Task-driven dynamic lane-level path planning for intelligent transfer vehicles in container terminals

Intelligent transfer vehicles in container terminals operate in dense and dynamic traffic environments, where conflicts at intersections and handling lanes can cause arrival-time deviations and potential deadlocks. This study addresses the dynamic lane-level path planning problem by proposing a hybrid multi-agent reinforcement learning framework under a centralized training and decentralized execution paradigm. The framework uses a lane-time grid with semaphore-based control to model space-time occupancy and incorporates a task-driven conflict resolution mechanism that accounts for task type, remaining time, and load status, with the objective of minimizing the sum of deviations between actual and expected vehicle arrival times. To improve learning efficiency and solution feasibility, an A* search planner provides warm-start trajectories, while a graph attention encoder captures structured traffic states. We further introduce TaskNet, a learned arbitration module that dynamically determines vehicle priorities in contested situations based on vehicle states and historical rewards, thereby enabling rapid conflict resolution in dense traffic environments. Extensive experiments on a realistic terminal layout show that the proposed algorithm significantly reduces path conflicts and task delays relative to benchmark methods, while maintaining sub-second decision times suitable for real-time deployment. The results demonstrate the value of combining structure-aware learning and task-driven coordination for scalable, real-time intelligent transfer vehicle scheduling in complex container terminal environments.