Learning fish-like swimming with a CPG-based locomotion controller

This paper presents a learning method to acquire fish-liking swimming with a CPG-based locomotor controller. The proposed method converts the related CPG parameters into dynamical systems that evolve as part of the CPG network dynamics. The teaching signals are derived from the kinematic model of carangiform swimming with trajectory approximation method. A novel coupling scheme for the CPG network, which are modeled as a chain of coupled Hopf oscillators is proposed to eliminate the influence of afferent signals on amplitude of the oscillator. The learning rules of intrinsic frequency, coupling weight and amplitude are formulated with phase space representation of the oscillators. The frequency, amplitudes and phase relations of the teaching signals can be encoded by the CPG network with the adaptation mechanisms. Numerical experiments are carried out to validate the effectiveness of the proposed learning rules.