Comparative Analysis of Artificial Neural Network Training Methods for Inverse Kinematics Learning

摘要

A method for obtaining an approximate solution to the inverse kinematics of a articulated chain is proposed in this paper. Specifically, the method is applied to determine the joint positions of a humanoid robot in locomotion tasks, defining the successive stable robot configurations needed to achieve the final foot position in each step. Our approach is based on the postural scheme method, using artificial neural networks to solve the problem. In this paper we define the restrictions that must be accomplished by the networks and make an exhaustive study about the learning algorithms, transfer functions, training sets composition, data normalization and artificial neural network topologies.