Pano2BIM: A Geometry-Constrained Vision-Language Agent for As-Built BIM Generation from a Floor Plan and a Single 360$^\circ$ Image
Digital twins require accurate, up-to-date building information models (BIM) that reflect physical conditions, yet generating and maintaining as-built models remains costly and labor-intensive. Emerging image-based methods can capture visual appearance without inherently producing structured BIM information, while vision-language models (VLMs) struggle with reliable metric object placement. This paper presents Pano2BIM, a geometry-constrained VLM agent that generates approximate IFC4 models at Level of Development (LOD) 200 from a floor plan with known room dimensions and a single 360-degree photograph. The VLM identifies landmarks and objects, while deterministic tools estimate camera pose, compute metric positions through ray casting, and check spatial constraints. An optional second stage enriches the presentation model through render-and-compare feedback without modifying the primary furniture geometry. In a preliminary evaluation on one laboratory, the proposed Measure workflow achieved a mean furniture-footprint intersection over union (IoU) of 0.53, compared with 0.44 for a single-response baseline using identical modeling instructions, and matched an average of 4.0 versus 2.3 of seven fixed furniture elements across three runs. Centroid disagreements were similar for the four elements consistently identified by both approaches. The measurement stage averaged 3.8 minutes and USD 1.63 in API costs per run, while the complete demonstrated workflow required approximately 42 minutes and USD 11. These findings demonstrate the feasibility of rapid, low-cost as-built BIM generation, with potential applications in maintaining digital twins that reflect changing physical conditions.