Unmanned aerial vehicles offer fast, flexible last-mile and emergency delivery, but routing them is challenging when the number of clients changes. Classical approaches based on mixed-integer programming and heuristics are common, yet they can struggle as problem sizes scale or shift.
The paper, posted on arXiv, focuses on pointer networks trained with reinforcement learning for capacitated drone delivery routing. These networks often fail to generalize when trained on small instances and then applied to larger ones. The authors propose changing the network's decision mechanism "from comparison to selection," aiming to make the learned policy more robust across different problem scales.
The work is specific to the capacitated variant of the problem, where each drone has a payload limit. The contribution is a mechanism-level change rather than a new architecture, and the stated goal is enhanced cross-scale generalization for learned routing policies.