Reinforcement Learning-Based 3D Indoor Furniture Arrangement Using Dataset Distribution-Based Rewards

This paper addresses the problem of 3D indoor furniture arrangement, where the position and orientation of each furniture object are determined from a given room floor plan and the furniture arrangement distributions and relationships observed in indoor scene datasets. Existing reinforcement learning-based indoor furniture arrangement methods are limited in that their reward criteria and constraints rely on hand-crafted rules or explicit information extracted from reference scenes. We propose a dataset distribution-based reward method that extracts category-specific rear and side wall-distance distributions from 3D-FRONT, a large-scale 3D scene dataset, and uses them as observation and reward signals for a multi-agent reinforcement learning policy. General agents learn furniture arrangement policies using reference distances derived from category-specific wall-distance distributions, while child agents are trained to converge to relative position and orientation targets defined with respect to their parent furniture objects. We evaluate bedroom, living room, and dining room scenes using FID on orthographic top-view renderings of furniture OBBs. The results show that the proposed method achieves distribution-matching performance comparable to or better than ATISS and DiffuScene. This demonstrates that directly converting dataset statistics into reinforcement learning rewards can be effective for learning indoor furniture arrangement policies.