Using Neural Networks for Classifying Human-Robot Contact Situations

Abstract

State-of-the-art robotic manipulators come with functions that detect unforeseen contact situations with the environment. However, for collaborative manipulators that are specifically designed to allow physical Human-Robot Interaction (pHRI), contact situations can be either intended (the operator may intentionally interact with the robot for instance to modify its trajectory) or undesired (a human accidentally runs into the robot). An appropriate reaction strategy for the robot manipulator depends on a correct classification and localization of the contact situation. To this end, this work proposes an approach using supervised learning techniques to distinguish between unintended contact situations (labeled collisions) and foreseen ones (labeled interactions), and to infer whether a contact occured on the upper or lower arm of the robot. A neural network is trained on measurement data gathered from different contact situations between a human subject and an ABB YuMi robot. The proposed method is then evaluated online on the robot using simple reaction strategies on both the person who trained the network and on other human individuals.