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

TL;DR

The paper introduces a method to learn continuous traversability maps from overhead data like satellite imagery and LiDAR, using human-driven GPS trajectories and geometric priors for supervision. It enables long-horizon off-road navigation by predicting safe, efficient paths beyond sensor range, validated across multiple real-world routes with significant reductions in path deviation and operator intervention.

Why It Matters

Long-range off-road navigation is critical for applications like agriculture, disaster response, and exploration, where vehicles must operate autonomously over vast, unstructured terrain. This approach reduces reliance on real-time sensing and human oversight, enabling more robust and scalable autonomy in challenging environments.

Key Findings

• The model learns traversability maps from overhead data using GPS trajectories and self-supervised LiDAR priors, eliminating the need for dense manual labels.
• Field tests on a Clearpath Warthog showed trajectories within 3.66% of human path length and an 85% reduction in required operator interventions.
• A public dataset of 299 scenes covering over 1,200 km² and 1,130 km of driving data was released to support future research.