Classification of components of affective touch using rapidly-manufacturable soft sensor skins

Abstract

Affective touches play an important role in everyday communication because a significant amount of human interactions are through physical contacts. To facilitate robust, safe human-robot interactions (HRIs), we require robots with soft skins that can interpret affective touches. In this paper, we describe the fabrication of rapidly manufacturable, soft sensor skins using liquid metal embedded silicone elastomer as a resistive element and trained a recurrent neural network (RNN) to distinguish between a variety of pokes and rubs from human users. On a $2 \times 2$ sensor array, we obtained an average classification accuracy of 97% for ten different types of pokes and rubs, demonstrating that the combination of a soft sensor and machine learning can classify the interactions. Our approach is a step towards intelligent soft robots that can understand social interactions through touches.