Abstract
In this work, preparations for the nonparametric calibration of a 7 DOF light-weight robot using machine learning techniques are presented. The approach was developed to satisfy the requirements on absolute accuracy for robot-assisted surgery. With the kinematic and non-kinematic properties in mind, we showed that a decomposition of the robot's kinematic chain can drastically reduce the number of necessary samples for a sophisticated training set. Thus, the data acquisition can be accomplished in a feasible time frame. Furthermore, we cope with the problem of data registration between the robot's internal model and the external measurements. We can show that by carefully choosing the split point for the decomposition, errors caused by the dependency between sub-chains of the robot are small enough to yield satisfying results.