UCLA-VAIL/NavBenchGS

NavBench-GS: 3D Gaussian Splatting Scenarios for Visual Navigation

NavBench-GS is a collection of real-world 3D Gaussian Splatting (3DGS) scenarios for benchmarking visual navigation models. Each scenario contains a reconstructed outdoor scene with paired USD meshes, 3DGS point clouds, and navigation waypoints.

Paper: S2E: From Seeing to Experiencing (ICLR 2026)

Project Page: https://vail-ucla.github.io/S2E/

Dataset Structure

NavBenchGS/
├── vid2sim_raw/                    # USD mesh scenarios
│   ├── 0000/
│   │   ├── tsdf_fusion_post.usd   # Reconstructed mesh (Isaac Sim compatible)
│   │   ├── tsdf_fusion_post.ply   # PLY mesh
│   │   ├── config.yaml            # Scenario configuration
│   │   └── point_cloud.ply        # Raw point cloud
│   ├── 0001/
│   └── ...
├── vid2sim_torch/                  # 3DGS point clouds (PyTorch format)
│   ├── 0000/                      # gsplat-compatible 3DGS data
│   ├── 0001/
│   └── ...
└── vid2sim_starting_ending_position.json   # Start/target waypoints per scenario

Usage

With S2E / URBAN-SIM (Isaac Sim)

# Clone S2E
git clone https://github.com/VAIL-UCLA/S2E.git

# Download this dataset to S2E/assets/NavBenchGS/
# Then run navigation in a scenario:
python vid2sim/main.py --scenario_id 0 --nav_mode forward --enable_cameras

Scenario Info

Each scenario has start and target positions defined in vid2sim_starting_ending_position.json:

{
  "0": {"start_x": 0.0, "start_y": -20.0, "target_x": 0.0, "target_y": 2.0},
  "1": {"start_x": ..., "start_y": ..., "target_x": ..., "target_y": ...},
  ...
}

Requirements

• NVIDIA Isaac Sim 5.x
• IsaacLab extension (3DGS camera support)
• gsplat for Gaussian Splatting rendering

Citation

@inproceedings{he2025seeing,
  title={From Seeing to Experiencing: Scaling Navigation Foundation Models with Reinforcement Learning},
  author={Honglin He and Yukai Ma and Brad Squicciarini and Wayne Wu and Bolei Zhou},
  booktitle={International Conference on Learning Representations},
  year={2026}
}

License

Apache 2.0