Research on the application of an improved CNN-guided A* algorithm in mobile robot path planning

Abstract To address the low search efficiency, poor path smoothness, and limited adaptability of the conventional A* algorithm in global path planning for mobile robots, this paper proposes a CNN-guided improved A* approach. First, a convolutional neural network (CNN) is embedded into the heuristic function to estimate the probability that a node lies on the optimal path, thereby providing learning-based guidance for node expansion. Second, to reduce redundant expansions caused by fixed-neighborhood exploration, a 32-neighborhood hybrid search strategy is introduced by combining rectangular boundary constraints with azimuth-angle pruning. In addition, an obstacle-density-adaptive threshold pruning mechanism is incorporated to further suppress invalid node expansions. Finally, a quadratic Bézier-curve-based smoothing method is applied to enhance trajectory continuity and smoothness. Simulation results demonstrate that the proposed algorithm consistently outperforms the conventional A* algorithm in path length, computation time, number of expanded nodes, and path smoothness. Comparative experiments on five map categories-Channel-type, Rectangular-type, Simple Maze-type, Complex Maze-type, and Sawtooth-type-show that, relative to the conventional A* algorithm, the proposed method reduces search time by 13.6%-36.3%, shortens path length by 1.2%-5.3%, decreases expanded nodes by 37.9%-76.5%, and reduces turning points by 16.7%-71.4% across different scenarios.