Neural Network based CPG Control Method of Undulating Fin in Underwater Biomimetic Robot

Abstract

This paper presents a deep neural network (DNN) based mapping and inverse mapping method to control an undulating fin developed in soft material. The undulating fin can reproduce the periodic biological motion by applying a central pattern generator (CPG) oscillating model and constructing networks between oscillators. Due to the dynamic complexity of the soft material and the dependency on the CPG oscillator, the dynamic model of the system has complex and highly non-linear properties. Here, to model the complex non-linear relationship between CPG control input and output thrust, we propose two DNN models for mapping (recognition network) and inverse mapping (generation network). Water tank experiments were conducted to obtain the output thrust according to the various CPG inputs, and the experimental data were used for learning the proposed networks. After learning, we compared the accuracy of the output thrust estimation between the proposed mapping method using the recognition network and the existing linear regression method. The tendency of CPG input recommendation using the generation network was also confirmed. Based on the proposed approach, we designed a robot control methodology integrating the proposed DNN and CPG network.