Vision-based reinforcement learning for robot navigation

We present a novel vision-based learning approach for autonomous robot navigation. A hybrid state-mapping model, which combines the merits of both static and dynamic state assigning strategies, is proposed to solve the problem of state organization in navigation-learning tasks. Specifically, the continuous feature space, which could be very large in general, is first mapped to a small-sized conceptual state space for learning in static. Then, ambiguities among the aliasing states, i.e., the same conceptual state is accidentally mapped to several physical states that require different action policies in reality, are efficiently eliminated in learning with a recursive state-splitting process. The proposed method has been applied to simulate the navigation learning by a simulated robot with very encouraging results