DRL-Based Adaptive Dynamic Window Approach for Path Planning of UAV in Dynamic Environments

The path planning and obstacle avoidance in dy-namic environments are crucial challenges for UAV autonomous navigation. While the dynamic window approach (DWA) is widely used for such tasks, it suffers from poor robustness and a high dependency on parameter tuning. To overcome these limitations, we propose DBADWA, an adaptive algorithm that integrates deep reinforcement learning with DWA to dynamically predict optimal weighting parameters, enhancing local navigation. In DBADWA, the action space is restructured using key parameters of DWA, including target heading deviation, distance to the nearest obstacle, and velocity. To address the issue of sparse rewards, we enhance the reward function of the advantage actor-critic model by incorporating the distance to obstacles alongside DWA's metrics, such as distance, velocity, and heading. Exper-imental results across various dynamic scenarios demonstrate that DBADWA significantly reduces path length, shortens arrival times, and improves adaptability, outperforming existing methods in complex environments.