Learning from limited experience: Real-world robot navigation from a single traversal

If you want to do something, first you have to go somewhere. Navigation is a crucial capability for any intelligent agent that is expected to operate in the world, be they human, animal, or robot. In this work, we present techniques for teaching robots to navigate in the real world, given only a single prior traversal of the environment. We employ an online adaptive approach to sequence-based localization, and a novel framework for model-based deep reinforcement learning, to develop robot systems that can successfully localize and navigate in a city