Roomer Turns 3D Indoor Layout Generation into a Repairable Workflow

Lead

Modern 3D indoor layout generators can create scenes that look plausible at a glance. Yet small geometric failures often remain: a chair intersects a table, a cabinet slips beyond a wall boundary, furniture blocks a door, or a walking path becomes unusable. Roomer focuses on these residual local violations. Instead of treating a faulty scene as something to regenerate from scratch, it asks a more operational question: which object caused the issue, what local edit should be made, and can the repaired scene pass a full verification step?

Core ideas

• Repair at the object level: Roomer casts layout errors as sparse, object-grounded repair problems. This is different from approaches that optimize or regenerate the whole scene without explicitly identifying the responsible object instances.
• Structured layout representation: The system encodes a scene as RoState, giving the repair process a structured model of the layout rather than only a visual impression.
• Violation-to-object binding: RoReview connects measured violations—such as collisions, out-of-bounds placements, obstructed openings, and blocked circulation—to the objects likely responsible for them.
• Planner plus deterministic solver: A geometry-conditioned vision-language model proposes a structured local edit. When necessary, a deterministic solver validates the proposed action and can generate a finite set of candidate edits.
• Commit only after verification: A candidate edit is accepted only if full-scene checking confirms that it fixes the target violation without adding new hard violations or breaking protected constraints.
• Training and evaluation assets: The planner is trained on Roomer-CC, a controlled-corruption dataset pairing faulty layouts with object-level evidence and known feasible inverse StatePatches. The authors also introduce Roomer-Eval to assess distributional quality, physical validity, and practical usability.

Why it matters

Roomer is interesting because it moves indoor layout synthesis from a one-shot generation problem toward a diagnose-repair-verify loop. For design tools, this could reduce the need for manual cleanup after a generated layout appears mostly correct but contains local failures. For embodied AI and simulation, usability constraints such as reachable openings and unblocked circulation are not cosmetic details; they affect whether agents can move and interact inside the scene.

The paper also highlights a broader evaluation gap. A layout may be physically plausible in a narrow sense while still being impractical: a door might be blocked, or a route through the room might be unusable. By proposing Roomer-Eval, the authors argue that scene generation should be judged not only by distributional realism or collision statistics, but also by whether the resulting space can actually function.

The available summary does not detail how Roomer behaves under more complex floor plans, user preferences, or production CAD/BIM pipelines. Still, the method points to a useful direction: combine generative planning with deterministic geometric checks, keep edits local when possible, and make repair decisions traceable to specific objects.

Source: Hugging Face Daily Papers