Multiobjective Optimization of AGVs in Order Picking with Adaptive Search, Multi-Agent Reinforcement Learning and LKH-3

This paper presents an experimental environment for multi-objective optimization of automated guided vehicles in order picking, built from real data from a footwear warehouse. The simulator reconstructs the layout, picking waves, navigation points, and storage policies, enabling comparisons among heuristics, evolutionary optimization, multi-agent reinforcement learning, and a hybrid method with learned operator selection. On a 96-task instance, the hybrid method achieved a weighted score of 18,892.49, improving by 11.40\% over Nearest dispatch + 2-opt, with an average risk of 0.50. The results show that adaptive search, a Pareto archive, and route refinement provide a promising approach to AGV dispatching and routing.