Dataset Card for STONE

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 fiftyone.utils.huggingface import load_from_hub

# Load the dataset
# Note: other available arguments include 'max_samples', etc
dataset = load_from_hub("Voxel51/STONE")

# 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:

Slice	Media type	Content
CAM_FRONT	image	1904 × 1200 JPEG, front-facing camera
CAM_FRONT_LEFT	image	1904 × 1200 JPEG
CAM_FRONT_RIGHT	image	1904 × 1200 JPEG
CAM_BACK	image	1904 × 1200 JPEG
CAM_BACK_LEFT	image	1904 × 1200 JPEG
CAM_BACK_RIGHT	image	1904 × 1200 JPEG
LIDAR_TOP	3d	.fo3d scene (LiDAR + Traversability + Trajectory layers)

Sample Fields

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

Identity & Provenance

Field	Type	Description
channel	StringField	Sensor name: CAM_FRONT, CAM_BACK, …, LIDAR_TOP
sample_token	StringField	nuScenes sample token (shared across all 7 slices in a group)
scene_token	StringField	nuScenes scene token
scene_name	StringField	Human-readable scene ID, e.g. scene-0053
location	StringField	Recording site: siheung_lake, siheung_farmland, siheung_land, kwangmyeong_land
vehicle	StringField	Vehicle ID: n001 – n004
timestamp	IntField	Unix timestamp in microseconds

nuScenes Metadata (matching the official nuScenes guide)

Field	Type	Description
token	StringField	sample_data token for this specific sensor record
ego_pose_token	StringField	Token into ego_pose.json — vehicle pose at this timestamp
calibrated_sensor_token	StringField	Token into calibrated_sensor.json — intrinsics & extrinsics
is_key_frame	BooleanField	Always True (STONE only contains keyframes)
prev	StringField	Previous sample_data token for this sensor (empty at scene start)
next	StringField	Next sample_data token for this sensor (empty at scene end)
sample_prev	StringField	Previous nuScenes sample token in the scene
sample_next	StringField	Next nuScenes sample token in the scene

Labels

Field	Type	Slices	Description
ground_truth	fo.Detections	LIDAR_TOP	3D obstacle annotations. Each fo.Detection carries location=[x,y,z], rotation=[roll,pitch,yaw], dimensions=[l,w,h] in the LiDAR sensor frame, plus num_lidar_pts and instance_token
cuboids	fo.Polylines	cameras	3D bounding boxes projected onto each camera as wireframe outlines using fo.Polyline.from_cuboid(). Filtered to boxes with all corners in front of the camera
ground_truth_2d	fo.Detections	cameras	Flat 2D bounding boxes from the pre-computed bbox_2d field in sample_annotation.json. Normalised [x, y, w, h] in [0, 1] space
terrain	fo.Classification	all	Dominant traversability class in the frame's voxel grid. label ∈ {free, traversable, potentially_traversable, non_traversable}. confidence = fraction of labeled voxels in that class
trajectory_2d	fo.Polylines	cameras	Projected path of the next 30 ego-pose waypoints (~3 seconds ahead) into the camera image plane. Present on ~83% of frames (absent near scene end)

Traversability Fractions

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

Field	Type	Description
pct_free	FloatField	Fraction of labeled voxels classified as Free (class 0)
pct_traversable	FloatField	Fraction classified as Traversable (class 1)
pct_potentially_traversable	FloatField	Fraction classified as Potentially Traversable (class 2)
pct_non_traversable	FloatField	Fraction classified as Non-Traversable (class 3)

---

LIDAR_TOP .fo3d Scene

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

Layer	Shading	Source	Description
LiDAR	height	samples/LIDAR_TOP/*.pcd	230,400-point raw scan from Hesai OT128. Points coloured by Z elevation via the viridis colorscale
Traversability	rgb	samples/VOXEL_OVERLAY/*_voxels.pcd	~140K points from the same scan, coloured by traversability class. Each point's class is looked up from the voxel grid after transforming from LiDAR sensor frame to ego frame
Trajectory	rgb	samples/TRAJECTORY/*_traj.pcd	All 200 ego-pose waypoints for the scene, transformed to the current frame's LiDAR sensor frame. Blue = past · White = current · Yellow = future

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

Class ID	Label	terrain.label value	Colour in viewer
0	Free	free	🟢 green rgb(50, 230, 50)
1	Traversable	traversable	🟡 yellow rgb(230, 230, 50)
2	Potentially Traversable	potentially_traversable	🟠 orange rgb(255, 153, 0)
3	Non-Traversable	non_traversable	🔴 red rgb(230, 25, 25)

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}
}