Building maintenance, particularly corrective maintenance that requires a rapid response to unexpected defects, remains a challenging area to automate. While Building Information Modeling (BIM) provides rich data for maintenance purposes, its direct application to robotics poses challenges due to differing approaches in 3D environment representation. This research presents a framework to address this gap by utilizing Large Language Models (LLMs) and introducing a novel BIM-driven 3D Scene Graph (BIM3DSG) for automated robot task planning in corrective building maintenance. The proposed framework comprises five LLM-based modules that handle semantic searching, robot selection, navigation planning, scanning planning, and task execution. To enable seamless integration between building data and robotics, this research introduces BIM3DSG, which bridges volumetric BIM data and robotics-focused 3D Scene Graphs. BIMSegGraphs, representing room-wise segmented BIM data, are mapped to an extended 3D Scene Graph structure, facilitating the direct use of BIM data in robotic applications while addressing token limitations inherent to LLMs. Evaluation results indicate that the baseline GPT-4o model performs well on straightforward tasks but struggles with complex spatial reasoning. Conversely, the advanced o1-preview model demonstrates superior performance in handling intricate scenarios, although challenges persist in determining optimal scanning positions. The framework showcases the potential for automating corrective maintenance tasks, despite remaining limitations in scanning optimization. This research contributes to the field by: (1) developing a comprehensive high-level planning framework for robotic corrective maintenance, (2) introducing BIM3DSG as a unified representation linking BIM and robotics, and (3) conducting a detailed evaluation of LLMs’ spatial reasoning capabilities in the Architecture, Engineering, and Construction (AEC) domain, integrating a real-world robotics library.     «

Building maintenance, particularly corrective maintenance that requires a rapid response to unexpected defects, remains a challenging area to automate. While Building Information Modeling (BIM) provides rich data for maintenance purposes, its direct application to robotics poses challenges due to differing approaches in 3D environment representation. This research presents a framework to address this gap by utilizing Large Language Models (LLMs) and introducing a novel BIM-driven 3D Scene Graph...     »