STONE Dataset: A Scalable Multi-Modal Surround-View 3D Traversability Dataset for Off-Road Robot Navigation

Konyul Park; Daehun Kim; Jiyong Oh; Seunghoon Yu; Junseo Park; Jaehyun Park; Hongjae Shin; Hyungchan Cho

cs.ROMar 10, 2026

STONE Dataset: A Scalable Multi-Modal Surround-View 3D Traversability Dataset for Off-Road Robot Navigation

Authors: Konyul Park, Daehun Kim, Jiyong Oh, Seunghoon Yu, Junseo Park, Jaehyun Park, Hongjae Shin, Hyungchan Cho, +2 more

Organizations: Interdisciplinary Program in Artificial Intelligence, Seoul National University, Seoul, 08826, Korea. · Department of Electrical and Computer Engineering, Seoul National University, Seoul, 08826, Korea.

Abstract

Reliable off-road navigation requires accurate estimation of traversable regions and robust perception under diverse terrain and sensing conditions. However, existing datasets lack both scalability and multi-modality, which limits progress in 3D traversability prediction. In this work, we introduce STONE, a large-scale multi-modal dataset for off-road navigation. STONE provides (1) trajectory-guided 3D traversability maps generated by a fully automated, annotation-free pipeline, and (2) comprehensive surround-view sensing with synchronized 128-channel LiDAR, six RGB cameras, and three 4D imaging radars. The dataset covers a wide range of environments and conditions, including day and night, grasslands, farmlands, construction sites, and lakes. Our auto-labeling pipeline reconstructs dense terrain surfaces from LiDAR scans, extracts geometric attributes such as slope, elevation, and roughness, and assigns traversability labels beyond the robot's trajectory using a Mahalanobis-distance-based criterion. This design enables scalable, geometry-aware ground-truth construction without manual annotation. Finally, we establish a benchmark for voxel-level 3D traversability prediction and provide strong baselines under both single-modal and multi-modal settings. STONE is available at: https://konyul.github.io/STONE-dataset/

Figures & tables

Dataset	Sensors	4D Radar	Camera FOV	# of Cameras	Image Resolution	LiDAR Resolution	Annotation
Freiburg Forest [ 13 ]	Camera, NIR	✗	Front-view	2	1024 $\times$ 768	-	2D semantic
YCOR [ 14 ]	Camera, LiDAR, INS	✗	Front-view	1	1024 $\times$ 544	64	2D semantic
RUGD [ 15 ]	Camera, LiDAR, IMU, GPS	✗	Front-view	1	688 $\times$ 550	32	2D semantic
RELLIS-3D [ 6 ]	Camera, LiDAR, INS	✗	Front-view	1	1920 $\times$ 1200	32, 64	2D/3D semantic
ORFD [ 7 ]	Camera, LiDAR	✗	Front-view	1	1280 $\times$ 720	40	2D traversability
TartanDrive 2.0 [ 16 ]	Camera, LiDAR, INS	✗	Front-view	1	1024 $\times$ 512	2 $\times$ 32, 70	-

TABLE I : Comparison of several off-road datasets

Fig. 2 : Sensor setup and coverage of the Bunker Pro UGV platform. (a) shows the sensor placement including LiDAR, cameras, 4D radars, IMU and GPS. (b) illustrates the sensing range and coverage in off-road environments.

Fig. 3 : Various conditions in which the dataset was collected. (a) illustrates off-road terrains such as sandy, rocky, gravelly, puddled, and muddy. (b) shows diverse illumination conditions in the dataset: sunny, cloudy, shadow, sunset, and night.

Fig. 4 : Dataset composition across different environments.

Fig. 5 : Overview of automated 3D traversability map generation process. (a) LiDAR point clouds are accumulated over multiple time steps in the global coordinate system and reconstructed into a 3D mesh using Poisson surface reconstruction. (b) For each vertex of the 3D mesh, geometric features such as elevation, slope, and roughness are extracted to construct a traversability map. (c) A reference distribution is computed from the geometric features of voxels along the robot’s driving trajectories using a multivariate Gaussian, and a 3D traversability ground-truth label is automatically generated for each voxel whose Mahalanobis distance falls below a predefined threshold.

Fig. 6 : Visualization of the STONE dataset. The top shows 360-degree images of the environment captured around the robot, the bottom left presents temporally aggregated LiDAR points, and the bottom right illustrates the automatically generated 3D traversability ground-truth map.

Fig. 7 : Examples of regions with the same semantic class (vegetation) but different traversability labels. (a) presents a front-view image, and (b) presents the corresponding ground-truth. Green and magenta regions denote traversable and non-traversable regions, respectively. Although the cyan and yellow boxes both correspond to vegetation, the cyan box is correctly marked as non-traversable, and the yellow box remains labeled as traversable.

Model	Sensor	IoU occ	IoU T	IoU P	IoU N	mIoU
C-OpenOcc [ 26 ]	C	35.0	14.5	23.8	33.7	24.0
C-CoNet [ 26 ]	C	37.0	14.8	24.3	34.2	24.4
OccFormer [ 27 ]	C	37.8	16.3	29.0	34.1	26.5
TPVFormer [ 28 ]	C	39.6	16.9	29.6	35.5	27.3
L-OpenOcc [ 26 ]	L	61.2	21.1	33.4	55.2	36.5
L-CoNet [ 26 ]	L	61.6	22.5	34.8	57.0	38.1

TABLE II : Evaluation of 3D traversability baselines based on IoU metrics (%)

Explore similar work

Learning Traversability for Long Horizon Off-Road Navigation
Jul 26, 2026Kasi Viswanath, Jason M. Gregory, Shaunak Kolhe +1Off-Road Robot Navigation Geospatial Representation Learning
Department of Mechanical Engineering, Texas A&M University, College Station, TX 77840, USA · DEVCOM Army Research laboratory, Adelphi, MD, USA
Distilling Global Traversability Priors for Image-based Affordance Prediction in Off-road Environments
Jul 20, 2026Matthew Sivaprakasam, Samuel Triest, Micah Nye +5Off-Road Robot Navigation Embodied Navigation
Robotics Institute, Carnegie Mellon University · Field AI
Learning to Drive on Mars: Visual Multimodal Traversability Estimation for Off-World Navigation
Sep 21, 2026Darren Chiu, Cole Wilson, Andrei Tumbar +2Off-Road Robot Navigation Space Robotics
University of Southern California, Los Angeles, CA 90089, USA. · Jet Propulsion Laboratory, California Institute of Technology, Pasadena, CA 91109, USA. · Washington State University, Pullman, WA 99164, USA.

Jul 26, 2026J/K move · Enter open · S save

Learning Traversability for Long Horizon Off-Road Navigation

Kasi Viswanath, Jason M. Gregory, Shaunak Kolhe, Srikanth Saripalli

Department of Mechanical Engineering, Texas A&M University, College Station, TX 77840, USA · DEVCOM Army Research laboratory, Adelphi, MD, USA

Off-Road Robot Navigation Geospatial Representation Learning Robot Navigation Safe Robot Navigation

Open paper AskPDF arXiv

Autonomous navigation across large off-road environments remains a challenging problem. Onboard sensors perceive only the immediate surroundings, yet safe and efficient routes depend on terrain features that extend well beyond the sensor horizon. Geo-spatial data sources such as satellite imagery, aerial LiDAR, and vector maps can close this gap, but learning traversability from them is difficult: dense labels are unavailable at scale, and existing methods rely on short-range sensing. We propose an efficient formulation that learns a continuous traversability map from overhead data, supervised directly by human-driven GPS trajectories and shaped by supervised geometric priors from LiDAR. Alongside the model, we release a dataset, curated from public sources, consisting of 299 scenes spanning

\sim\!1{,}244\,\mathrm{km}^{2}

of diverse terrain, paired with

1{,}130\,\mathrm{km}

of human driving. In field trials on a Clearpath Warthog across seven routes at two sites,our method achieves trajectories within

5.5\%

of human path length and reduces operator interventions by

\sim\!85\%

compared to local-planner-only autonomy.

Abstract

Explore similar work

Learning Traversability for Long Horizon Off-Road Navigation

Distilling Global Traversability Priors for Image-based Affordance Prediction in Off-road Environments

Learning to Drive on Mars: Visual Multimodal Traversability Estimation for Off-World Navigation