RadDQN: A Deep Q Learning-Based Architecture for Finding Time-Efficient Minimum Radiation Exposure Pathway

Recent advances in deep reinforcement learning (DRL) have expanded its use in various automation sectors, including the nuclear industry. While DRL shows promise for optimizing radiation exposure, the development of radiation-aware autonomous unmanned aerial vehicles (UAVs) is hindered by inefficient reward functions and exploration strategies. In this article, we introduce a radiation-aware deep Q-learning network (RadDQN), designed to provide time-efficient, minimum radiation-exposure pathways in radiation zones. RadDQN operates on a radiation-sensitive reward function considering surrounding radiation intensity through the inverse square law and prioritizes reaching the final destination. Departing from the traditional $\epsilon $ -greedy algorithm, RadDQN implements unique exploration strategies that guide the agent to transform random actions into model-directed ones if transitioning to a future state projects higher radiation exposure compared to its current state. This approach ensures minimal radiation exposure while efficiently progressing toward the goal. We validate RadDQN's accuracy against a grid-based deterministic method. Our results demonstrate that the formulated reward function and exploration strategy effectively manage various radiation field distributions. In addition, RadDQN shows superior convergence rates and higher training stability compared to the baseline model, indicating its effectiveness in optimizing radiation-aware UAV navigation and potential for real-world applications.