Offroad-global-nav Geospatial Dataset

Overview

This repository contains the dataset introduced in:

“Learning Traversability-Aware Global Planners for Long Horizon Off-Road Navigation”

The dataset is designed to support long-range off-road navigation using multi-modal geospatial data, combining large-scale overhead sensing with real-world human driving behavior.

Unlike traditional off-road datasets that focus on local perception, this dataset enables global planning over kilometer-scale environments.

---

📊 Dataset Summary

• Scenes: 299 geographically diverse locations (more will be added)
• Coverage: ~1,244 km²
• Human driving data: ~1,130 km of GPS trajectories

Modalities

• 🛰️ Satellite imagery (GeoTIFF)
• 🌐 Aerial LiDAR point clouds (LAZ)
• 🗺️ OpenStreetMap vectors (OSM XML)
• 📍 Human trajectories (KML)

Each scene is fully geo-referenced and co-registered.

---

🗺️ Geographic Coverage

The dataset spans diverse terrain types across the United States, including:

• Deserts
• Grasslands
• Forests
• Mountains
• Quarries and mines

Red markers indicate sampled regions used for dataset construction.

*Satellite Images from different locations with diverse terrain.

🧱 Data Structure

🔍 Modalities Explained

Satellite Imagery

High-resolution RGB imagery capturing:

• Vegetation
• Trails
• Water bodies
• Terrain appearance

---

LiDAR Point Clouds

Dense aerial LiDAR (5–27 pts/m²) providing:

• Elevation (height)
• Surface normals (slope)
• Intensity (surface reflectivity)

---

OpenStreetMap (OSM)

Vector priors including:

• Roads and trails
• Waterways

Used as weak semantic supervision.

---

Human GPS Trajectories

Real-world driving paths used as:

• Implicit supervision for traversability
• Ground truth for path preference

---

🚀 Getting Started

from datasets import load_dataset

dataset = load_dataset("anony-008/offroad-global-nav")

sample = dataset["data"][0]

---

🤝 Acknowledgements

This dataset builds upon publicly available geospatial data sources including:

USGS LiDAR

ArcGIS satellite imagery

OpenStreetMap