ADSKAILab/floora-0.6b

Overview

FLOORA (Floor Layout Optimization with RL Alignment) is a specialized language model developed by Autodesk Research for architectural floor-plan generation. This 0.6 billion parameter variant, based on Qwen3-0.6B, has an effective size of approximately 440M parameters due to a reduced domain-specific vocabulary. It is trained to generate structured multifamily residential layouts from architectural prompts, outputting a compact, human-readable, and token-efficient architectural Domain-Specific Language (DSL).

Key Capabilities

• Domain-Specific Generation: Autoregressively generates labeled floor-plan polygons for living units, corridors, and vertical circulation cores within multifamily residential buildings.
• Structured Output: Produces a custom architectural DSL that is deterministically parseable, allowing for validation, normalization, rendering, and conversion into geometric floor plans.
• Advanced Training: Utilizes domain-specific pretraining, architect-edited supervised fine-tuning, and reinforcement learning with learned architect preferences and verifiable geometric/functional rewards.
• Efficiency: The custom DSL is significantly more compact than traditional formats like IFC, requiring 642 DSL characters compared to over 10,000 for an equivalent example.

Good For

• Rapid Conceptual Design: Ideal for quick generation and exploration of multifamily residential layouts.
• Automated Layout Exploration: Supports automated design exploration and architect review workflows.
• Architectural Research: Useful for research in geometry-aware design optimization and domain-specific language models for engineering.
• Structured AEC Generation: Facilitates experiments in structured Architecture, Engineering, and Construction (AEC) generation.