Learning to Retrieve Containers: A Scale-Diverse Deep Reinforcement Learning Approach for the Container Retrieval Problem

Woo-Jin Shin, Inguk Choi, Sang-Hyun Cho, Hyun-Jung Kim

Abstract:

This study addresses the container retrieval problem (CRP), a key challenge in the storage yards of automated container terminals where operational efficiency directly affects vessel turnaround time and yard congestion. In storage yards, containers are stacked vertically to maximize space utilization; however, accessing one located below others requires relocating the blocking containers, leading to additional crane movements and delays. The CRP involves retrieving containers from multiple bays in a specified order while minimizing the total working time of the yard crane, with relocation position decisions being critical. The CRP poses several practical challenges: despite
being NP-hard, real-world instances often involve hundreds of containers, requiring high-quality solutions in real time; yard configurations also vary widely and change frequently, demanding methods that adapt effectively to arbitrary layouts. We propose a novel deep reinforcement learning approach incorporating (1) a size-agnostic network architecture, enabling a single trained network to handle diverse yard configurations, and (2) a scale-diverse learning framework, which trains on a various yard scales using a normalized loss to improve generalization and scalability. Experiments on well-known benchmarks with several hundred containers show that the proposed
method substantially outperforms existing baselines across a wide range of yard sizes. It also scales to instances with thousands of containers and maintains strong performance in dynamic settings where retrieval orders are revealed online. Solutions are produced within a second for realistic instances, confirming its effectiveness and practical applicability in real-world automated container terminals.