Project Details
[Return to Previous Page]AI-Driven BIM-Graph Agents for Intelligent and Adaptive Indoor Robotic Inspection
Company: Penn State University Cocoziello Institute of Real Estate Innovation
Major(s):
Primary: CMPEN
Secondary: CMPSC
Optional: EGEE
Non-Disclosure Agreement: NO
Intellectual Property: NO
Facility inspection and maintenance are among the most resource-intensive phases of a building’s life cycle, representing 30–50% of total operational costs in large commercial and institutional facilities. In the U.S., maintenance technicians typically earn $25–45 per hour, with complex inspections, such as HVAC, fire safety, and electrical systems—requiring 2–3 hours per site visit and often involving multiple specialists. For a mid-sized university building portfolio, this translates to annual inspection costs exceeding $2 million. More critically, traditional inspections are largely reactive, issues are identified only after occupant complaints or performance degradation, leading to average repair delays of 3–7 days and downtime costs exceeding $0.50–$1.00 per square foot per day. Manual workflows further cause data fragmentation: inspectors frequently rely on disconnected mobile devices and handwritten notes, resulting in 20–30% of recorded issues lacking precise spatial localization within Building Information Modeling (BIM) or facility management systems. These inefficiencies collectively hinder data-driven maintenance planning and inflate life-cycle costs. To address these challenges, this project proposes a novel AI-driven multi-agent framework for autonomous indoor robotic inspection that seamlessly integrates Building Information Modeling (BIM) with large foundation models to achieve spatial reasoning, adaptability, and real-time decision-making. In the planning stage, the building is represented as a hierarchical BIM-Graph extracted from its digital twin, enabling structured scene understanding. A Large Language Model (LLM) Planner interprets this graph to infer spatial and functional relationships (e.g., adjacency between air handling units and diffusers) and generates inspection waypoints optimized for coverage and efficiency. In the execution stage, a Vision-Language Model (VLM) Agent dynamically updates the inspection plan based on real-time observations, performing adaptive re-planning when encountering obstructions or anomalies. The proposed framework is distinctly novel in three ways: 1) It introduces a BIM-Graph as a domain-specific scene representation, enabling large foundation models to reason about architectural and mechanical hierarchies, capabilities absent in general robotic reasoning systems. 2) It establishes a cooperative multi-agent loop between LLM and VLM agents, coupling symbolic planning with perceptual grounding for continuous adaptation in dynamic indoor environments. 3) It advances domain-adaptive autonomy by embedding building semantics directly into AI reasoning, bridging the long-standing gap between BIM data and embodied robotic intelligence. Evaluation will focus on waypoint optimization accuracy, task completion rate, defect localization precision, and projected labor-hour reduction. The anticipated outcome is a 50–70% reduction in manual inspection time, significantly improved localization accuracy, and a validated pathway toward scalable, cost-effective, and intelligent facility management across the built environment.
