An observation model and segmentation algorithm for skill acquisition of a deburring task

Abstract

In robotic deburring applications it is desirable to have sensor feedback. The control strategies for this sensor feedback have to be adapted frequently to work piece and work tool parameters. The paper discusses a method for transferring the skill of a human operator to a control strategy that can cope with these changes in parameters. The human skill is transferred by an indirect learning method. The human actions are modeled as an impedance controller whose parameters are adapted by observations of the deburring process state. The nonlinear relation between the process state and the controller parameters is learnt by a neural network. To apply this method it is necessary that the observations are independent of the controller actions. This is shown for an observation model that is derived from a process model for deburring and experimentally verified. Segmentation of the training data is done by analyzing the summed normalized innovation squared value of a static Kalman filter.