Automated Learning of Operation Parameters for Robotic Cleaning by Mechanical Scrubbing

Abstract

The task of cleaning surfaces where foreign particles are removed by mechanical scrubbing requires oscillatory motions of the cleaning tool. Selecting the optimal operation parameters is important to automate this task with robots. The operation parameters can be the tool speed, force applied to the surface, frequency and amplitude of tool oscillation, stiffness offered by the robot, etc. The optimal set of parameters will be different for different surface/stain profiles and physical limitations of the robot. A large number of cleaning experiments need to be done if we try to find the optimal parameters exhaustively in a high dimensional space. It will also take a significant number of experiments to find the right model for the cleaning function and predict the optimal cleaning parameters under supervised learning settings. Conducting large number of experiments is often not feasible. We describe a semi-supervised learning approach to reduce the number of cleaning experiments to automate the process of finding the optimal cleaning parameters for arbitrary surface/stain profiles. This generalized method is also applicable for the tasks of grinding and polishing. Results from experiments with two Kuka robots performing cleaning tasks show the validity of our approach.