Evolutionary Robotics, Anticipation and the Reality Gap

Abstract

Evolutionary Robotics provide efficient tools and approach to address automatic design of controllers for autonomous mobile robots. However, the computational cost of the optimization process makes it difficult to evolve controllers directly into the real world. This paper addresses the key problem of transferring into the real world a robotic controller that has been evolved in a robotic simulator. The approach presented here relies on the definition of an anticipation-enabled control architecture. The anticipation module is able to build a partial model of the simulated environment and, once in the real world, performs an error estimation of this model. This error can be reused so as to perform in-situ online adaptation of robot control. Experiments in simulation and real-world showed that an evolved robot is able to perform on-line recovery from several kind of locomotion perturbations.