Central pattern generator parameter search for a biped robot walking using nonparametric estimation based particle swam optimization

Abstract

Although bipedal walking based on Central Pattern Generator (CPG) is promising, parameter search of CPG is hard because there is no methodology to set the parameters and the search space is too big. Therefore, evolutionary computation(EC) methods such as Genetic Algorithms(GAs), multi-objective Genetic Algorithms, and Genetic Programming(GP) are often used to optimize the parameters. However, when EC is used to find parameters of CPG, the fitness of the parameters is evaluated by applying to a robot and the evaluation takes long time. So fast convergence is a important factor for selecting a method to prevent robot from too much iterations of the method. In this thesis, nonparametric estimation based Particle Swarm Optimization (NEPSO) is suggested to search parameters of CPG for bipedal walking. Canonical Particle Swarm Optimization (PSO) generally converges faster than the other EC method such as GAs and GP and the suggested algorithm converges faster than the PSO. The NEPSO uses a concept experience repository to store previous position and fitness of particles in PSO and estimated best position to accelerate convergence speed. The proposed algorithm is compared with PSO variants in numerical experiments and a tested in a three dimensional dynamic simulator for bipedal walking.