spatialverse/SAGE-3D_Collision_Mesh

SAGE-3D Collision Mesh: Physics-Enabled Collision Bodies for 3D Gaussian Scenes

Paper | Project Page | Code

High-precision collision geometry dataset extracted from 1,000 indoor Mesh scenes, enabling physically accurate navigation and interaction in virtual environments.

Collision Mesh of InteriorGS data captured on Issac Sim 5.0.

---

📢 News

• 2025-12-15: Released SAGE-3D Collision Mesh dataset with collision bodies for 1000 InteriorGS scenes.

---

📋 Overview

3D Gaussian Splatting (3DGS) has emerged as a powerful representation for photorealistic scene rendering. However, a critical limitation prevents its use in interactive applications: 3DGS lacks physical collision geometry, making it impossible for embodied agents to interact realistically with the environment. Without collision detection, agents can pass through walls, float through furniture, or exhibit other physically implausible behaviors.

To address this challenge, we present SAGE-3D Collision Mesh, a comprehensive dataset of physics-enabled collision bodies extracted from the InteriorGS dataset. This dataset provides:

• 554,000+ collision bodies corresponding to object instances across 1000 indoor scenes
• Convex hull decomposition for efficient and accurate collision detection
• USD format compatible with modern physics simulation platforms (Isaac Sim, Omniverse, etc.)
• Tight integration with InteriorGS 3DGS scenes for seamless visual-physical alignment

Each collision body is carefully decomposed into convex hulls, ensuring computational efficiency while maintaining geometric fidelity. This enables:

Real-time collision detection - Optimized convex hulls enable fast physics simulation
Embodied AI research - Support for manipulation, navigation, and interaction tasks
Sim-to-real transfer - Realistic physics simulation reduces the sim-to-real gap

Key Features

Feature	Description
Scenes	1000 indoor environments (homes, offices, museums, restaurants, etc.)
Objects	554,000+ collision body instances
Format	USD (Universal Scene Description)
Compatibility	Isaac Sim 5.0+, NVIDIA Omniverse

---

🗂️ Dataset Structure

Collision_Mesh/
├── 839873/                    # Scene-specific collision data
│   └── 839873_collision.usd   # Collision mesh in USD format
├── 839874/
│   └── 839874_collision.usd
├── 839875/
│   └── 839875_collision.usd
└── ...

---

💡 Sample Usage

This dataset provides physics-enabled collision bodies in USD format, which are crucial for enabling physically accurate navigation and interaction in virtual environments. These meshes are primarily used to integrate with 3D Gaussian Splatting (3DGS) scenes (e.g., from spatialverse/SAGE-3D_InteriorGS_usdz) to create complete USDA scene files compatible with physics simulation platforms like NVIDIA Isaac Sim.

1. Download the SAGE-3D Collision Mesh Dataset

First, clone this dataset repository using git lfs:

git lfs install
git clone https://huggingface.co/datasets/spatialverse/SAGE-3D_Collision_Mesh

2. Prepare Isaac Sim Environment

The full SAGE-3D project, including the use of these collision meshes, relies on NVIDIA Isaac Sim. Please refer to the official SAGE-3D GitHub repository for detailed instructions on setting up Isaac Sim (version 5.0+).

3. Build USDA Scene Files with Collision Data

Once you have downloaded this collision mesh dataset and obtained the USDZ scene files (e.g., from spatialverse/SAGE-3D_InteriorGS_usdz), you can integrate the collision bodies into USDA scene files using the sage3d_usda_builder.py script provided in the SAGE-3D Official repository.

# Example command to build USDA Scene Files
python Code/benchmark/scene_data/sage3d_usda_builder.py \
    --usdz-dir /path/to/usdz \
    --out-dir /path/to/output/usda \
    --template Data/template.usda \
    --usdz-placeholder "@usdz_root[gauss.usda]@" \
    --collision-placeholder "@collision_root@" \
    --usdz-path-template "/path/to/usdz/{scene_id}.usdz[gauss.usda]" \
    --collision-path-template "/path/to/Collision_Mesh/{scene_id}/{scene_id}_collision.usd" \
    --overwrite

Parameters:

• --usdz-dir: Directory containing USDZ scene files (e.g., from spatialverse/SAGE-3D_InteriorGS_usdz)
• --out-dir: Output directory for USDA files
• --template: Path to USDA template file (found in the SAGE-3D_Official GitHub repository)
• --usdz-placeholder: Placeholder string in template for USDZ reference (default: @usdz_root[gauss.usda]@)
• --collision-placeholder: Placeholder string in template for collision payload (default: @collision_root@)
• --usdz-path-template: Template for USDZ reference path (use {scene_id} as placeholder)
• --collision-path-template: Template for collision payload path (use {scene_id} as placeholder to point to files in this dataset)
• --overwrite: Overwrite existing USDA files

---

🔗 Related Datasets

This dataset is part of the SAGE-3D project:

1. InteriorGS: Source 3DGS scenes with semantic annotations
   → spatialverse/InteriorGS

2. SAGE-3D InteriorGS USDZ: USDZ format scenes for Isaac Sim
   → spatialverse/SAGE-3D_InteriorGS_usdz

3. SAGE-3D Collision Mesh (This dataset): Physics-enabled collision bodies
   → spatialverse/SAGE-3D_Collision_Mesh

4. SAGE-3D VLN Data: Navigation trajectories and instructions
   → spatialverse/SAGE-3D_VLN_Data

---

📄 License

This dataset is released under CC-BY-NC-4.0.

---

📜 Citation

If you use SAGE-3D Collision Mesh in your research, please cite our paper:

@misc{miao2025physicallyexecutable3dgaussian,
  title={Towards Physically Executable 3D Gaussian for Embodied Navigation}, 
  author={Bingchen Miao and Rong Wei and Zhiqi Ge and Xiaoquan Sun and Shiqi Gao and Jingzhe Zhu and Renhan Wang and Siliang Tang and Jun Xiao and Rui Tang and Juncheng Li},
  year={2025},
  eprint={2510.21307},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2510.21307}, 
}

Please also cite the InteriorGS dataset:

@misc{InteriorGS2025,
  title        = {InteriorGS: A 3D Gaussian Splatting Dataset of Semantically Labeled Indoor Scenes},
  author       = {SpatialVerse Research Team, Manycore Tech Inc.},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/datasets/spatialverse/InteriorGS}}
}

---

SAGE-3D: Semantically and Physically-Aligned Gaussian Environments for 3D Navigation