Learning to Prevent Failure States

To achieve robust autonomy, robots must avoid getting stuck in states from which they cannot recover without external aid. While this is the role of the robot's control algorithms, these are often imperfect. We examine how to detect failures by observing the robot's internal sensors over time. For such cases, triggering a response when detecting the onset of a failure can increase the operational range of the robot. Concretely, we explore the use of supervised learning techniques to create a classifier that can detect a potential failure and trigger a response for a dynamically balancing robot. We present a fully implemented system, where the results clearly demonstrate an improved safety margin for the robot