Chun-Hao FanRIH-TENG WUYung-I Chang2025-05-012025-05-012025https://www.scopus.com/record/display.uri?eid=2-s2.0-85217925268&origin=recordpagehttps://scholars.lib.ntu.edu.tw/handle/123456789/728765Cracks are critical indicators of the structural safety of infrastructure. Conventional crack inspection relies on human power, which is time-consuming and labor-intensive. Recently, Artificial Intelligence (AI) promotes more advanced crack identification approaches. However, existing studies still lack of ability to explore cracks in an autonomous onboard manner. To address this challenge, this paper presents a framework based on deep reinforcement learning. The framework consists of an inspection and an exploration agent, where the former utilizes U-Net to segment cracks, while the latter then employs Double Deep Q-network to navigate itself to find more unseen cracks. Built upon DeepCrack dataset, the proposed agent achieves a capture rate of 75.7% and 72.1% on training and testing environments, respectively. By visualizing the inspection trajectories, results demonstrate that the proposed agent successfully explores and collects more crack regions by itself after observing only a fraction of the crack, showcasing promising potentials for robotic inspection.Crack explorationCrack identificationDeep reinforcement learningRobotic inspectionSemantic segmentation[SDGs]SDG3journal article10.1016/j.autcon.2025.106009