HomeWorld: A Unified Floorplan-to-Furnished Framework for Generating Controllable, Densely Interactive Whole-Home Scenes

Researchers introduce HomeWorld, a unified framework for generating complete, furnished home scenes from floorplans using hierarchical AI models. The system combines large language models for floorplan generation, image models for furniture layout, and vision-language models for iterative refinement, producing simulation-ready indoor environments with a dataset of 300K real floorplans and 5K fully furnished scenes.

HomeWorld addresses a significant bottleneck in embodied AI development and virtual environment creation by automating the generation of complex, realistic indoor spaces. Rather than treating floorplan design and furniture placement as separate problems, the researchers developed an integrated pipeline that maintains coherence across the entire home. This hierarchical approach—progressing from macro layout decisions to micro object placement—mirrors how human designers think about interior spaces, enabling both fine-grained control and emergent realism.

The release of 300K floorplans and 5K furnished scenes represents substantial value for the robotics and simulation communities, which have historically struggled with data scarcity for training embodied AI systems. The use of vision-language models to iteratively refine placements demonstrates how modern foundation models can enforce design principles and physical feasibility without explicit rule engineering. This shift from hand-crafted constraints to learned refinement patterns reduces engineering overhead while improving quality.

For the broader AI ecosystem, this work signals maturation in scene generation technology. Virtual environment generation has direct applications in robot training, game development, and architectural visualization—markets that previously required significant manual labor. The framework's modularity, allowing asset replacement and customization, makes it accessible to downstream applications. The public release of datasets and trained models will likely accelerate research in embodied AI, potentially reducing development costs for companies building simulation platforms. The emphasis on physics-aware generation and surface properties indicates the field is moving toward production-ready outputs rather than visual approximations.

• →HomeWorld provides an end-to-end framework for generating realistic, simulation-ready home environments with controllable design parameters.
• →A dataset of 300K real floorplans and 5K fully furnished scenes will be publicly released, addressing major data scarcity in embodied AI research.
• →Vision-language model refinement replaces hand-crafted rules with learned constraints, improving both quality and design flexibility.
• →The hierarchical decomposition of home generation enables fine-grained control while maintaining global coherence across whole-home layouts.
• →Physics-aware object placement and basic texture/lighting setups move generated scenes closer to practical simulation and visualization use cases.