ends and near misses between paths that don't cross cancel out on average. Relative to the flux term, the missing piece is (πR / 2|w|)(1/D_a + 1/D_b): it grows linearly with R and is larger for short legs and slow relative speeds.

On the 24 layouts (Manhattan pickup locations standing in for deliveries, 240 flights an hour, 400,000 sampled flight pairs each):

  • cross-operator pairs carry 99.5% of the deficit, and the start term accounts for 104% of it;
  • the deficit is 3.7%, 7.2% and 13.1% at R = 25, 50 and 100 m, as the linear-in-*R* prediction says;
  • adding the start chords brings the closed form from 0.928 of the exact count to 1.0025 (within 2% in all 24 layouts). It picks the same two-layer split as before in 23 of 24 layouts and a 0.4% better one in the last.

All six pre-registered predictions held. In practice, a planner using a crossing-path formula for drone (or any straight-line traffic) conflict rates should add a πR² term for every take-off and turnaround. At realistic separation minima it is not a rounding error.

The model has straight legs, Poisson launches, no avoidance and no wind, and covers one city. Code, tests and results: research/aerolog-scaling, milestone 11s.

All writing · More in Research findings · All tools and research