Roomer frames indoor-layout failures such as collisions, out-of-bounds placements, blocked openings, and obstructed circulation as sparse, object-grounded repair problems. Its RoState representation and RoReview module connect measured violations to responsible objects. A geometry-conditioned vision-language planner proposes structured local edits, while a deterministic solver validates or enumerates candidate edits. Candidates are committed only after full-scene verification confirms that the target violation is removed without introducing new hard violations or violating protected constraints. The paper also introduces Roomer-CC for controlled corruption training and Roomer-Eval for distributional quality, physical validity, and usability, claiming transfer across external layout generators.
No heat snapshots are available in the last 24 hours.