Voxel51/STONE

Dataset Card for STONE

[image]

STONE is a large-scale multi-modal dataset for off-road 3D traversability prediction, collected by autonomous ground vehicles across four outdoor environments in South Korea. It provides 7,000 keyframes with surround-view imagery from 6 cameras (1904×1200), 128-channel LiDAR scans (230K points), and voxel-level traversability annotations classifying terrain into free, traversable, potentially traversable, and non-traversable regions. Following the nuScenes format, the dataset includes 3D obstacle bounding boxes, ego-pose trajectories, and synchronized multi-sensor data at ~10 Hz. This FiftyOne version contains a stratified sample of 35 scenes (200 frames each) from the full 279-scene collection, organized as grouped samples with 7 slices per keyframe (6 cameras + 1 LiDAR 3D scene).

This is a FiftyOne dataset with 7000 samples.

Installation

If you haven't already, install FiftyOne:

pip install -U fiftyone

Usage

import fiftyone as fo
from huggingface_hub import snapshot_download


# Download the dataset snapshot to the current working directory

snapshot_download(
    repo_id="Voxel51/STONE", 
    local_dir=".", 
    repo_type="dataset"
    )

# Load dataset from current directory using FiftyOne's native format
dataset = fo.Dataset.from_dir(
    dataset_dir=".",  # Current directory contains the dataset files
    dataset_type=fo.types.FiftyOneDataset,  # Specify FiftyOne dataset format
    name="STONE"  # Assign a name to the dataset for identification
)

# Launch the App
session = fo.launch_app(dataset)

STONE — FiftyOne Dataset Card

STONE is a large-scale multi-modal dataset for off-road 3D traversability prediction, collected by an autonomous ground vehicle (UGV) across four outdoor environments in South Korea. The dataset follows the nuScenes format and provides surround-view camera imagery, 128-channel LiDAR scans, and voxel-level traversability annotations.

• —Paper: Park et al., "STONE: A Scalable Multi-Modal Surround-View 3D Traversability Dataset for Off-Road Robot Navigation", ICRA 2026
• —arXiv: https://arxiv.org/abs/2603.09175
• —License: CC BY-NC-ND 4.0 (dataset) · Apache 2.0 (code)
• —Format: nuScenes / Occ3D-nuScenes
• —Project Page: https://konyul.github.io/STONE-dataset/

FiftyOne Dataset Structure

The dataset is a grouped dataset — one group per keyframe, with seven slices:

Sample Fields

These fields are present on every sample across all seven slices.

Identity & Provenance

nuScenes Metadata (matching the official nuScenes guide)

Labels

Traversability Fractions

These fields are on all slices, derived from gts/<scene>/<token>/labels.npz.

LIDAR_TOP .fo3d Scene

Each LIDAR_TOP sample points to a .fo3d scene file containing three stacked point cloud layers:

Camera configuration: defaultCameraPosition = {x: -15, y: 0, z: 10} (15 m behind, 10 m above), up = "Z" (NuScenes Z-up convention), set via dataset.app_config.plugins["3d"].

Traversability Classes

The voxel grid has shape (200, 200, 16) — a 40 m × 40 m × 3.2 m volume centred on the vehicle at 0.2 m resolution. Value 255 = unoccupied.

Citation

@inproceedings{park2026stone,
  title={STONE: A Scalable Multi-Modal Surround-View 3D Traversability Dataset for Off-Road Robot Navigation},
  author={Park, Konyul and Kim, Daehun and Oh, Jiyong and Yu, Seunghoon and Park, Junseo
          and Park, Jaehyun and Shin, Hongjae and Cho, Hyungchan and Kim, Jungho and Choi, Jun Won},
  booktitle={Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)},
  year={2026}
}