Integrated Order Dispatching and Routing for Last-Mile Pickup via Deep Reinforcement Learning
The paper pairs a learned routing oracle with real-time dispatch heuristics for last-mile pickup decisions.
The authors argue dispatching and routing are too interdependent to optimize separately, but too costly to learn fully end to end at large scale. Their framework uses a Dynamic-Residual Graph Attention Network encoder and a look-ahead, courier-personalized decoder for routing. The dispatch side uses the routing oracle to pick candidate couriers, with local search to preserve real-time scalability. Tests on Cainiao Logistics datasets, including offline evaluation and rolling-horizon simulation, beat benchmarks on solution quality and solving time. ArXiv · AI/CL/LG's note
The authors argue dispatching and routing are too interdependent to optimize separately, but too costly to learn fully end to end at large scale. Their framework uses a Dynamic-Residual Graph Attention Network encoder and a look-ahead, courier-personalized decoder for routing. The dispatch side uses the routing oracle to pick candidate couriers, with local search to preserve real-time scalability. Tests on Cainiao Logistics datasets, including offline evaluation and rolling-horizon simulation, beat benchmarks on solution quality and solving time. ArXiv · AI/CL/LG's note
score 4