Fitness Based Identification of a Robot Structure

Embodiment theory suggests that recurrent processes of sensorimotor activity give rise to cognitive structures. In the case of robots, internal sensorimotor activity generated with physics simulators can be exploited to expand the historical domain of action, however pre-engineered simulations are limited by the reality gap problem. Alternatively simulation might be inferred and self-constructed out of data collected during robot functioning. Fundamental to this line of research is defining a distance function to asses the potential of candidate robot simulations to reproduce real world activity. In this paper we study the characteristics of a distance function based on behavioral fitness measurements. We show how this function can be applied for the generation of behaviors using an algorithm that co-evolves a robot and its simulation. The experiments show how the monotonicity of the function increases with the number of behaviors being tested in reality and with the genotypic diversity of corresponding robot controllers. Moreover it allows for the accurate identification of behavior-relevant parameters contained in the simulation. The metric shows an advantage, when compared to other metrics, for assessing the quality of simulators over long time scales of robot behavioral evaluation.