MANSION

MANSION: Multi-floor lANguage-to-3D Scene generatIOn for loNg-horizon tasks

LLM-driven procedural multi-floor building generation. Given a natural-language description (e.g., "2-story office building with conference rooms"), Mansion generates complete building layouts including room segmentation, walls, doors, windows, furniture placement, lighting, and top-down rendered images.

🎉 MANSION has been accepted to CVPR 2026!

Installation

1. Create a virtual environment

conda create -n mansion python=3.10 -y
conda activate mansion

2. Install Python dependencies

cd mansion
pip install -r requirements.txt

# Install the required AI2-THOR version
pip install --extra-index-url https://ai2thor-pypi.allenai.org ai2thor==0+6f165fdaf3cf2d03728f931f39261d14a67414d0

3. Download scene assets (first run)

The project requires AI2-THOR and Objaverse-THOR model assets:

python -m objathor.dataset.download_holodeck_base_data --version 2023_09_23
python -m objathor.dataset.download_assets --version 2023_09_23
python -m objathor.dataset.download_annotations --version 2023_09_23
python -m objathor.dataset.download_features --version 2023_09_23

If the download is interrupted, you can manually download and extract the assets to ~/.objathor-assets/2023_09_23/:

https://pub-daedd7738a984186a00f2ab264d06a07.r2.dev/2023_09_23/assets.tar

4. Install Mansion patch assets (required)

Download mansion_patch.zip from our HuggingFace repository:

https://huggingface.co/datasets/superbigsaw/MansionWorld

Place the downloaded mansion_patch.zip in the project root directory (next to setup_mansion.py), then run:

python setup_mansion.py

This script will automatically extract the patch and install AI2-THOR local release, door database, annotations, and extra 3D assets to the correct locations.

LLM Configuration

Before running the pipeline, configure your LLM API credentials. We support OpenAI-compatible APIs, Azure OpenAI, and custom model endpoints. See config/constants.py for all available options.

# Example: OpenAI-compatible API
export OPENAI_API_KEY="your-api-key"
export OPENAI_API_BASE="https://api.openai.com"  # optional, for proxy/custom endpoints

Then specify llm_provider in mansion_quickstart.py:

cfg = make_config(
    requirement="...",
    llm_provider="openai",   # "openai" | "azure" | "mixed"
)

• "openai" — All nodes use OPENAI_CONFIG from config/constants.py
• "azure" — All nodes use AZURE_CONFIG from config/constants.py
• "mixed" — Per-node LLM routing via config/node_config.json, supporting different models per pipeline stage

Quick Start

python mansion_quickstart.py

MansionWorld

MansionWorld is our dataset of 1,000 pre-generated multi-floor buildings. You can download it directly from HuggingFace without running the full MANSION generation pipeline:

https://huggingface.co/datasets/superbigsaw/MansionWorld

Mansion Gym

Mansion Gym (mansion-api) is a gymnasium-compatible environment built on top of MansionWorld, providing cross-floor embodied AI capabilities — navigation, object interaction, and long-horizon task execution across multiple floors.

Installation

pip install mansion-api

Quick Start

If you haven't set up the full MANSION pipeline yet, download the MansionWorld dataset from HuggingFace and open mansionworld_quickstart.ipynb to explore the key skills and interfaces provided by Mansion Gym.

For full API documentation, see the project page:

https://agibotgeneral.github.io/mansion-site/

Citation

If you find this work useful, please cite our paper:

@misc{che2026mansionmultifloorlanguageto3dscene,
      title={MANSION: Multi-floor lANguage-to-3D Scene generatIOn for loNg-horizon tasks},
      author={Lirong Che and Shuo Wen and Shan Huang and Chuang Wang and Yuzhe Yang and Gregory Dudek and Xueqian Wang and Jian Su},
      year={2026},
      eprint={2603.11554},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2603.11554},
}

Acknowledgements

This repository includes code adapted from Holodeck (allenai/Holodeck), licensed under Apache-2.0. We modified the original implementation for scene instantiation and object placement. See the LICENSE file and third-party notices for attribution details.

License

This project is licensed under the Apache License 2.0. See LICENSE for details.