Wanderland Dataset

Dataset Description

Wanderland is a large-scale urban dataset designed for geometrically grounded simulation and open-world embodied AI research. The dataset contains diverse urban scenes captured with dual fisheye cameras, providing high-quality data for 3D reconstruction, novel view synthesis, and navigation tasks.

Key Features

• Urban Scenes: Diverse outdoor environments with varying complexity
• Multi-Modal Data: RGB images, depth, 3D point clouds, 3D Gaussian Splatting models
• Camera Data: Fisheye images + undistorted pinhole images (800×800, 90° FOV)
• 3D Reconstructions: COLMAP sparse models + dense point clouds + 3DGS models
• Navigation Data: Isaac Sim compatible scene files (USDZ) + episode configurations
• Public Manifest: Released scenes are listed in wanderland_public_manifest.csv with quality tiers and exact data paths

Supported Tasks

• 3D Reconstruction: Multi-view stereo, structure-from-motion, depth estimation
• Novel View Synthesis: NeRF, 3D Gaussian Splatting, view interpolation
• Embodied AI Navigation: Visual navigation, path planning, sim-to-real transfer
• Scene Understanding: 3D scene parsing, object detection, spatial reasoning

Current Public Release

Metric	Value
Release Versions	v1, v2
Released Scenes	335
Showcase Scenes	91 (49 v1 + 42 v2)
Evaluation-Ready Scenes	113
Training-Ready Scenes	131
Images per Scene	200-1000 (varies)
Image Resolution (Undistorted)	800×800
Image Resolution (Fisheye)	2K
Camera Model	Dual fisheye → Pinhole projection
Point Cloud Size	1-10M points per scene

The manifest is the source of truth for the currently released scenes. It may be updated as new scenes are added or existing scenes are repaired.

The v2 showcase scenes are preview releases focused on processed visual assets. They include 3DGS, mesh, LiDAR, fisheye imagery, camera metadata, and NVS splits, but currently do not include episodes.json or scene.usdz. The released v2 camera poses and geometry use the same Y-up, COLMAP-aligned world convention as the public Wanderland release.

Dataset Structure

Released scenes are organized by release version and quality tier:

Scene directories use opaque scene IDs. Use wanderland_public_manifest.csv and its data_path column as the canonical index.

data/
├── v1/
│   ├── 01_showcase/
│   │   └── <scene_id>/
│   ├── 02_evaluation_ready/
│   │   └── <scene_id>/
│   └── 03_training_ready/
│       └── <scene_id>/
└── v2/
    ├── 01_showcase/
    │   └── <scene_id>/
    └── README.md

v1 scene directories may contain:

<scene_id>/
├── fisheye.tar.gz           # Original fisheye images (JPG, 1920×1080)
├── fisheye_mask.tar.gz      # Validity masks for fisheye images
├── images.tar.gz            # Undistorted images (PNG, 800×800, 90° FOV)
├── images_mask.tar.gz       # Validity masks for undistorted images
├── raw_pcd.ply              # Dense 3D point cloud (PLY format)
├── 3dgs.ply                 # Pre-trained 3D Gaussian Splatting model
├── transforms.json          # Camera parameters (intrinsics + extrinsics)
├── scene.usdz               # Isaac Sim compatible scene file
├── episodes.json            # Navigation episode configurations
├── sparse/                  # COLMAP sparse reconstruction
│   └── 0/
│       ├── cameras.bin      # Camera intrinsics (PINHOLE model)
│       ├── images.bin       # Camera poses (quaternion + translation)
│       └── points3D.bin     # Sparse 3D points
└── nvs_split/               # Train/val splits for novel view synthesis
    ├── train.txt            # Training images (per-scene split)
    └── val.txt              # Validation images (per-scene split)

v2 showcase scene directories contain a smaller processed-asset profile:

<scene_id>/
├── fisheye.tar.gz           # Original fisheye images
├── 3dgs.ply                 # Pre-trained 3D Gaussian Splatting model
├── scene.ply                # Reconstructed mesh
├── transforms.json          # Camera parameters and poses
├── scene_manifest.json      # Per-scene processing metadata
├── lidar/
│   ├── colorized.las        # Colorized LiDAR point cloud
│   └── uncolorized.las      # Raw-color LiDAR point cloud
├── nvs_split/
│   ├── train.txt
│   └── val.txt
├── info/                    # Device and calibration metadata
├── core/                    # Processing logs and odometry files
└── reports/                 # Processing summaries

File Descriptions

Image Data:

• images/: Undistorted pinhole images (800×800, 90° FOV, PNG format)
• images_mask/: Validity masks indicating valid pixel regions
• fisheye/: Original fisheye images (JPG format)
• fisheye_mask/: Validity masks for fisheye images

3D Data:

• raw_pcd.ply: Dense point cloud with RGB colors (PLY format)
• 3dgs.ply: Pre-trained 3D Gaussian Splatting model
• sparse/0/: COLMAP sparse reconstruction (cameras, poses, sparse points)
• scene.ply: Reconstructed mesh in v2 showcase scenes
• lidar/*.las: Processed LiDAR point clouds in v2 showcase scenes

Camera Parameters:

• transforms.json: Complete camera parameters (intrinsics, extrinsics, distortion)
• v1 and v2 camera poses follow the public COLMAP-aligned release convention.

Navigation Data:

• scene.usdz: USD scene file for NVIDIA Isaac Sim
• episodes.json: Navigation episode configurations

Navigation files are included for released v1 scenes where listed. The v2 showcase release currently does not include scene.usdz or episodes.json.

Data Splits:

• nvs_split/: Per-scene image splits for novel view synthesis
• Paper scene-level split files are provided with the GitHub download tool.

Camera Models

Fisheye Camera (Original):

• Distortion: 4-parameter fisheye model (k1, k2, k3, k4)
• Dual camera setup (left + right)

Undistorted Camera (Processed):

• Model: PINHOLE (rectilinear projection)
• Intrinsics: fx=fy=400.0, cx=cy=400.0
• Resolution: 800×800 pixels
• Field of view: 90 degrees

Coordinate System:

• v1 and v2 camera poses follow the public COLMAP-aligned release convention
• Y-up, right-handed world coordinate system
• Units: Meters

Download Instructions

For the current public manifest, use the download tool. It supports both v1 and v2 manifest entries.

# Preview one v2 showcase scene first
python download.py --modality nvs --quality-tier showcase --release-version v2 --count 1 --dry-run

# Download v2 showcase scenes
python download.py --modality nvs --quality-tier showcase --release-version v2

# Download evaluation-ready v1 scenes
python download.py --modality nvs --quality-tier evaluation_ready

We provide a public manifest at wanderland_public_manifest.csv. It lists each released scene with a quality_tier, optional quality_tags, release metadata, and a data_path pointing to the scene directory.

Note that the manifest will be updated as new scenes are added and existing scenes are repaired.

License

This dataset is released under the Apache 2.0 License. See the LICENSE file for details.

Citation

If you use the Wanderland dataset in your research, please cite:

@article{liu2025wanderland,
  title={Wanderland: Geometrically Grounded Simulation for Open-World Embodied AI},
  author={Liu, Xinhao and Li, Jiaqi and Deng, Youming and Chen, Ruxin and Zhang, Yingjia and Ma, Yifei and Guo, Li and Li, Yiming and Zhang, Jing and Feng, Chen},
  journal={arXiv preprint arXiv:2511.20620},
  year={2025}
}

Links

• Paper: arXiv:2511.20620
• Project Page: ai4ce.github.io/wanderland
• GitHub Repository: github.com/ai4ce/wanderland
• Download Tool: Download README