Architectural drawings are powerful tools for representing spatial, functional, and relational knowledge about buildings and environments. While architects intuitively interpret these drawings as cohesive wholes, drawing on experience and tacit knowledge, machines still struggle to achieve the same depth of understanding. This research explores how learning machines can be enabled to read and analyse floorplans in ways that more closely resemble human architectural reasoning. 

A central challenge lies in translating architectural knowledge into formalised representations that machines can process—capturing geometry, topology, and semantics with clarity and precision. To address this, the paper proposes a theoretical framework that deconstructs the multiple layers of information embedded in floorplans. Through a case study of a house, it links visual features to computer vision tasks such as detection, segmentation, and classification, while also connecting architectural meaning to structured knowledge models. 

By defining 15 distinct layers of spatial information, this work lays the foundation for richer architectural datasets and more transparent machine interpretation. Ultimately, the study contributes to bridging the gap between human perception and computational analysis, opening new possibilities for design support, knowledge representation, and collaboration in the era of multimodal artificial intelligence.

This study employs a qualitative methodology to explore the types of information that should be encoded in floorplans. It combines two approaches: (1) a historical analysis of floorplan understanding from the 1960s to the present, highlighting the evolution from early CAD to advanced AI technologies capable of capturing complex qualitative aspects of floorplans, and (2) a case study analysis to illustrate what is meant by qualitative and complex semantics. 

The case study analysis focuses on Château du Val, a 17th-century house in the French countryside, which has undergone several renovations and extensions over the centuries as it served various families and associations. The case study floorplan is analysed systematically across four distinct categories: visual perception, spatial perception, machine perception, and knowledge representation (See Appendix). Visual and spatial perception are grounded in architectural education, machine perception draws from the fields of computer vision and document analysis, and knowledge representation is informed by established frameworks in design research. The building was chosen based on several factors: its illustrative potential, covering a range of geometric, topological, and semantic concepts; its historical and contextual relevance, having been repurposed over centuries; its comparability to previous studies in shape grammars and rectangular arrangements; and its residential function, making it comparable to floorplans in contemporary AI datasets. 

More information about the study results and layer visualizations will be added once the official publication is released in 2026. 

This text is based on the abstract of the conference paper 'Reading Between the Lines: Bridging AI and Architectural Knowledge in Floorplan Understanding', presented at the CAAD Futures 2025 Conference in Hong Kong. The full paper received an Honourable Mention and will be published by Springer in 2026.