A bi-level priority sorting framework for flexible AGV service scheduling in smart warehouses

TL;DR

This paper introduces a bi-level optimization framework for AGV scheduling in smart warehouses, balancing order fulfillment, cost, and efficiency through real-time routing and heuristic rules.

Contribution

It presents a novel bi-level optimization approach integrating heuristic rules and reinforcement learning for flexible AGV operations in smart warehouses.

Findings

Reduces average order delay and system costs by over 50% during peak demand.

Maintains service level above 90% and maximizes AGV utilization.

Demonstrates robustness and flexibility across diverse scenarios.

Abstract

This paper proposes a bi-level optimization framework to coordinate Automated Guided Vehicle (AGV) flexible operations in smart independent warehouses, addressing the critical challenge of balancing high-throughput order fulfillment with stringent cost control. The framework is designed to simultaneously optimize flexible customer service level, system cost, and operational efficiency. The first level dynamically adjusts real-time scheduling parameters, such as order commitment times and delay tolerance, based on predefined customer priority categories. The second level performs real-time routing optimization for each AGV by identifying the shortest feasible paths while avoiding conflicts. For complex multi-capacity package picking tasks, two heuristic rules, priority, deadline, with shortest path (PDSP) and delay cost with shortest path (DCSP), are applied to multi-capacity package picking tasks and further training is carried out using the reinforcement learning algorithm of A* guided deep Q-learning (AGDQN). Comprehensive simulation experiments, conducted across diverse warehouse layouts and order demand patterns, demonstrate that the proposed framework equipped with both heuristic rules consistently reduces average order delay and total system costs by over 50% during peak demand periods. This is achieved while maintaining a service level above 90% and maximizing AGV utilization. The method also exhibits superior flexibility and sustained efficiency under normal and fluctuating demand scenarios. Additional ablation studies confirm that the proposed priority sorting mechanism delivers robust performance advantages when tested with various other reinforcement learning baselines.