System identification of a robot arm with extended Kalman filter and artificial neural networks

Abstract

Abstract The development of an algorithm to describe a dynamic system and to predict its future behaviour in further consequence is the aim of the present study. Non parametric models provide a general description of object dynamics and artificial neural networks (ANN), which are a very flexible and universal learning method, belong to it. However, the standard estimation procedures for ANN like Levenberg-Marquardt (LM) do not consider that data is observed and consequently is uncertain. The combination with the extended Kalman filter (EKF) enables the consideration of the uncertainty in the estimation process. The analogies between EKF and LM are discussed and thereon the advantages of the EKF are outlined. The integration of ANN into EKF will be evaluated on an industrial robot arm. At first, a simplified model is determined; the ANN describes the robot position deviations as a function of the joint encoder values. The robot reference positions are measured by a laser tracker. In order to compare it with the robot outputs, the observations need to be transformed to the robot frame and the offset between the end-effector and the robot flange has to be determined. A method to estimate both parameters simultaneously is developed and the results are verified on basis of simulated data. This paper comprises two novel approaches. First, uncertainty is considered in the ANN estimation on basis of the combination with the EKF. Considering the full covariance matrix of the robot deviations leads to a better prediction of the robot’s behaviour. Second, an integrated transformation and lever arm determination is introduced and the robot’s repeatability presents the limiting factor of the achievable parameter uncertainty.