Autonomous Learning for Long Range Vision in Mobile Robots

Designing a vision-based autonomous robot that can navigate through complicated outdoor environments is an extremely challenging problem that is far from solved. However, through use of a self-supervised realtime learning strategy and a trained deep belief net for compact feature representations, we have developed a long-range vision system that succeeds in accurately classifying obstacles and traversable areas that are 200 meters or more distant, bringing us closer to the goal of human-level autonomous driving.