superbigsaw/MansionWorld

Dataset

MansionWorld

10

9 commits

2 linked in READMEs

updated Mar 14, 2026

See the code

README

MansionWorld

MansionWorld is a large-scale multi-floor indoor scene dataset for embodied AI research, built on top of the AI2-THOR simulator.

Key Features

  • 1,000 interactive multi-floor buildings with 2–10 floors each
  • 10,000+ rooms in total across all buildings
  • Diverse building types: Residential (50%), Office (30%), Public (20%)

Download

The dataset is provided as zip archives for easier downloading:

FileSizeDescription
mansionworld.zip~2.9 GBAll 1,000 building scenes
mansion_patch.zip~789 MBAdditional assets (elevators, stairs, etc.)

Dataset Structure

After extracting mansionworld.zip:

mansionworld/
β”œβ”€β”€ {building_type}_{floors}f_{area}_fp{id}#{variant}/
β”‚   β”œβ”€β”€ floor_1.json      # Floor configuration (doors, walls, objects, rooms)
β”‚   β”œβ”€β”€ floor_1.png       # Floor plan visualization
β”‚   β”œβ”€β”€ floor_2.json
β”‚   β”œβ”€β”€ floor_2.png
β”‚   └── ...
└── ...

Building Naming Convention

Example: office_corporate_hq_6f_300_fp001#1

  • office_corporate_hq: Building type (office - corporate headquarters)
  • 6f: Number of floors (6 floors)
  • 300: Approximate floor area (mΒ²)
  • fp001: Floorplan template ID
  • #1: Variant index

Usage

For loading MansionWorld scenes and using the Mansion API, please refer to the documentation.

Quick Download

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

Citation

If you use MansionWorld in your research, please cite:

@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}, 
}

License

This dataset is released under the CC BY 4.0 License.

3d-scene
ai2thor
embodied-ai
indoor-environment
multi-floor
simulation

Contributors

superbigsaw

6 commits

车李

superbigsaw/MansionWorld

Dataset

MansionWorld

10

9 commits

2 linked in READMEs

updated Mar 14, 2026

See the code

README

MansionWorld

MansionWorld is a large-scale multi-floor indoor scene dataset for embodied AI research, built on top of the AI2-THOR simulator.

Key Features

  • 1,000 interactive multi-floor buildings with 2–10 floors each
  • 10,000+ rooms in total across all buildings
  • Diverse building types: Residential (50%), Office (30%), Public (20%)

Download

The dataset is provided as zip archives for easier downloading:

FileSizeDescription
mansionworld.zip~2.9 GBAll 1,000 building scenes
mansion_patch.zip~789 MBAdditional assets (elevators, stairs, etc.)

Dataset Structure

After extracting mansionworld.zip:

mansionworld/
β”œβ”€β”€ {building_type}_{floors}f_{area}_fp{id}#{variant}/
β”‚   β”œβ”€β”€ floor_1.json      # Floor configuration (doors, walls, objects, rooms)
β”‚   β”œβ”€β”€ floor_1.png       # Floor plan visualization
β”‚   β”œβ”€β”€ floor_2.json
β”‚   β”œβ”€β”€ floor_2.png
β”‚   └── ...
└── ...

Building Naming Convention

Example: office_corporate_hq_6f_300_fp001#1

  • office_corporate_hq: Building type (office - corporate headquarters)
  • 6f: Number of floors (6 floors)
  • 300: Approximate floor area (mΒ²)
  • fp001: Floorplan template ID
  • #1: Variant index

Usage

For loading MansionWorld scenes and using the Mansion API, please refer to the documentation.

Quick Download

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

Citation

If you use MansionWorld in your research, please cite:

@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}, 
}

License

This dataset is released under the CC BY 4.0 License.

3d-scene
ai2thor
embodied-ai
indoor-environment
multi-floor
simulation

Contributors

superbigsaw

6 commits

车李