RL-ANA*: A Reinforcement-Learning-Guided Adaptive-Neighborhood A* Algorithm for Mobile Robot Path Planning in Complex Environments

Abstract Mobile-robot path planning balances search efficiency and path quality, while fixed neighborhoods and hand-crafted switching rules can limit adaptability. To address these issues, we propose a reinforcement-learning-guided adaptive-neighborhood A* algorithm (RL-ANA*), in which a deep Q-network (DQN) dynamically switches among three neighborhood tiers {8, 16, 24} based on a 23-dimensional spatial-perception feature vector, combined with a position-aware dynamically weighted heuristic that balances search aggressiveness and path quality. A five-stage curriculum-learning strategy spanning 30×30 to 128×128 grids supports policy learning across map scales. Experimental results demonstrate that, compared with standard A*, the proposed method reduces expanded nodes by 48.1%, 40.5%, and 59.7% on 50×50, 64×64, and 128×128 maps, respectively (all p < 0.001), with smaller observed standard deviations and mean LOS-smoothed path-length increases of 1.1%–4.7%. Sequential configuration comparisons show an additional node reduction of up to 11.5% of the Basic A* mean node count relative to the hand-crafted switching control. Cross-topology generalization tests on unknown maps achieve node-reduction rates of 17.8%–56.5% relative to Basic A*, supporting transferability under the tested conditions. Feature-ablation and parameter-sensitivity experiments further examine the state representation and reward settings. Physical deployment under ROS evaluates seven start–goal configurations on each of two maps, each repeated five times per planner. Both planners completed all 70 trials per planner successfully, with no observed collision. RL-ANA* reduced mean planning time by 13.3% and 20.1% on the two maps, whereas a matched MATLAB benchmark showed higher runtime owing mainly to DQN inference and feature extraction. These results demonstrate that data-driven adaptive-neighborhood strategies can reduce search effort in the tested mobile-robot navigation tasks, while runtime benefits remain implementation- and scenario-dependent.