Shade-Aware Pedestrian Routing Without Online Shadow Generation: How Much Model Do We Actually Need?

Abstract

Extreme heat is becoming a routine summer hazard, and the people most exposed are those who must walk: to transit, to work, to a cooling centre. Time in direct sun is the largest term a route can change in a pedestrian’s heat load, so shaded routing is an adaptation lever needing no new infrastructure. It does require knowing how sunlit each street segment is at the moment of travel, which today means running shadow geometry online. Existing systems route through an explicit shade image; we ask whether an offline geometric teacher can instead supervise a surrogate that maps scene and time straight to a scalar cost per edge. On a Toronto walking graph under a hard 10% detour cap it can: mean sun-exposed distance falls by about a fifth versus shortest paths and lands within a metre of the geometric teacher across 42,000 sealed-holdout queries. Three results temper this. A Transformer gives no advantage over a small multilayer perceptron on identical features, and the perceptron is the fastest of the three, well ahead of the geometry it replaces; entity tokenisation does not help; and at held-out times the learned models beat a nearest-time lookup but not a cache that interpolates between clock times. The regime where learning is genuinely needed therefore lies beyond dense, single-graph, fixed-schedule supervision — a boundary these negative results sharpen rather than undermine.