Model-free execution monitoring by learning from simulation

Abstract

Autonomous robots need the ability to plan their actions and to execute them robustly and in a safe way in face of a changing and partially unpredictable environment. This is especially important if we want to design autonomous robots that can safely co-habitate with humans. In order to manage this, these robots need the ability to detect when the execution does not proceed as planned, and to correctly identify the causes of the failure. An execution monitoring system is a system that allows the robot to detect and classify these failures. In this work we show that pattern recognition techniques can be applied to realize execution monitoring by classifying observed behavioral patterns into normal or faulty behaviors. The approach has been successfully tested on a real robot navigating in an office environment. Interesting, these tests show that we can train an execution monitor in simulation, and then use it in a real robot.