Deep Learning of Structure-Borne Sound for Contact Classification in Robotics

Abstract

A complete perception is necessary for robots in unstructured surroundings. This is especially important for manipulation tasks in which the robot is in contact with the environment. Since sensing modalities have advantages and disadvantages in different situations, sensory information from multiple sources may help achieve a more complete and accurate perception. The possibility of using structure-born sound for the purpose of gaining information about a contact is introduced in this work. A method for material classification is developed and demonstrated. Additionally, the possibility to estimate the position of the contact is analyzed. In order to avoid the difficulty of modeling the behavior of structure-borne sound, a machine learning algorithm is applied. The algorithm used is the DLR-VAE, which is an improved version of the variational autoencoder. Various parameters, which influence the accuracy of the classification are identified and analyzed. Furthermore, the resulting model is tested on the real robot system.