A Neural Network System for Calculation of Inverse Dynamics for Manipulators

This paper proposes a method to calculate manipulator inverse dynamics using a neural network. In this method, two sets of neural networks are prepared. One is for the elements of inertia moment matrix, and the other is for gravitational force. Each input for the network is only a joint position. Teacher signals of each network are also calculated using only a joint position, and therefore learning of each network is fast. The neural network, which acquires a model of inertia moment matrix, is used to calculate inertial force, centrifugal force and Coriolis force. In particular, the terms of centrifugal force and Coriolis force are calculated using a characteristic of manipulator dynamics and structure of the neural network. This method can be applied to the wide area data of joint positions, joint velocities and joint accelerations to calculate manipulator inverse dynamics. To show the validity of this method, the inverse dynamics of a two-dimensional manipulator are calculated.