Integrating developmental and conventional Markov decision processes: An application to robotic navigation

This paper proposes an architecture for developmental robot that can learn abstract concepts early on and use these concepts to reason and make decisions. We introduce a frame work of two macro-layers. The bottom layer takes the desired information (e.g., desired heading direction) as the input. The top macro-layer enables human teachers to interactively inject a representation of abstract concepts (e.g., location) into the developmental process. This architecture is applied to a navigation problem, and its superiority over one-layer architecture is confirmed in comparative experiments using simulated Lidar sensor data. The robotic navigation demonstrates its robustness in accomplishing complicated task in clutter outdoor environments.