A Hybrid Sparse Graph Attention Framework for UAV Path Planning

Abstract

Uncrewed aerial vehicle (UAV) path planning is essential for applications, such as infrastructure inspection, environmental monitoring, and logistics distribution. A typical mission requires the UAV to depart from a base, visit designated targets, and return, forming a closed route. However, large-scale targets, high-dimensional environments, and strict timing constraints make it challenging to achieve both accuracy and computational efficiency. To address these issues, we propose sparse graph attention network for LKH (SGAT-LKH), a hybrid path planning framework that integrates a sparse graph attention encoder with the Lin–Kernighan–Helsgaun (LKH) solver. By using a sparse graph encoder with multihead attention to extract informative neighborhood features, SGAT-LKH generates a data-driven candidate edge set, which alleviates reliance on manual heuristics. Extensive UAV path planning simulations demonstrate the effectiveness of SGAT-LKH: on 2-D random TSP500 tasks, it achieves an average runtime of 12.8 s, about 70% faster than the LKH solver. On TSPLIB benchmarks, the average gap stays below 0.8%, and the optimality gap is only 0.085% on extended 3-D tasks. These results demonstrate SGAT-LKH’s ability to balance accuracy and efficiency, providing a scalable and practical solution for large-scale and time-critical UAV missions.