Data-Based Learning for Control and Prediction of Robot-Structure Interactions with Comparable Robot and Structure Stiffnesses

Abstract

Efficient and accurate robotic interactions with flexible structures are critical for many manufacturing processes where the elasticity of the workpiece and the robot must be accounted for. In particular, for clamping and drilling flexible structures, maintaining tool-workpiece normality and limiting shear forces are essential. The main contributions of this thesis are to show: Firstly, that data acquired during a robotic clamping operation can be learned and used to speed up the process for similar operations. Utilizing the learned parameters, a map between the measured forces and robot joint positions is used to develop time-based robot-joint (velocity) trajectories to achieve a specified robot-workpiece interaction. Experimental results show that the operating speed can be increased while maintaining interaction forces and torques within acceptable levels. Secondly, with a minimum number of learned measurement locations controlled interactions such as clamping operations can be efficiently predicted at unmeasured locations. Gaussian Process Regression methods are presented as a suitable means to predict interactions, ultimately resulting in increased interaction speeds at workpiece locations where controlled-interactions have not been conducted. Additionally, to minimize the number of measurement locations, a method is developed using existing interaction data to optimize measurement location selection by minimizing prediction uncertainty.