Automatic furnishing of architectural floor plans is a longstanding challenge for real-estate visualization, interior design, and architectural workflows. The core difficulty lies in the scarcity of real-world data and the need to satisfy multiple interacting geometric and functional constraints simultaneously.
To address this, researchers have released AntPlan, a curated dataset of 505 real professional architectural floor plans with dense furniture annotations covering 92 object classes and ten residential room categories. Alongside it, they present Architect-Ant, a framework that generates furniture layouts using an editable coordinate-based domain-specific language (DSL).
Architect-Ant first learns professional furnishing patterns through supervised fine-tuning of a vision-language model, then refines its output with GRPO (Group Relative Policy Optimization) guided by a Layout Rule Score (LRS). This score aggregates geometric and functional constraints derived from professional plans, providing outcome-level supervision without requiring prescribed reasoning traces.
According to the authors, experiments against diverse state-of-the-art baselines show that Architect-Ant achieves the highest functional completeness and overall layout-rule scores while maintaining low geometric violation rates. Because no search or critique loop is needed at inference time, the method generates layouts substantially faster than iterative agentic approaches.
The resulting layouts remain object-level editable and can be converted directly into furnished 3D scenes. The team believes the work and the AntPlan dataset may benefit researchers working on floor-plan furnishing, indoor scene synthesis, and constraint-aware layout generation more broadly. The project page and dataset card are publicly available.