Using novelty detection in HRI: Enabling robots to detect new poses and actively ask for their labels

Active robot learners take an active role in its own learning by asking queries to its human teachers when they receive new data. However, not every received input is useful for the robot, and asking for non-informative inputs or asking too many questions might produce a negative impact on how the human perceives the robot. We present a novelty detection system that enables a robot to ask questions only when it decides that a stimuli is both novel and interesting. Our system is based in separating the decision process in two steps: first discriminating novel from known stimuli to the robot's model and second by discriminating if this stimuli is likely to happen again. Our approach uses the notion of curiosity, which controls the eagerness in which the robot asks questions to the user. We evaluate our approach in the domain of pose learning by training our robot with 20 instances of poses and then showing it some novel ones. Our approach is able to detect novel poses with a 84% F-Score. Also, tuning our curiosity parameter we have been able to control the minimum number of times in which the robot needs to be exposed to a new stimuli before it asks the user a label for this stimuli. Our approach enables robots to keep learning continuously, even after its training is finished. Also, the introduction of the curiosity parameter, allows to tune which are the conditions in which the robot should want to learn more.