Scour and structural damage caused by hydraulic-bridge interaction threaten cross-river structures. Three-dimensional sonar provides direct underwater inspection, but complex aquatic environments and ship-mounted acquisition systems introduce significant noise, limiting data usability. This paper proposes SIREND-Net, a deep learning framework for denoising bridge sonar point clouds. The method employs a three-stage strategy: outlier removal, suppression of surface-adjacent fuzzy noise, and density restoration through inverse manifold reconstruction. By flexibly switching between supervised and unsupervised modes, SIREND-Net adapts to diverse field conditions. To address data scarcity, USPCDenoise, a field-measured dataset tailored for bridge sonar point cloud denoising, is developed. Based on real-world data, sonar noise morphology is analyzed using equivalent thickness as a surrogate metric under a manifold-like approximation. Comparisons with ten representative methods show that SIREND-Net achieves superior denoising accuracy, generalization, and efficiency. Field experiments further confirm improved measurement precision, demonstrating its practical value for bridge underwater inspection.
Informed decision-making for built assets requires a comprehensive, spatially integrated overview of available documentation. This documentation, however, is typically heterogeneous, siloed, and loosely connected. The spatial relationship of each data source to the structure thereby remains implicit and accessible only to individual experts. To interlink these sources and query them automatically, their spatial relations to the asset must be made machine-readable, yet no vocabulary exists to express them. To address this, the Space Types Ontology (SPOT) is developed to explicitly express these relationships in a format-agnostic way, including the qualitative references that prevail in legacy documentation. The applicability of SPOT is demonstrated by semi-automatically interlinking the documentation of the Nibelungen Bridge in Worms, Germany, into a spatial knowledge graph. This approach makes heterogeneous asset data queryable across data types, reducing dependence on individual expert knowledge and providing a format-independent foundation for informed, data-driven asset management.
The integration of Vision-Language Models (VLMs) into robotic welding is transforming quality inspection for industrialized construction from traditional pattern matching to high-level semantic reasoning. However, a critical challenge remains: how can small-scale VLMs be enabled to perform reliable, interpretable, and robot-specific welding-quality inspections? This paper proposes Visual-Physics Reference-Guided Welding (VPR-Weld), which leverages retrieved multimodal samples as anchors for in-context defect diagnosis and physics-guided parameter optimization. Using Qwen3.5-VL-2B as a representative backbone, binary standard/defect classification reached a peak accuracy of 0.8673 and an F1-score of 0.8194. Across nine architectures (0.8B–4B), LLM-judge scores ranged from about 1.0 without references to above 3.4 for the best configuration, while adaptive calibration reached 3.89 and judge scores correlated with expert ratings (r = 0.79). The 2B model maintained a VRAM footprint below 4 GiB, supporting memory-efficient local inference. Without data-intensive fine-tuning, VPR-Weld offers scalable decision support for high-mix manufacturing.
Building-instance meshing from multi-view UAV imagery is essential for digital twins and 3D geographic information modeling. However, reconstructing individual buildings in large aerial scenes remains challenging because existing methods often process scenes globally without preserving building identities. Consequently, neighboring buildings may become incorrectly connected, resulting in inaccurate boundaries and background structures incorporated into building meshes. This paper proposes Proxy-Coupled Masked Gaussian Fields, an instance-aware framework designed to generate separated and structurally reliable building meshes in large-scale aerial scenes. Unlike existing methods that optimize and fuse the scene globally, the proposed method preserves building identities from initialization and constrains reconstruction with conservative proxy supports and mask evidence. This design reduces interference between adjacent buildings, mitigates occlusion and mask errors, and prevents spurious mesh connections during extraction. Experiments on large-scale aerial datasets demonstrate that the proposed method produces cleaner boundaries, less background noise, and more reliable instance-level building meshes than existing baselines.
Building Information Modeling (BIM) authoring in the Architecture, Engineering, and Construction (AEC) sector entails navigating complex professional software, demanding significant manual effort and domain expertise. Current automation approaches rely on rigid Application Programming Interfaces (APIs), which detach agents from human-centric visual workflows and limit generalizability across software platforms. This paper presents BIMgent, a multimodal LLM-based agentic framework designed to perform autonomous BIM authoring directly via Graphical User Interface (GUI) interactions. To address high visual noise and long-horizon dependencies, BIMgent incorporates documentation-guided hierarchical planning and dynamic GUI grounding. The framework is evaluated on a benchmark derived from real-world designs of residential buildings. BIMgent achieves a 77.8% end-to-end success rate on long-horizon BIM authoring tasks, significantly outperforming leading generalist computer-use models, including Claude-Sonnet-4.5 (2.2%) and GPT-5.4 (11.1%). By bridging abstract design inputs and concrete modeling actions without API constraints, BIMgent offers a viable path towards autonomous architectural modeling assistants.
