English | 中文

InternScenes: A Large-scale Interactive Indoor Scene Dataset with Realistic Layouts

     

🏡 Introduction

InternScenes comprises approximately 40,000 diverse scenes and 1.96M 3D objects that cover 15 common scene types and 288 object classes, which is roughly 10 times larger than existing datasets.

💡 Highlights

Existing 3D scene datasets often suffer from:

• ❌ Limited diversity or simulatability
• ❌ Sanitized layouts lacking small items
• ❌ Severe object collisions

Accordingly, InternScenes integrates a wide variety of scenes, and particularly, preserves small items for complex layouts, resolve collisions, and further incorporates interactive objects, thus ensures:

• 📊 Large scale: 40,000 diverse scenes including 1.96M 3D objects covering 288 object classes.
• 🚪 Realistic layouts: preserving massive small objects to strictly align with real-world scanned scenes.
• 🕹️ Interactivity: 20% interactive objects inside covering 16 common types, such as cabinets, microwaves, ovens, and fridges.

Which tasks will benefit from our dataset?

• ✅ 3D scene reconstruction
• ✅ 3D scene understanding
• ✅ Scene layout generation
• ✅ Embodied navigation

What's included to support these tasks?

1. Convenient and efficient scene rendering scripts;
2. Detailed object semantic information;
3. Unified format and coordinate system for scene layouts;
4. Various trajectories for embodied navigation.

For the usage of our dataset, please refer to the tutorials.

🪄 News

• 2025/07 InternScenes-Real2Sim v1.0 released.

📋 Table of Contents

• 🏡 Introduction
• 💡 Highlights
• 🪄 News
• 📋 Table of Contents
• ⚙️ Getting Started
• 📖 Tutorial
• 📋 TODO List
• 🧷 Citation
• 📄License
• 🥰 Acknowledgements

⚙️ Getting Started

Installation

1. Clone this repository.

git https://github.com/InternRobotics/InternScenes.git
cd InternScenes

2. Create an environment and install basic dependencies.

conda create -n internscenes python=3.10 -y  
conda activate internscenes
pip install -r requirements.txt

3. (optional) Install Isaac-Sim 4.1.0 for rendering *.usd files and convertion *.glb files.

# Make sure your conda environment is activated.
pip install torch==2.4.0 --index-url https://download.pytorch.org/whl/cu118
pip install isaacsim==4.1.0 isaacsim-extscache-physics==4.1.0 isaacsim-extscache-kit==4.1.0 isaacsim-extscache-kit-sdk==4.1.0 --extra-index-url https://pypi.nvidia.com

Data Preparation

Please refer to the data guide for downloading and organizing the dataset.

InternScenes uses two data groups:

1. Real2Sim layout data: data/Layout_info/{dataset}/{scan_id}/ stores the scene structure mesh and layout.json annotations. See the Real2Sim data guide for the full directory layout and layout schema.
2. Trajectory and rendering data: trajectory_tools reads USD scene assets from data/source/{dataset}/part{part}/{usd}_usd/, uses rendering resources from data/env/, and writes trajectories, caches, and rendered frames under output/. See Trajectory and Rendering Data and the trajectory tools for details.

📖 Tutorial

We provide a simple tutorial here as a guideline for the visualization and basic usage of our dataset. Welcome to try and post your suggestions!

Trajectory Generation & Rendering

For generating camera trajectories and rendering scenes in Isaac Sim, please refer to the trajectory tools.

📋 TODO List

• Release the InternScenes-Real2Sim.
• Release trajectories for each scene and rendering scripts.
• Release the paper.
• Polish the codes of building the InternScenes-Real2Sim.
• Polish the codes of building the InternScenes-Synthetic.
• Release the InternScenes-Synthetic.

🧷 Citation

@inproceedings{InternScenes,
  title={InternScenes: A Large-scale Interactive Indoor Scene Dataset with Realistic Layouts},
  author={Zhong, Weipeng and Cao, Peizhou and Jin, Yichen and Li, Luo and Cai, Wenzhe and Lin, Jingli and Lyu, Zhaoyang and Wang, Tai and Dai, Bo and Xu, Xudong and Pang, Jiangmiao},
  year={2025},
  booktitle={arXiv},
}

📄License

This work is under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

🥰 Acknowledgements

• EmbodiedScan: The scenes we retrieve are based on the annotations from EmbodiedScan, which include a large number of 9-DoF bounding boxes for small objects.
• InternUtopia (Previously GRUtopia ): Some of the high-quality 3D asset files (*.usd) in this repository are sourced from here.
• Hunyuan3D-2.1: The textures for some of the 3D assets were generated using this model.
• HSSD: A curated selection of object assets from this project has been used to construct our asset library.
• PartNet-Mobility: A curated selection of object assets from this project has been used to construct our asset library.
• Objaverse: A curated selection of object assets from this project has been used to construct our asset library.