Smart Robotic Inspection of Levees and Culverts: A Scoping Review of Robotic Platforms, Sensing Technologies, AI Methods, and Multi-Robot Coordination for Flood-Control Infrastructure

Abstract Earthen levees, dikes, flood-control embankments, culverts, and stormwater drainage conduits represent vital lifelines for mitigating catastrophic inun-dation in urban and rural watersheds. However, conventional manual walk-downs and tethered closed-circuit television (CCTV) surveys remain severely labor-intensive, hazardous, subjective, and difficult to scale across vegetated, flooded, submerged, and GPS-denied environments. Autonomous robotics, multimodal remote sensing, deep-learning-based computer vision, and multi-robot coordination offer transformative opportunities to enhance inspection frequency, spatial accuracy, and maintenance decision-making. Following the Joanna Briggs Institute (JBI) methodology and the PRISMA extension for Scoping Reviews (PRISMA-ScR), this study maps the multidisciplinary land-scape across six major scholarly databases: Scopus, Web of Science, IEEE Xplore, ScienceDirect, SpringerLink, and Google Scholar (2010–2026). Out of 1,860 initially retrieved records, 1,314 unique studies were screened, 212 full texts were assessed, and 64 primary empirical studies were systemati-cally synthesized. We chart the evidence across seven interconnected the-matic streams: (1) unmanned aerial vehicle (UAV) corridor surveying; (2) unmanned ground vehicles (UGVs) and pipe crawlers; (3) amphibious and autonomous underwater vehicles (AUVs/ROVs); (4) multimodal sensor fusion (RGB, thermal infrared, LiDAR, sonar, and GPR); (5) deep learning for defect detection and semantic segmentation; (6) degraded-environment SLAM and GPS-denied navigation; and (7) multi-robot coordination, com-munication relays, and digital-twin maintenance integration. Quantitative findings show that while aerial thermal/RGB levee inspection and in-pipe YOLO/U-Net defect detection achieve high localized accuracy (mAP ex-ceeding 75–86%, segmentation IoU > 0.71), end-to-end systems integration remains scarce. Only 10.9% (n = 7) of the literature validates heterogeneous multi-domain robot teams in operational flood-control environments, and fewer than 5% close the loop from automated anomaly detection to predic-tive maintenance prioritization. To bridge these gaps, we propose a resilient six-tier reference architecture for communication-aware, heterogeneous robot teams and establish a concrete ten-point research agenda for operational flood-control infrastructure management.