Novelty detection system for a social robot

This Thesis presents a system that allows a social human-interactive robot to be able to actively learn novel stimuli presented to it, expanding its knowledge base. The system detects unknown patterns and decides when those patterns are worth being learned by the robot. The system architecture leans on novelty detection algorithms, that serve to implement two steps: rst ltering noise entries; and then evaluating if the entries that have passed the noise lter are known by the existing model, or on the contrary, they are novel entries that are worth learning. When a novel entry is identi ed, the system activates the learning process to update the model with this new data. The novelty detection system is evaluated in the pose learning domain. The dataset is composed by 28 users that teach 3 di erent poses to the system. In the experiments, we compare the performance of four di erent novelty detection algorithms for this task. We rst evaluate the noise lter by analyzing how many entries from the same pose have to be shown to the robot to pass the lter. The second step is tested training the system with one of the poses, then evaluating if the algorithms are able to detect test entries from other poses as novel. A third experiment tests our system for detecting in-class novelties. The results show that the performances vary between the novelty detection algorithms. The best performance is achieved by GMM, with a 86% F score for detecting new poses and a 80% F score to detect variations within poses. This novelty detection system opens the door for robotic systems to be able to act as active learners, making their own decisions about when it is worth to learn from new stimuli. Additionally, to the extent of our knowledge, there is no reference on Novelty Detection for pose recognition in a Human-Robot interactive application, so this work is a novelty itself.Ingeniería Electrónica Industrial y Automátic