Characterizing Anomalous Behaviors and Revising Robotic Controllers

This paper is concerned with revising autonomous robotic controllers. A proof of concept of the proposed machine learning-based approach is presented, aimed at characterizing and avoiding the wobbling phenomenon incurred by a Braitenberg controller. Based on the global assessment of a few trajectories by the expert, the goal is to identify erroneous sub-behaviors. The success criterion is to be able to identify as soon as possible (early alarm) such behaviors when they occur, in order e.g. to trigger an emergency controller.