Learning-Based Ground Vehicle Navigation in Outdoor Unstructured Environments

Autonomous navigation in outdoor unstructured environments is still an open challenge in field robotics, due in part to the difficulty to recognize and evaluate distances from obstacles and to identify type and slope of terrain. We present our current research on autonomous ground robot navigation in outdoor environments. Lying at the intersection of robotics and artificial intelligence, we investigate vision-based methods, integrating unsupervised learning and domain adaptation techniques, for improved sim-to-real capabilities. We validate the proposed methods with on-field experiments on real unmanned ground vehicles, thus assessing the feasibility of the developed navigation methods.