Dynamic storage space optimization for multi-AGV systems: a multi-task proximal policy optimization approach

Multi-Automatic Guided Vehicle (AGVs) are core components of intelligent warehousing and logistics. In current deployments, storage locations are usually fixed, which decouples storage from vehicle scheduling and causes inefficient routing, congestion, and poor space utilization. The resulting coupled storage scheduling problem is NP-hard and challenging for classical optimization and heuristic methods.This study proposes a deep reinforcement learning framework based on Multi-Task Proximal Policy Optimization (MTPPO) that jointly optimizes dynamic storage allocation and AGV dispatching in an end-to-end manner. The framework decomposes the joint decision into two coordinated subtasks, AGV assignment and storage location selection, and employs a hybrid CNN–GNN architecture to fuse local congestion patterns with global topological relationships in the warehouse. An adaptive action-masking mechanism and a Monte Carlo trajectory-level reward reconstruction scheme are introduced to enforce feasibility constraints and stabilize training.Simulation studies on multi-AGV warehousing scenarios with varying grid sizes, obstacle densities, and task loads show that MTPPO shortens task completion time by about 10% and significantly reduces waiting-time variance compared with rule-based and metaheuristic baselines. In the storage dimension, the learned policy reduces actual transport and waiting time by an average of 8.7% relative to the best shortest-distance strategy, with gains rising to 11–18% under high-density obstacle layouts. These results demonstrate that jointly learning storage allocation and AGV scheduling yields more efficient, stable, and robust operation than optimizing each component in isolation. Also, the policy can scale to larger layouts and remain robust to moderate temporal noise. The proposed MTPPO framework thus provides a practical and scalable solution for dynamic storage space optimization in multi-AGV systems.