Online Prediction of Robot to Human Handover Events Using Vibrations

Abstract

One of the main issues for a robotic passer is to detect the onset of a handover, in order to avoid the object from being released when the human partner is not ready or if some impact occurs. This paper presents the methodology for a robotic passer, that is potentially able to estimate the interaction forces by the receiver on the object, thus to achieve fluent and safe handovers. The proposed system uses a vibrator that energizes the object and an accelerometer that monitors vibration propagation through the object during the handover. We focused on the machine-learning technique to classify between four states during object handover. A neural network was trained for these four states and tested online. In experimental trials an accuracy of 85.2 % and 93.9% were obtained respectively for four classes and two classes of actions by a neural network classifier.