Optimized Path Planning and Safe Obstacle Avoidance for UAVs Based on Enhanced DQN

This article introduces an enhanced Deep Q-Network (DQN) approach, termed the Double Annealing Cosine Q-Network (DACQN), for unmanned aerial vehicle (UAV) path planning and obstacle avoidance. Unlike conventional algorithms that prioritize single objectives, such as path length, while neglecting factors like energy consumption or altitude constraints, DACQN employs a comprehensive cost function that integrates flight distance, obstacle risks, energy expenditure, and altitude constraints. A tailored reward function guides the DACQN model’s learning, producing optimized, practical flight paths. To address Q-value overestimation in traditional DQN, DACQN incorporates a Double DQN mechanism, utilizing separate neural networks for action selection and Q -value estimation to reduce bias and enhance policy stability. Additionally, a Cosine Annealing learning rate scheduler mitigates the risk of local optima entrapment, improving convergence efficiency. In virtual simulations, DACQN outperforms traditional DQN, achieving faster convergence, greater stability, and enhanced adaptability in complex environments.