SaferSAC: A Deep Reinforcement Learning Framework for Autonomous Obstacle Avoidance Navigation in UAVs

Abstract

Autonomous obstacle avoidance and navigation for Unmanned Aerial Vehicles (UAVs) in dynamic environments presents a significant challenge in the field of autonomous planning. Traditional sampling- and optimization-based methods typically rely on global maps and necessitate frequent replanning, rendering them often ill-suited for real-time applications in dynamic scenarios. While deep reinforcement learning (DRL) approaches offer end-to-end perception-decision mapping, they often suffer from limitations in three-dimensional (3D) perception accuracy, sample efficiency, and action space adaptability. To address these challenges, this paper proposes SaferSAC, a DRL framework specifically tailored for robust UAV obstacle avoidance and navigation. First, we design a depth-semantic fusion-based 3D obstacle detection module that achieves precise spatial awareness by jointly processing depth images and semantic segmentation results. Second, we introduce a four-buffer prioritized experience replay mechanism that differentially stores and samples experiences based on distinct categories (e.g., success, obstacle, and precision), thereby significantly enhancing sample efficiency. Finally, we propose an optimization-based adaptive action space planning method. By solving a constrained optimization problem to dynamically adjust action boundaries for velocity and yaw rate, this method enhances the safety and flexibility of local avoidance maneuvers. Experimental results demonstrate that, compared to baseline methods, our approach yields more robust navigation strategies and substantially improves success rates in complex environments.