Back-Propagation Neural Network based predictive control for biomimetic robotic fish

Abstract

This paper presents a practical swimming data prediction method for a free-swimming, three-link robotic fish. Since a full hydrodynamic model for fish swimming is very complex and intractable, the primitive swimming data generated by a Central Pattern Generator controller is fed into a Back-Propagation Neural Network (BPNN) for trimming. After the process of training, the BPNN is able to predict the actual swimming data for various swimming patterns without dynamic modeling. Preliminary simulation and experimental results on swimming control show the effectiveness of the proposed prediction method as well as its potential for other flexible link-based robots.