AI Agents Now Draft Floor Plans That Pass Municipal Building Codes

Generative artificial intelligence has conquered images, music and text, but it has long stumbled over a deceptively simple challenge: drawing a floor plan that a municipal building department would actually approve. A research team at RV College of Engineering in Bengaluru now reports a framework called LayoutGen that closes this gap, coupling the creative capacity of large language models with the unyielding rigor of deterministic geometric verification. Published in Discover Artificial Intelligence, the study shows that an evaluation-guided multiagent loop can slash regulatory violations in generated residential layouts by 92 percent, dropping from 3.7 violations per plan in a single-shot baseline to just 0.3 across 200 automated generation runs.

The problem LayoutGen attacks is what the authors call a fundamental validity gap in probabilistic generation. Diffusion models and adversarial networks such as HouseGAN, HouseGAN++ and HouseDiffusion can produce layouts whose statistical and visual properties closely resemble expert-designed plans, yet they routinely violate the regulatory, functional and safety requirements that constitute real architectural expertise. A diffusion model can enforce that a wall is straight, the paper notes, but it lacks the cognitive apparatus to know that a bedroom needs a window or that a habitable room must exceed a statutory minimum area. Zoning regulations, structural constraints and life-safety codes impose strict requirements that a layout appearing architecturally reasonable may still breach, from minimum egress widths to setback rules and ventilation provisions.

LayoutGen’s answer is a neuro-symbolic architecture in which a large language model serves as the primary reasoning engine for spatial planning, coordinate proposal and natural language feedback, while a deterministic rule checker, implemented in Python, verifies every quantitative constraint against a structured regulatory repository. The division of labor is absolute: no threshold, score or grade originates from the language model. Every scored quantity is computed by geometric checks over the generated room coordinates, and the model’s output is treated as a proposal, never a verdict. A plan the model believes to be compliant is scored non-compliant if the geometry says otherwise, and the refinement loop closes through geometry rather than through assertion, bounding the effect of model error on reported compliance.

Orchestrating the workflow is a state-management protocol the authors call the Chain, which represents the design process as a sequence of explicit, verifiable state transitions. Each design state is an immutable tuple containing a topological bubble diagram, a geometric vector plan, the active constraint set and the evaluation report. The Chain admits a transition only when an intermediate state clears a quality threshold, generally a grade of B− or 65 percent, so corruption cannot propagate through the pipeline. This evaluation-gated gating also mitigates the context drift problem that plagues long-horizon language model agents, in which constraints specified early in an interaction are forgotten later, because every constraint active at a given step is carried in the state itself rather than in conversational memory.

Context arrives through a dual retrieval-augmented generation scheme. A Visual RAG stream searches a curated repository of reference floor plans indexed by interpretable metadata such as typology captions, dimensional attributes and style categories, giving the Architect Agent realistic geometric priors without learned embeddings. A Legal RAG stream injects structured municipal building by-laws, pre-processed as deterministic JSON assertions, each carrying the clause number it transcribes, the threshold value, the affected room categories and the violation severity. The three implemented repositories cover Bengaluru’s BBMP, Chennai’s CMDA and Hyderabad’s GHMC regimes, and each bye-law document was ingested in full, so every clause is present and retrievable. Constraints are injected directly into agent system prompts as high-priority fixed values, ensuring the model cannot negotiate away critical safety requirements and providing full auditability of which regulations were applied.

The Multi-Dimensional Evaluation Engine scores each candidate layout along three weighted criteria: regulatory compliance at 40 percent, design quality at 35 percent and sustainability at 25 percent, yielding a final grade from A+ to F. Compliance is measured over the provisions decidable from single-floor room geometry, principally minimum room areas, minimum room widths and a critical circulation width below which a plan is disqualified outright as a life-safety failure. Design quality decomposes into circulation efficiency, Golden Ratio room proportioning, spatial balance measured by the offset between geometric center and weighted centroid, and functional flow between key room pairs such as kitchen and dining. Sustainability combines window-to-floor ratio with material efficiency. Notably, corridor share appears in both the design and sustainability composites, making it the most heavily weighted single sub-metric in the engine. An advisory Vaastu Shastra assessment, covering traditional placements such as kitchens in the Southeast and master bedrooms in the Southwest, informs the feedback loop without entering the grade.

The experimental results trace a striking trajectory of improvement. A component ablation shows that Visual RAG alone lifts the compliance score by 2.3 points by providing typological priors, while Legal RAG alone contributes a far larger gain of 17.2 points by injecting deterministic regulatory constraints. Combining both retrieval streams yields a further 2.9 points, and adding the Evaluation Engine with iterative refinement pushes compliance to 93.5 after one pass and saturates at 97.8 after three, with violations falling to 0.3 per plan. The refinement proceeds in two phases: a steep correction phase resolving critical safety violations, followed by an optimization phase polishing Golden Ratio proportioning and centroid balance. The Chain guarantees monotonic improvement by construction, since the state carried forward is always the highest-scoring state observed, and the procedure always terminates within its iteration budget.

Portability emerges as a headline finding. The same multiagent architecture ran unchanged across all three municipal rule repositories through structured rule substitution alone, without retraining or prompt redesign, and across four commercial language model APIs with byte-identical prompts. The compliance column varied by only 4.2 points across providers against 7.0 points on design quality, the pattern expected when one dimension is enforced by an external deterministic checker and the other depends on how well the model exploits semantic feedback. Every backbone tested, from Claude Opus 4.6 to GPT-4o and Gemini 2.5 Flash, reached a deployable configuration, suggesting the capability does not hinge on a single proprietary model. Across all 200 runs the framework also achieved a 100 percent geometric validity rate, with no overlapping room polygons or boundary violations in any final output.

Compared against a deliberately minimal baseline built on the Z3 satisfiability modulo theories solver, the contrast is revealing. The solver consistently satisfies hard constraints but produces rigid, grid-aligned layouts scoring below 60 on design quality, and when the specification becomes overconstrained it simply reports unsatisfiability, leaving the designer to decide which constraint to relax. LayoutGen instead returns a near-feasible layout, reports which constraint the current geometry violates, and relaxes the soft preference, never the statutory threshold. In a qualitative case study, a 3BHK plan for a 30-by-40-foot plot initially misplaced the kitchen in the Northeast, scoring an A−; after semantic feedback relocated it to the Southeast per Vaastu requirements and improved dining adjacency, the weighted final score rose from 82.2 to 90.6, an A+ grade.

The framework is not without trade-offs. The full three-iteration pipeline requires roughly 38.4 seconds and about 0.70 dollars per plan against 6.8 seconds and 0.12 dollars for single-shot generation, a factor of 5.8 in cost that the authors argue is negligible relative to professional design timescales and the expense of correcting violations by hand. The compliance measure itself is scoped to dimensional provisions decidable from room geometry; floor area ratio, setbacks, egress path length, staircase geometry, parking and ventilation are present in the repositories but checked outside the generation loop, since a single-floor room layout does not determine them. The BBMP repository encodes the 2003 bye-laws, and regulatory currency is maintained by an external clause-by-clause update process. Still, the reference implementation is openly available under the MIT licence, and because the Chain protocol and Evaluation Engine contain no architecture-specific logic, the authors identify warehouse zoning, printed circuit board floorplanning and user-interface layout as natural targets for the same generate-evaluate-refine machinery. The study, they conclude, marks a shift from purely probabilistic floor plan generation toward compliance-aware generation, in which rule violations actively shape synthesis rather than merely auditing its outcomes.

Subject of Research: Artificial Intelligence

Article Title: An evaluation guided multiagent framework for generating floor plans that comply with municipal building regulations

Article References: An evaluation guided multiagent framework for generating floor plans that comply with municipal building regulations. (n.d.). https://doi.org/10.1007/s44163-026-02481-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02481-9

Keywords: LayoutGen, multiagent systems, neuro-symbolic AI, floor plan generation, building regulations, retrieval-augmented generation, automated compliance checking, large language models, Vaastu Shastra, deterministic evaluation, evaluation, guided