TartanGround Dataset

Welcome to the TartanGround dataset documentation — a large-scale dataset for ground robot perception and navigation.

Quick Links

🌐 Dataset Webpage	TartanGround
📄 Paper	arXiv:2505.10696
💻 GitHub Repository	castacks/tartanairpy
📊 Metadata	Google Sheet

Installation

# 1. Create and activate conda environment
conda create -n tartanground python=3.9
conda activate tartanground

# 2. Clone repository with all submodules
git clone --recursive git@github.com:castacks/tartanairpy.git
cd tartanairpy

# 3. Ensure submodules are up to date
git submodule update --init --recursive

# 4. Install the package
 pip install -e .

Note

Make sure you have git and conda installed on your system before proceeding.

Dataset Structure

The TartanGround dataset is organized hierarchically by environment and robot type. Each environment contains data for multiple robot platforms with comprehensive sensor modalities.

Directory Layout

TartanGround_Root/
├── AbandonedCable/
│   ├── AbandonedCable_rgb.pcd      # Global RGB point cloud
│   ├── AbandonedCable_sem.pcd      # Global Semantic point cloud
│   ├── seg_label_map.json          # Semantic segmentation label map
│   ├── Data_omni/                  # Omnidirectional robot data
│   │   ├── P0000/
│   │   │   ├── image_lcam_front/
│   │   │   ├── depth_lcam_front/
│   │   │   ├── seg_lcam_front/
│   │   │   ├── imu/
│   │   │   ├── lidar/
│   │   │   ├── pose_lcam_front.txt
│   │   │   ├── P0000_metadata.json
│   │   │   ├── image_lcam_left/
│   │   │   └── ...
│   │   └── P00XX/
│   ├── Data_diff/                  # Differential drive robot data
│   │   ├── P1000/
│   │   └── P10XX/
│   └── Data_anymal/                # Quadrupedal robot data
│       ├── P2000/
│       └── P20XX/
├── AbandonedFactory/
│   └── (same structure as above)
└── ...

Robot Platforms

Supported Robot Types

Robot	Type	Trajectory IDs	Description
omni	Omnidirectional	P0000, P0001, …	Holonomic movement in all directions
diff	Differential Drive	P1000, P1001, …	Differential wheeled robot
anymal	Quadrupedal	P2000, P2001, …	Legged robot for complex terrains

Camera Configuration

Each robot is equipped with a stereo 6-cam setup providing full 360° coverage (similar to Tartanair-v2: Modalities).

Camera Specifications

Parameter	Value
Camera Positions	front, left, right, back, top, bottom
Field of View	90° (each camera)
Resolution	640×640 pixels
Stereo Configuration	Left (lcam_*) and Right (rcam_*) pairs available

Available Camera Names

# Left cameras
['lcam_front', 'lcam_right', 'lcam_left', 'lcam_back', 'lcam_top', 'lcam_bottom']

# Right cameras
['rcam_front', 'rcam_right', 'rcam_left', 'rcam_back', 'rcam_top', 'rcam_bottom']

Sensor Modalities

The dataset provides multi-modal sensor data to support various robotic perception tasks:

• RGB Images (image): Color images for visual perception

• Depth Maps (depth): Accurate depth information for 3D scene understanding

• Semantic Segmentation (seg): Pixel-wise semantic labels

• IMU Data (imu): Inertial measurement unit data

• LiDAR Point Clouds (lidar): 3D point cloud data (32 Beam simulated LiDAR)

• Robot Poses (meta): Ground truth 6-DOF poses and metadata including robot height

• Global Point Clouds:

  • RGB Point Clouds (rgb_pcd): Colored 3D representations of entire environments

  • Semantic Point Clouds (sem_pcd): Point clouds with semantic labels

• Segmentation Labels (seg_labels): Label mappings for semantic segmentation tasks

• ROS Bags (rosbag): Proprioceptive data with joint states (available for anymal robot only)

Available Modalities

# Complete list of available modalities
['image', 'meta', 'depth', 'seg', 'lidar', 'imu', 'rosbag', 'sem_pcd', 'seg_labels', 'rgb_pcd']

Note

The rosbag modality is only available for the anymal (quadrupedal) robot version.

Download Dataset

The TartanGround dataset can be downloaded using the tartanairpy Python toolkit. The repository includes ready-to-use examples in examples/download_ground_example.py.

Download Examples

Example 1 – Download all modalities for specific environments and robots:

import tartanair as ta

ta.init('/path/to/tartanground/root')

env = ["AbandonedFactory", "ConstructionSite", "Hospital"]

ta.download_ground(
    env = env,
    version = ['omni', 'diff', 'anymal'],
    traj = [],
    modality = [
        'image', 'meta', 'depth', 'seg', 'lidar', 'imu',
        'rosbag', 'sem_pcd', 'seg_labels', 'rgb_pcd'
    ],
    camera_name = ['lcam_front', 'lcam_right', 'lcam_left', 'lcam_back'],
    unzip = False
)

Example 2 – Download one trajectory from each environment (Omnidirectional robot only):

import tartanair as ta

ta.init('/path/to/tartanground/root')

ta.download_ground(
    env = [],
    version = ['omni'],
    traj = ['P0000'],
    modality = [],
    camera_name = ['lcam_front'],
    unzip = False
)

Example 3 – Download semantic occupancy data only:

import tartanair as ta

ta.init('/path/to/tartanground/root')

ta.download_ground(
    env = [],
    version = [],
    traj = [],
    modality = ['seg_labels', 'sem_pcd'],
    camera_name = [],
    unzip = False
)

Example 4 – Download entire dataset (~15 TB):

import tartanair as ta

ta.init('/path/to/tartanground/root')

ta.download_ground(
    env = [],
    version = [],
    traj = [],
    modality = [],
    camera_name = [],
    unzip = False
)

Multi-threaded Download

For faster downloads, use the multi-threaded version:

import tartanair as ta

ta.init('/path/to/tartanground/root')

ta.download_ground_multi_thread(
    env = env,
    version = ['omni', 'diff', 'anymal'],
    traj = [],
    modality = [
        'image', 'meta', 'depth', 'seg', 'lidar', 'imu',
        'rosbag', 'sem_pcd', 'seg_labels', 'rgb_pcd'
    ],
    camera_name = ['lcam_front', 'lcam_right', 'lcam_left', 'lcam_back'],
    unzip = False,
    num_workers = 8
)

Download Process Notes

Important Download Information

• Pre-download Analysis: The script lists all files and calculates total space requirements before downloading

• File Format: Data is downloaded as zip files

• Post-processing: Use examples/unzip_ground_files.py to extract the downloaded data

• Multi-threading: Adjust num_workers parameter based on your system capabilities and network bandwidth

See also

Full Examples: Comprehensive download scripts are available in examples/download_ground_example.py in the GitHub repository.

Dataset Statistics

TartanGround Dataset Overview

Metric	Value
Total Environments	63 diverse scenarios
Total Trajectories	878 trajectories
Robot Platforms	3 types (Omnidirectional, Differential, Quadrupedal)
Trajectory Distribution	440 omni, 198 diff, 240 legged
Total Samples	1.44 million samples
RGB Images	17.3 million images
Dataset Size	~16 TB
Samples per Trajectory	600-8,000 samples
Camera Resolution	640×640 pixels
Data Collection Frequency	10 Hz

Environment Categories

Environment Types

Category	Description
Indoor	Indoor spaces with complex layouts and lighting
Nature	Natural environments with varied vegetation, natural terrain and seasons
Rural	Countryside settings with varied topography
Urban	City environments with structured layouts
Industrial/Infrastructure	Construction sites and industrial facilities
Historical/Thematic	Heritage sites and specialized environments

Semantic Occupancy Maps

The TartanGround dataset supports semantic occupancy prediction research by providing tools to generate local 3D occupancy maps with semantic class labels. These maps are essential for training and evaluating neural networks that predict semantic occupancy from sensor observations.

Use Cases

• Semantic Occupancy Prediction: Train networks to predict 3D semantic structure from 2D observations

• 3D Scene Understanding: Evaluate spatial reasoning capabilities of perception models

• Navigation Planning: Generate semantically-aware path planning datasets

• Multi-modal Fusion: Combine RGB, depth, and LiDAR data for 3D semantic mapping

Workflow Overview

The semantic occupancy map generation follows a two-step process:

Step 1: Generation - Extract local occupancy maps around each robot pose

Step 2: Visualization - Inspect and validate the generated occupancy maps

Step 1: Generate Semantic Occupancy Maps

Use the GPU-accelerated script to extract local semantic occupancy maps from global point clouds:

# Generate occupancy maps for a specific trajectory
python examples/subsample_semantic_pcd_gpu.py \
    --root_dir /path/to/tartanground/dataset \
    --env ConstructionSite \
    --traj P0000

Key Parameters:

# Customize occupancy map properties
python examples/subsample_semantic_pcd_gpu.py \
    --root_dir /path/to/dataset \
    --env ConstructionSite \
    --traj P0000 \
    --resolution 0.1 \                    # Voxel size in meters
    --x_bounds -20 20 \                   # Local X bounds [min, max]
    --y_bounds -20 20 \                   # Local Y bounds [min, max]
    --z_bounds -3 5 \                     # Local Z bounds [min, max]
    --subsample_poses 10                  # Process every 10th pose

Output:

The script generates .npz files in {trajectory}/sem_occ/ containing:

• occupancy_map: 3D voxel grid with semantic class IDs

• pose: Robot pose [x, y, z, qx, qy, qz, qw]

• bounds: Local coordinate bounds

• resolution: Voxel resolution in meters

• class_mapping: Semantic class ID to RGB color mapping

GPU Requirements

This script requires:

• NVIDIA GPU with CUDA support

• CuPy library for GPU acceleration

• Sufficient GPU memory (8GB+ recommended for large environments)

Step 2: Visualize Occupancy Maps

Use the interactive viewer to inspect generated occupancy maps and verify their quality:

# Visualize occupancy maps with navigation controls
python examples/visualize_semantic_occ_local.py \
    --root_dir /path/to/tartanground/dataset \
    --env ConstructionSite \
    --traj P0000 \
    --skip_samples 100

Customization Options:

# Customize visualization appearance
python examples/visualize_semantic_occ_local.py \
    --root_dir /path/to/dataset \
    --env ConstructionSite \
    --traj P0000 \
    --skip_samples 50 \                   # Show every 50th occupancy map
    --point_size 12.0 \                   # Larger point visualization
    --background white                    # White background

Dataset Integration

File Structure:

After running the generation script, your dataset structure will include:

TartanGround_Root/
├── ConstructionSite/
│   ├── ConstructionSite_sem.pcd          # Global semantic point cloud (input)
│   ├── Data_omni/
│   │   ├── P0000/
│   │   │   ├── pose_lcam_front.txt       # Robot poses (input)
│   │   │   ├── sem_occ/                  # Generated occupancy maps
│   │   │   │   ├── semantic_occupancy_000000.npz
│   │   │   │   ├── semantic_occupancy_000001.npz
│   │   │   │   └── ...
│   │   │   └── (other sensor data)
│   │   └── P0001/
│   └── seg_label_map.json                # Semantic class names

Loading Occupancy Maps in Python:

import numpy as np

# Load a single occupancy map
data = np.load('semantic_occupancy_000000.npz', allow_pickle=True)

occupancy_map = data['occupancy_map']     # 3D array with class IDs
pose = data['pose']                       # Robot pose [x,y,z,qx,qy,qz,qw]
bounds = data['bounds']                   # Local bounds [x_min,x_max,y_min,y_max,z_min,z_max]
resolution = data['resolution']           # Voxel size in meters
class_mapping = data['class_mapping']     # Class ID to RGB mapping

print(f"Occupancy map shape: {occupancy_map.shape}")
print(f"Resolution: {resolution}m per voxel")
print(f"Occupied voxels: {np.sum(occupancy_map > 0)}")

See also

Complete Examples: Full parameter documentation and advanced usage examples are available in the GitHub repository examples folder.

Citation

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

@article{patel2025tartanground,
  title={TartanGround: A Large-Scale Dataset for Ground Robot Perception and Navigation},
  author={Patel, Manthan and Yang, Fan and Qiu, Yuheng and Cadena, Cesar and Scherer, Sebastian and Hutter, Marco and Wang, Wenshan},
  journal={arXiv preprint arXiv:2505.10696},
  year={2025}
}

Support and Contact

For technical issues, questions, or bug reports, please open an issue on the GitHub repository.

For applications and interesting dataset uses, visit the dataset webpage.