Drone Navigation and Target Interception Using Deep Reinforcement Learning: A Cascade Reward Approach

This article proposes an architecture for drone navigation and target interception, utilizing a self-supervised, model-free deep reinforcement learning approach. Unlike the traditional methods relying on complex controllers, our approach uses deep reinforcement learning with cascade rewards, enabling a single drone to navigate obstacles and intercept targets using only a forward-facing depth–RGB camera. This research has significant implications for robotics, as it demonstrates how complex tasks can be tackled using deep reinforcement learning. Our work encompasses three key contributions. First, we tackle the challenge of partial observability when employing nonlinear function approximators for learning stochastic policies. Second, we optimize the task of maximizing the overall expected reward. Finally, we develop a software library for training drones to track and intercept targets. Through our experiments, we demonstrated that our approach, incorporating cascade reward, outperforms state-of-the-art deep
Q
-network algorithms in terms of learning policies. By leveraging our methodology, drones can successfully navigate complex indoor and outdoor environments and effectively intercept targets based on visual cues.