Autonomous off-road vehicles struggle to plan efficient routes because their onboard sensors only see the immediate surroundings. To close that gap, researchers at Texas A&M and NASA's Jet Propulsion Laboratory propose a method that learns a continuous traversability map from overhead sources such as satellite imagery, aerial LiDAR, and vector maps. The model is supervised directly by human-driven GPS trajectories and shaped by geometric priors from LiDAR, avoiding the need for dense terrain labels at scale.
The team also released a dataset curated from public sources, containing 299 scenes spanning roughly 1,244 square kilometers of varied terrain, paired with 1,130 kilometers of human driving. In field trials on a Clearpath Warthog across seven routes at two sites, the method achieved trajectories within 5.5% of human path length and reduced operator interventions by about 85% compared with local-planner-only autonomy.
These results suggest that overhead data can effectively extend the planning horizon for off-road robots, making them safer and more autonomous in large, unstructured environments. The paper is available on arXiv under reference 2607.23743.