Detecting Conversational Failures in Task-Oriented Human-Robot Interactions

Sammanfattning: In conversations between humans, not only the content of the utterances but also our social signals provide information on the state of communication. Similarly, in the field of human-robot interaction, we want the robots to have the capability of interpreting social signals given by human users. Such social signals can be operationalised in order to detect unexpected behaviours of robots. This thesis work aims to compare machine learning based methods to investigate robots’ recognition of their own unexpected behaviours based on human social signals. We trained support vector machine, random forest and logistic regression classifiers with a guided task human-robot interaction corpus that includes planned robot failures. We created features based on gaze, motion and facial expressions. We defined data points of different window lengths and compared effects of different robot embodiments. The results show that there is a promising potential in this field and also that the accuracy of this classification task depends on different variables that require careful tuning.