Removing two moves: what else pushes drones into walls on the worst link.

The question

The previous two CITY cycles found that taking the speed-matching push (copying the speed of craft on the crossing street’s level) away from a craft that was about to hit a wall saved about half of those losses on the worst radio tier and nearly all on the better one. What sends the other half into walls? The candidate was the craft’s own avoidance of craft on the other level. Five predictions were registered before any re-flight.

The flight

The CITY campaign re-used the test crossing and the wider stack of lanes at exact heights, on 48 fresh layouts per setting. From each run it sampled wall losses and re-flew each from the same start, with every other craft flying live, five ways from up to 10 seconds before the hit: speed-matching removed (last cycle’s way); only the dodge away from other-level craft removed; both removed; and two checks that keep the moves but add a sideways nudge along the lane of the dodge’s size. A loss counts as saved if the craft is not lost by any cause in the 10 seconds after the original hit. 1,530 losses were re-flown five ways (7,714 re-flights with 64 null checks identical to the original), no errors, no simulator change.

What happened

Removing only the dodge saved about four in ten wall losses on the worst tier (38 and 41 per cent, crossing and wider stack), and still a quarter to a third on the better tier, where the plan had bet on no effect. The same-size nudge saved 2 to 7 per cent, so it is the direction of that dodge, not just any push. With both moves removed, 97 and 96 per cent of worst-tier wall losses were saved; of the losses speed-matching removal alone leaves, 93 to 94 per cent, against 14 to 18 per cent for the nudge. The saves sit where the dodge carried the bigger push toward the wall.

The previous result held on three of four settings; on the crossing it came out a little higher (53 against 48 per cent), just outside the range registered, so that check reads as not replicated. Cost: without the dodge the saved craft passed about 9 to 10 m closer to other-level craft, with no contact loss of its own in the next 10 seconds, and other craft’s losses in that window were not clearly more than with speed-matching alone removed. Each saved craft flew about three minutes longer, and over the rest of the run about one more other craft was lost for every two saved; per minute flown that rate is not clearly different. One of five predictions held; the misses all say the dodge mattered more, and more widely, than bet.

The next question

Nothing is adopted: no setting was changed and nothing is switched on. Whether to keep speed-matching with the crossing level as the default stays the human’s decision. A new question for the human is whether to test removing the sideways dodge from other-level craft as a candidate setting, off by default and on fresh layouts; any such setting needs the own-height error axis.