Human-operational 3D indoor layout generation with LLM-driven anthropometric simulation

Abstract

Three-dimensional (3D) indoor layout generation is essential for interior design automation, digital twin construction, human behavior simulation, and occupant activity modeling. Although recent layout-generation methods have achieved significant advances in visual realism, they have overlooked anthropometric considerations, such as human scale and behaviors, producing layouts unsuitable for human interactions and unusable for activities in physical and virtual environments. This study introduces a human-operational 3D indoor layout-generation approach driven by a large language model–based anthropometric simulation. The proposed pipeline employs a human-action predictor that associates atomic actions with objects, and a spatial rule generator combining the scene context, predicted actions, and anthropometric data to ensure clearance, accessibility, and alignment. These anthropometric spatial rules are integrated into a diffusion-based scene-graph model during training and inference. This method reduces anthropometric rule violations, supporting the automated generation of human-operational spaces in which occupants can perform activities in digital twins and real-world implementations.