Abstraction as a Mechanism to Cross the Reality Gap in Evolutionary Robotics

Abstract

One of the major challenges of Evolutionary Robotics is to transfer robot controllers evolved in simulation to robots in the real world. In this article, we investigate abstraction on the sensory inputs and motor actions as a potential solution to this problem. Abstraction means that the robot uses preprocessed sensory inputs and closed loop low-level controllers that execute higher level motor commands. We apply abstraction to the task of forming an asymmetric triangle with a homogeneous swarm of MAVs. The results show that the evolved behavior is effective both in simulation and reality, suggesting that abstraction can be a useful tool in making evolved behavior robust to the reality gap. Furthermore, we study the evolved solution, showing that it exploits the environment (in this case the identical behavior of the other robots) and creates behavioral attractors resulting in the creation of the required formation. Hence, the analysis suggests that by using abstraction, sensory-motor coordination is not necessarily lost but rather shifted to a higher level of abstraction.