Accelerating A Path Planning for Unmanned Aerial Vehicles in Complex Terrain with a Convolutional Neural Network*

Abstract

Unmanned aerial vehicles are increasingly used in complex field missions such as infrastructure inspection and terrain mapping. Path planning is a core technology for ensuring operational quality and efficiency. Complex terrain environments often contain no-fly zones and mountain obstacles. In such environments, the classical A* algorithm uses distance-based heuristics, which severely underestimate the true traversal cost and cause a large number of unnecessary node expansions. The problem is particularly pronounced when substantial detours are required, in which case planning can take tens of seconds or even minutes, seriously limiting practical usability. This paper proposes a convolutional neural network-guided A* path planning method. It accelerates planning in such scenarios and reduces both the mean planning time and its variability across problem instances. The network is trained to predict the remaining cost-to-go from candidate nodes to the goal based on the global grid map. During the search process, the maximum of the network prediction and the Euclidean distance is used as the heuristic, which approximates the true optimal cost-to-go more closely than the Euclidean-distance heuristic alone. This effectively reduces unnecessary node expansions. The method is evaluated on twenty 200×200×40 grid maps, giving forty test instances in two complex-terrain scenarios. Compared with the A* baseline with line-of-sight smoothing, the proposed method reduces the mean planning time by 59.5% and the cross-scenario standard deviation from 23.7s to 4.7s, with a path-length increase of only 0.4%.