Single-Class Discrimination for Robot Anomaly Detection

Abstract

Early system error detection is essential for safe human-robot interaction. Detecting errors in the large amount of sensory and status data is, however, difficult, especially since the amount of fault data is very small if at all available; a problem for which no sufficiently stable methods exist. This thesis introduces a new Robot Anomaly Detection System (RADS) consisting of a multi-stage solution, that adapts to high dimensional input and detects errors without previous records of them. Application to real robot data proves its satisfying performance.