State Space Construction by Attention Control

In order to understand cognitive aspects of autonomous robots, it is fruitful to develop a mechanism by which the robot autonomously analyzes physical sensor data and construct a state space. This paper proposes a coherent approach to constructing such a robot oriented state space by statistically analyzing sensor patterns and rewards given as the result of task executions. In the state space construction, the robot creates sensor pattern classifiers called Empirically Obtained Perceivers (EOPs) which, when combined, represent internal states of the robot. A novel feature of this method is that the EOP directs attention to select necessary information, and the state space is obtained with the attention control mechanism using EOPs. We have confirmed that the robot can effectively construct state spaces through its vision sensor and execute a navigation task with the obtained state spaces in a complicated simulated world.