Floor plan compliance: AI framework for residential checks

Subash Gautam, Debaditya Acharya, Alexandra Kleeman and Sarah Foster submitted a paper to arXiv (arXiv:2607.00015) on 26 May 2026 proposing a conceptual, AI-driven framework for automated floor plan compliance checking in multi-apartment residential buildings. The authors frame the work against Australian policy reforms including SEPP65, BADS and SPP7.3 and argue current manual checks do not scale for large inventories of apartments.

What does the paper propose?

The paper proposes a three-part automated pipeline: a Rule Engine that converts textual building codes into executable, explainable rules using a Large Language Model (LLM), a Data Extraction Engine that segments floor plan images into elements, and a Compliance Check Engine that evaluates the extracted, structured representation against the rules. The Data Extraction Engine produces a structured building graph with topological relationships by identifying walls, rooms, fixtures, text and symbols in floor plans. The Compliance Check Engine then applies the LLM-generated rules to that graph to determine compliance.

The authors position this framework as addressing fragmentation in existing methods, which they say typically focus on single apartments and lack a unified approach for multi-unit compliance checking. They describe the framework as intended to support consistent enforcement across jurisdictions and to scale assessments across many apartments.

How does the framework work in practice?

At a high level the flow is: floor plan images are fed into a Data Extraction Engine that segments raw images into discrete elements; those elements are transformed into a structured building graph capturing topology; an LLM-driven Rule Engine converts textual building codes into executable rules; the Compliance Check Engine evaluates the structured graph against those rules. The paper lists the Data Extraction Engine outputs explicitly as walls, rooms, fixtures, text and symbols.

The LLM is embedded within the Rule Engine to translate human-readable code into rules that can be executed and explained. The structured building graph records topological relationships between elements so the Compliance Check Engine can perform geometric and spatial analyses relevant to health-related features named by the authors, including daylight access, natural ventilation, privacy and space efficiency. The authors present the design as a conceptual architecture rather than a finished, field-tested system.

Why it matters

Automating floor plan compliance could reduce the time and labour the authors identify as barriers to evaluating apartment design, especially as regulations evolve. The paper highlights that current manual checks are time-intensive and that evolving policies limit scalability for large-scale assessments across thousands of apartments. Embedding an LLM into a rule-conversion role aims to make textual codes machine-executable while retaining explainability, which matters for regulatory transparency and enforcement.

Shifting routine, geometric compliance checks into an automated pipeline would change where human expertise is focused: from repetitive measurement to validation of rule translations, edge cases and policy interpretation. For jurisdictions grappling with apartment design quality reforms such as SEPP65, BADS and SPP7.3, the framework targets a practical bottleneck in enforcement and large-scale policy assessment.

What to watch

Whether the conceptual framework moves to implementation and empirical validation is the next concrete signal to follow; the paper presents a design but not a deployed system. A decisive test will be demonstrations that the Data Extraction Engine and LLM-driven Rule Engine can together assess many apartments at scale and handle jurisdictional variations in code language, as the authors emphasise scalability for large-scale assessments across thousands of apartments.

Authors and identifiers: the paper is arXiv:2607.00015, submitted 26 May 2026, by Subash Gautam, Debaditya Acharya, Alexandra Kleeman and Sarah Foster. The abstract situates the work in Computers and Society, Artificial Intelligence and Computer Vision research areas.