Touch Interaction Classification Using Upper-Body Sensor Suit

In this study, we evaluated different machine learning algorithms for touch classification using data obtained with the upper part of a whole-body sensor suit for human-looking robots and gathered human touch interaction data, comparing the results with human classification based on video footage. The data collection experiment included eight different touch behaviors: hit, hug, pat, poke, push, shake, tap, and sexual harassment (SH). Two different datasets were defined based on the data gathered: instant dataset which is the exact moment when the touch is performed and average dataset which is the average of ~1 s before and after of the touch interaction. Seven different machine-learning algorithms were implemented: logistic regression (LR), linear support vector machine (LSVM), random forest classifier (RFC), one-vs-rest (OVR) classifier, kernelized support vector machine (KSVM), K-nearest neighbors (KNNs), and fully connected neural network (FCNN). The results obtained showed that the FCNN has the best performance in both datasets, having the higher score with the average dataset. Finally, this result was compared with the human classification results, which had a considerably lower accuracy, showing that the implementation of a neural network with the sensor suit can lead to efficient touch behavior classification in human-robot interaction.