Automated robot function recovery after unanticipated failure or environmental change using a minimum of hardware trials

Recovering functionality after unanticipated damage or environmental change, using a minimum amount of hardware testing, is a desirable and under-explored topic in evolutionary hardware and evolutionary robotics. In a previous paper, we introduced a two-stage evolutionary algorithm, which we call the estimation-exploration algorithm, that evolves a robot simulator to accurately describe what damage a ´physical´ robot has undergone, and then evolves a compensatory neural network in the evolved simulator that, when downloaded to the ´physical´ robot, restores functionality. Here we introduce a new fitness metric that allows the algorithm to correctly describe not only complete but also partial failures, and also allows the algorithm to disambiguate between internal damage and external environmental change, based solely on sensory feedback. In most cases only four hardware evaluations are necessary in order to restore complete functionality to the ´physical´ robot.