Mixed Compositional Pattern-Producing Network-NeuroEvolution of Augmenting Topologies Method for the Locomotion Control of a Snake-Like Modular Robot

Abstract

Snake robots, which are a type of bionic robots, have high adaptability to various environments. Because of the unique structure and many degrees of freedom (DoFs) of snake robots, these robots can move on rough terrain or in narrow spaces. However, developing a controller for the locomotion of snake robots with many DoFs is challenging. A control model developed for a snake robot moving in one environment might be unsuitable in other environments. Thus, a mixed compositional pattern-producing network (CPPN)-NeuroEvolution of Augmenting Topologies (NEAT) method is proposed in this paper. This method is based on a neuroevolution algorithm and serpentine locomotion, and it can be used to control the locomotion of a snake-like modular robot in multiple environments with different obstacles. In the proposed method, a group of neural networks is constructed for each environment, and these networks are trained using a neuroevolution algorithm for several generations. Some pretrained neural networks for each environment are selected and then integrated to obtain mixed neural networks. These mixed neural networks are then evolved using a multiobjective genetic algorithm (MOGA) to improve the locomotion of a snake-like modular robot in multiple environments. The results of this study indicated that the mixed CPPN-NEAT method outperformed the combined CPPN-NEAT method and use of MOGA alone (without pretraining, combined networks, and mixed networks) in locomotion control for a snake-like modular robot in three environments with different obstacles.