Recent advances in artificial intelligence (AI) and cyber-physical systems have fostered innovative approaches to performance assessment and management of existing building stock. This study presents an AI-assisted digital twin (DT) framework for the automated and high-precision detection of façade defects in large-scale buildings. Leveraging unmanned aerial vehicles (UAVs) for visual data acquisition, the proposed framework integrates building information modeling (BIM) and geographic information systems (GIS) into a GeoBIM-assisted DT environment. An end-to-end pipeline is developed for defect localization and semantic registration, in which a virtual building model and camera geometry are constructed using geographic metadata. Synthetic views are generated to simulate real image capture conditions, enabling depth-based inference of each defect’s spatial location. This facilitates the projection of defect data into georeferenced DT models. A dual-verification method combining image and geographic features is employed to eliminate duplicate detection across overlapping images, and structural context is retrieved via GeoBIM for semantic enrichment of defect information. The proposed system exemplifies the fusion of DT technologies with deep learning and cyber intelligence to enhance defect detection accuracy, resilience optimization, and timely building health monitoring. Experimental validation on a high-rise building in Hong Kong demonstrates the robustness and scalability of the framework, indicating strong potential for smart building maintenance and operation.

The complexity and variability of the deep-sea environment present significant challenges for autonomous underwater robots, particularly in dynamic modeling considering environmental disturbances. This paper presents a novel environment uncertainty-aware dynamic modeling approach for autonomous deep-sea robots. First, a robot multibody dynamic model is established under ideal environmental conditions. Then, environmental disturbances, including pressure, temperature, and density, are incorporated to capture their environment–robot coupling effects. Finally, a neural network compensator is designed to predict pose deviations caused by uncertain ocean disturbances. Experimental studies under deep-sea conditions show that the proposed model can accurately predict the motion state of the deep-sea robot with most depth errors within 5 m and pitch angle errors within 0.1 rad, providing a solid foundation for future autonomous deep-sea robot motion control and autonomous navigation.