Application of Deep Reinforcement Learning in Obstacle Avoidance and Inspection Path Planning for UAVs in Complex Vegetation Environments

Traditional UAV path planning methods without energy consumption constraints often fail to meet the requirements of obstacle avoidance and inspection in complex vegetation environments. To address this challenge, an improved ε-greedy strategy is proposed, which effectively balances the efficiency trade-off between the exploration and exploitation phases. Within the framework of deep reinforcement learning based on Monte Carlo Tree Search (MCTS), the state and action spaces for task planning are constructed in stages, and a unified objective function is used to design the reward mechanism. This approach comprehensively considers the UAV’s maneuverability, multi-obstacle avoidance decision-making, and obstacle avoidance probability. Accordingly, optimization functions for obstacle target allocation and obstacle avoidance decision-making are established. By dynamically adjusting their weights, an efficient planning objective function is constructed, resulting in superior cumulative task rewards, data collection efficiency, and training stability throughout the training process.