John Huddleston, Jianna Zhang
Service animals undergo rigorous and lengthy training to fill many difficult and dangerous roles for the benefit of their human counterparts. With existing technology, a subset of these roles that depend on the ability to learn and intelligently respond to a variety of external stimuli can also be filled by robots with less training time and maintenance cost. This paper explores such an approach with a guide robot for the blind that is trained by feedback from both humans and the environment using a reinforcement learning model. This model will allow the guide robot to selectively obey human commands depending on its understanding of the safety of doing so. We expect the robot will quickly learn when to disobey after exposure to a diverse set of environments.