A Reinforcement-Learning-Driven Multi-Strategy Spherical-Vector Grey Wolf Optimizer for UAV 3D Path Planning

Unmanned aerial vehicles (UAVs) have been widely used in surveying and mapping, inspection, emergency rescue, and environmental monitoring. However, effective path planning remains a key challenge in complex three-dimensional terrain, where UAVs must simultaneously cope with terrain undulations, no-fly zones, safety-clearance requirements, and trajectory-smoothness constraints. In addition, conventional intelligent optimization algorithms often suffer from search instability and premature convergence. To address these challenges, this study proposes a reinforcement-learning-driven multi-strategy spherical-vector grey wolf optimizer, termed TLQ-SGWO, where TLQ denotes the combined use of Tent-Logistic hybrid initialization and Q-learning search-strategy scheduling. In the proposed method, candidate trajectories are encoded using spherical-vector increments; Tent-Logistic hybrid initialization is introduced to enhance population diversity; and Q-learning is incorporated to adaptively select search strategies, thereby dynamically balancing exploration and exploitation. A comprehensive cost function integrating path length, threat avoidance, terrain clearance, and trajectory smoothness is further constructed to improve the feasibility and safety of the planned trajectories. Experiments are conducted on the CEC2017 benchmark functions and artificially generated complex mountainous terrain scenarios. On the CEC2017 benchmark suite, TLQ-SGWO achieves the best average rankings in both mean error and standard deviation among the seven compared algorithms, indicating a stronger balance between optimization accuracy and robustness. In artificial mountainous scenarios, TLQ-SGWO obtains the lowest mean path cost in three of the four scenarios and remains statistically comparable to the strongest hybrid baseline in the remaining scenario, while maintaining stable feasible 3D trajectories under increasing no-fly-zone complexity.