Automated Optimisation of Collision-aware Contact Classification in Robotics

Abstract

Contact classification is of great importance when it comes to robotics and aeronautics. Robotic manipulators should be able to quickly register contact and act accordingly in order to prevent or limit damages through collisions. Hence, the robot needs to be able to discriminate unintended collisions from desired contacts. For this determination, a collision-aware contact classifier is needed. In this thesis, a Fourierbased approach is introduced, which implements a convolutional neural network. Moreover, a framework for collision-aware contact classification is optimised in its labeling and training process. Usually, these processes are time-consuming, hence, automation techniques will be proposed and integrated into the learning pipeline of the humanoid robot Rollin' Justin from the German Aerospace Center (DLR). The aim is to improve the ease of use of such classifiers in future applications and increase the autonomy of humanoid robots.