🤖 AI

Learning an Interior Layout Policy in a Domain Specific Language Action Space

This paper introduces LayoutDSL, a novel LLM-based framework that generates indoor scene layouts by learning a policy within a structured Domain-Specific Language (DSL) action space, leveraging a new dataset and reinforcement learning to overcome the limitations of direct coordinate prediction and significantly improve spatial plausibility and design logicality.

Original authors: Yuhao Lu, Weichen Zhang, Wenyi Xiao, Haohui Chen, Yiyun Fei

Published 2026-08-11

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuhao Lu, Weichen Zhang, Wenyi Xiao, Haohui Chen, Yiyun Fei

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a robot how to decorate a room. In the world of computer science, this is called "indoor scene layout generation." For a long time, scientists have tried to get computers to figure out where to put a sofa, a bed, or a bookshelf. They've taught machines to guess the exact X, Y, and Z coordinates (like a GPS for furniture) or to draw rough boxes around rooms. But here's the problem: real rooms aren't just empty boxes. They have doors that need to stay open, windows that shouldn't be blocked by a giant wardrobe, and weird corners that don't fit standard furniture. When computers try to guess exact numbers for every single piece of furniture, they often get lost, creating layouts that look like a game of Tetris where the pieces don't quite fit or, worse, float in mid-air.

This is where a new idea comes in: instead of asking a computer to do complex math to find a number, what if we asked it to speak a special language? Think of it like teaching a child to build with LEGOs. Instead of telling the child, "Place the red brick at coordinate 4.2, 7.1, 0," you say, "Put the red brick next to the blue brick, on the left side." This paper introduces a system called LayoutDSL, which teaches an Artificial Intelligence (AI) to "speak" this special language of interior design. The goal is to make the AI think more like a human designer, using logic and relationships ("next to," "facing," "centered") rather than just guessing numbers.

The Problem with Guessing Numbers

The authors of this paper noticed that many existing methods treat room design like a math problem where the computer has to predict a specific number for every piece of furniture. They call this "direct coordinate prediction." Imagine trying to describe the location of a coffee table by saying, "It is 3.613 meters from the left wall and 2.778 meters from the back wall." If you move the room slightly or change the shape of the wall, that number becomes useless. The computer has to relearn the number every time the room changes.

The paper argues that this approach is flawed because it ignores the "rules" of a room, like doors and windows. It's like trying to park a car by only looking at the GPS numbers without checking if there's a tree in the way. The computer might calculate a perfect spot, but if that spot is inside a wall or blocking a door, the design fails. The authors suggest that by forcing the AI to use a structured language (a Domain-Specific Language, or DSL) that describes relationships instead of just numbers, the AI can learn the actual logic of design.

The Solution: Teaching the AI a New Language

To fix this, the team created LayoutDSL, a framework where the AI doesn't output coordinates directly. Instead, it outputs a series of instructions in a special language. It's like giving the AI a recipe book. Instead of saying "Put the sofa at (5, 2)," the AI says: "Place the sofa facing the TV, with its back against the north wall, leaving a small gap."

Here is how they built it:

The Language (DSL): They invented a set of rules for describing furniture placement. A sentence in this language looks something like: place sofa1 orient:90.0 wall5 wall_center_left:0.028 distance:0.000. This tells the computer: "Put sofa number 1, turned 90 degrees, anchored to wall number 5, aligned slightly to the left of the center, with zero distance from the wall." This is much more stable than a raw number because it describes how the furniture relates to the room, not just where it is in a void.
The Training Data (3D-FrontDSL): To teach the AI this language, they couldn't just use old data. They had to create a new dataset called 3D-FrontDSL. They took thousands of real 3D room designs and used a computer program to translate the furniture coordinates into these new DSL sentences. This created a massive library of "Room + DSL Sentence" pairs for the AI to study.
The Two-Step Training:
- Step 1 (Supervised Fine-Tuning): They first taught the AI to read the room information (like "this is a living room with a door on the north wall") and output the correct DSL sentences. This was like teaching the AI the vocabulary and grammar of interior design.
- Step 2 (Reinforcement Learning): This is where the AI gets a "coach." The team created a set of rules (rewards) to check if the AI's design actually works. Did the furniture crash into the wall? (Collision). Is it blocking the door? (Forbidden Placement). Is there enough space to walk around? (Reachability). If the AI makes a mistake, it gets a "bad score." If it makes a good layout, it gets a "good score." The AI then practices over and over, trying to get a better score, learning to avoid mistakes just like a student learning from a teacher's red pen.

What They Found

The results were quite impressive. The team tested their new system, LayoutDSL, against some of the biggest and most powerful AI models available (including models with hundreds of billions of parameters). Even though their model was much smaller (only 4 billion parameters), it performed better than the giants.

Better Logic: The AI using LayoutDSL created layouts that were much more logical. It rarely blocked doors or windows, and the furniture felt like it belonged in the room.
Higher Success Rate: While other models sometimes failed to produce a usable layout (getting stuck or generating gibberish), LayoutDSL succeeded 100% of the time in their tests.
The Power of the Language: When they compared the new language-based method to the old "guess the number" method, the language approach won hands down. The paper suggests that by forcing the AI to think in relationships (DSL) rather than raw numbers, it learned the "reasoning" behind design much faster and more effectively.

Why It Matters

This paper suggests that the future of AI design isn't just about making bigger computers that can do more math. It's about teaching them to think in the right way. By using a structured language that mimics how humans describe space, the AI can understand the rules of a room. This means we might soon see AI tools that can help us redesign our homes, ensuring our furniture fits perfectly, our doors can open, and our rooms look great, all without the computer needing to be a supercomputer. The authors show that with the right "language" and a little bit of practice (reinforcement learning), even a smaller AI can become a master interior designer.

Technical Summary: Learning an Interior Layout Policy in a Domain-Specific Language Action Space

Problem Statement

Indoor scene layout generation aims to produce physically feasible and functionally meaningful furniture arrangements given a room's structural constraints. The paper identifies two fundamental limitations in existing approaches:

Oversimplified Room Modeling: Many methods reduce rooms to coarse 3D bounding boxes, neglecting critical structural elements like doors, windows, and holes, which are essential for realistic layouts.
Ineffective Reasoning Paradigms: Prior learning-based methods typically formulate layout generation as continuous regression over raw geometric parameters (e.g., direct coordinate prediction). This approach treats layout design as point estimation in Euclidean space, hindering the model's ability to learn the underlying logical reasoning of intelligent design. Furthermore, raw coordinates exhibit high variance across different room geometries, making generalization difficult.

Methodology: LayoutDSL

The authors propose LayoutDSL, a novel framework that shifts the paradigm from direct parameter prediction to policy learning within a structured, interpretable Domain-Specific Language (DSL) action space. The framework consists of three core components:

1. Layout DSL Design

The authors design a DSL that symbolically encodes spatial relationships and geometric constraints. Instead of predicting $(x, y, z)$ coordinates directly, the model generates natural-language-like statements. A single statement follows the schema:
action + instance + orient:θ + anchor + align:δ + distance:d

Anchor: Can be a wall segment (e.g., wall1) or a previously placed furniture item (e.g., bed0).
Alignment: Specifies the relative spatial relationship (e.g., wall_center_left, obj_front_center) with 26 predefined options.
Distance: A scalar offset from the anchor.

This representation incorporates structural priors, resulting in more stable positional descriptions compared to raw coordinates. A bidirectional Generator-Interpreter system converts between 3D layout parameters and DSL statements.

2. Dataset Construction: 3D-FrontDSL

To train the model, the authors constructed 3D-FrontDSL, a dataset derived from the 3D-Front dataset.

Preprocessing: Irregular floor plans are decomposed into rectangular "meta-rooms" using computational geometry.
Annotation: The dataset includes room-structure annotations (walls, doors, windows, holes) paired with synthetic DSL action sequences generated by the DSL generator.
Scale: The dataset contains 9,305 meta-room layout pairs.

3. Training Pipeline

The framework employs a two-stage training process:

Supervised Fine-Tuning (SFT): The LLM is aligned with the syntax and semantics of the Layout DSL using the 3D-FrontDSL dataset. The model learns to generate structured DSL action sequences conditioned on room structural information.
Reinforcement Learning (RL) with Verifiable Rewards: To enhance physical plausibility and design logic, the policy is optimized using Group Relative Policy Optimization (GRPO). The reward function is composed of five verifiable metrics grounded in interior design principles:
- Collision (COL): Penalizes overlapping bounding boxes.
- Out-of-Bounds (OOB): Penalizes furniture extending beyond room boundaries.
- Forbidden-Placement (FP): Penalizes placing furniture in restricted zones (e.g., blocking doors or window fronts).
- Route Reachability (REA): Ensures global connectivity and object accessibility.
- Space Logicality (LOG): Evaluates functional consistency (e.g., bed orientation relative to doors) and spatial harmony.
- Note: An anti-hacking penalty is included to prevent the model from generating sparse layouts to evade collision or reachability penalties.

Key Contributions

DSL-Based Policy Learning: The introduction of a layout domain-specific language that transforms indoor scene generation into a policy-learning problem over a structured, interpretable action space, moving away from direct coordinate regression.
3D-FrontDSL Dataset: The creation of a new dataset featuring room-structure annotations (including doors and windows) paired with synthetic DSL statements, enabling structure-conditioned layout generation.
Verifiable RL Rewards: The proposal of a comprehensive reward system incorporating geometric feasibility and human-aligned design principles, demonstrating that RL significantly improves layout quality.

Experimental Results

The authors evaluated LayoutDSL on RealHome60, a test set of 60 real-world residential rooms.

Performance vs. Baselines: A 4B-parameter LLM (Qwen3-4B-Instruct) fine-tuned with LayoutDSL and RL achieved a mean score of 0.878 across five metrics (COL, OOB, FP, REA, LOG).
Comparison with Large Models: This 4B model outperformed much larger, unfine-tuned industry-leading LLMs (e.g., Gemini-3.1, GPT-5.1) in the few-shot setting, surpassing Gemini-3.1 by 26.4 percentage points.
Comparison with Existing Methods: LayoutDSL outperformed the strongest existing baseline (I-Design) by 28.4 points, particularly excelling in structural constraint handling (FP and LOG metrics).
Ablation Studies:
- DSL vs. Direct Parameters: DSL-based generation consistently outperformed direct parameter prediction in both SFT and RL stages.
- SFT and RL: SFT improved the mean score from 0.357 (baseline) to 0.685, while RL further boosted it to 0.878.
- Anti-Hacking: Removing the anti-hacking penalty degraded performance, confirming its necessity for realistic object density.

Significance and Claims

The paper claims that by projecting spatial computation into a DSL action space, LayoutDSL enables LLMs to learn the reasoning logic of interior design rather than merely memorizing coordinate distributions. The authors assert that this approach:

Substantially improves spatial plausibility and design logicality.
Allows smaller models (4B) to outperform larger, untrained models by leveraging structured reasoning and verifiable feedback.
Provides a scalable and interpretable framework for automated interior design, with potential applications in simulation-based data generation and commercial scene synthesis.

The authors conclude that while the current work focuses on structured rooms, future directions include handling irregular room geometries and diverse user requirements, suggesting that large-scale RL could further facilitate deployment in complex intelligent design applications.