A comparison of neural networks and physics models as motion simulators for simple robotic evolution

Abstract

Robotic simulators are used extensively in Evolutionary Robotics (ER). Such simulators are typically constructed by considering the governing physics of the robotic system under investigation. Even though such physics-based simulators have seen wide usage in ER, there are some potential challenges involved in their construction and usage. An alternative approach to developing robotic simulators for use in ER, is to sample data directly from the robotic system and construct simulators based solely on this data. The authors have previously shown the viability of this approach by training Artificial Neural Networks (ANNs) to act as simulators in the ER process. It is, however, not known how this approach to simulator construction will compare to physics-based approaches, since a comparative study between ANN-based and physics-based robotic simulators in ER has not yet been conducted. This paper describes such a comparative study. Robotic simulators for the motion of a differentially-steered mobile robot were constructed using both ANN-based and physics-based approaches. These two approaches were then compared by employing each of the developed simulators in the ER process to evolve simple navigation controllers for the experimental robot in simulation. Results obtained indicated that, for the robotic system investigated in this study, ANN-based robotic simulators offer a promising alternative to physics-based simulators.