Unmanned aerial vehicles (UAVs) enable fast and flexible last‑mile and emergency delivery, yet routing becomes challenging as the number of clients varies. Classical mixed‑integer programming and heuristic methods scale poorly or require problem‑specific design, motivating learning‑based alternatives. Reinforcement learning is especially promising if a single policy can generalize across problem sizes without retraining. Although attention‑based Pointer Networks are inherently size‑agnostic, their solution quality degrades when scaling up or down.
We use the Capacitated Vehicle Routing Problem (CVRP) as a controlled proxy, fixing the capacity $Q$ to reflect drone payload limits, and identify three factors behind the degradation: shifts in node density, dispersion of attention, and dilution of the mean‑pooled graph context. We systematically explore modifications at the architecture level, encoder side, and decoder side, and propose a learned, permutation‑invariant graph‑context aggregation mechanism implemented as a weighted average and as Set2Set.
Results reveal distinct requirements for each transfer direction: entropy‑based attention scaling reduces the CVRP20→CVRP100 performance gap by 1.41 percentage points, while the proposed aggregation methods achieve the largest gain on the more challenging CVRP100→CVRP20 transfer, about 0.9 percentage points. Exploiting this complementarity, we introduce an approach selector that chooses the best model for each train‑test size pair. The selector lowers cost relative to the reference model across all nine seen train‑test pairs and eight of nine pairs involving unseen sizes (CVRP35, CVRP75, and CVRP200), achieving improvements of up to 3.2%.
Review: The paper provides a thorough analysis of why Pointer Networks lose performance across scales and offers concrete architectural tweaks and a model‑selection strategy that together deliver measurable gains, making the approach highly relevant for scalable drone delivery routing.