Airport shuttle vehicle path planning method combining graph neural network and reinforcement learning

To address the low computational efficiency and unstable solution quality of traditional heuristic methods in large-scale, complex pickup and drop-off scenarios, this paper proposes a path planning approach that integrates graph neural networks with reinforcement learning. The Airport Pickup and Drop-off Vehicle Routing Problem (APD-VRP) is formulated as a Markov Decision Process (MDP), enabling end-to-end learning of routing policies through a graph attention-based encoder and a sequence-generating decoder. To enhance the model’s capacity to capture sequential dependencies and local patterns, we extend the graph attention encoder with a positional encoding module and a one-dimensional convolutional layer, improving its joint perception of global and local features. The policy is trained using a policy gradient method to achieve efficient path planning in complex environments. Experimental results demonstrate that the proposed method significantly outperforms the unenhanced baseline model in terms of both route cost and path length.