Automating 2D Floor Plans with NLP
Final Year Project — From natural language to AutoCAD floor plans.
An AI-powered system that turns plain English descriptions into complete 2D house floor plans, helping architects and homeowners create layouts in minutes instead of drawing them manually. Built using an NLP pipeline with prompt refinement, a custom-trained Named Entity Recognition (NER) model, constraint satisfaction algorithms for layout optimization, and AutoCAD APIs to automatically generate multi-floor 2D floor plans directly in AutoCAD.
Architectural drafting is time-consuming and requires specialized skills. This project automates the initial floor plan generation phase purely from natural language input — converting plain English descriptions into structured, multi-floor residential layouts. Users can specify complete requirements like room count, floor count, and plot size, or just provide a plot size and let the system generate the most optimal layout on its own.
End-to-end automation
Goes from a plain English prompt straight to a finished AutoCAD drawing, no manual drafting in between.
Seconds, not hours
Generates a full layout in 15-90 seconds depending on prompt length.
Works with vague input
Handles both detailed specs and loose descriptions like "a large room next to the kitchen."
Scales to multi-floor plans
Tested up to 4 floors while keeping layouts consistent across levels.
90% layout accuracy
Constraint solver resolves conflicting requirements into a spatially sound plan.
User enters floor plan requirements in natural language.
Prompt is refined and missing values are filled in.
Architectural entities — rooms, floor count, adjacency — are extracted.
Entities are passed to a solver that generates an optimal layout.
Drawing engine converts the layout into a 2D vector floor plan.
Ambiguity in natural language
Text descriptions are inherently ambiguous — "a large room next to the kitchen" has many valid interpretations. Built a constraint resolution layer that picks the most spatially optimal arrangement.
Training data scarcity
Architectural NER datasets are rare. Built a custom annotated dataset by manually labeling architectural descriptions and augmenting with synthetic examples.
Producing conventional layouts
A layout that satisfies every constraint can still look haphazard. Tuned the CSP solver so rooms stay properly attached and adjacent, resulting in floor plans that read as realistic, buildable designs rather than a technically valid but unnatural arrangement.
Mapping the solution to AutoCAD
The CSP solver outputs room relationships, not geometry. Built a translation layer that converts the solved layout into precise x and y coordinates, then drives the AutoCAD API directly to draw walls and rooms in the correct positions.
Successfully demonstrated end-to-end automation from text to AutoCAD drawing — generating multi-floor plans (tested up to 4 floors) with 90% accuracy, completing the full process in 15-90 seconds depending on prompt length. Final Year Project at FAST-NUCES.
Team members
Supervisors
Syed Kashan Hussain, Haris Mohsin