A measure of good motor actions for active visual perception

The morphology of an embodied agent heavily influences the characteristics of sensory signals induced during sensorimotor activity. In order to facilitate later information processing, it is desirable to couple sensors and actuators in a way such that the agent's actions induce well structured sensory feedback. In this work, we empirically investigate the meaning of “well structured” with respect to the predictability of the induced stimulus and the complexity of the model required to predict the change. For this purpose, we consider a visual sensor consisting of an unknown topology composed of light sensitive receptors and we investigate a number of different actions. For each action, we learn a stimulus prediction function and we propose to qualify the action depending on the complexity of the prediction model. By visualizing this measure for two different sensors and over different action spaces, we observe that a given sensor topology implicitly imposes actions for which the provoked change in stimulus can be described with less effort, which we feel has direct implications for the behavior of an agent using that sensor.