From simulated to real robots

Abstract

Evolutionary robotics using genetic algorithms to evolve control systems for real robots is a powerful tool, since it allows an automatic evolution of control systems. However, evolutionary robotics has serious limitations because of the time involved. It is very time consuming to evolve whole populations of real robots for many generations. A simulated/physical approach where main parts of the evolution takes place in a simulator reduces the time consumption dramatically. We describe how the Khepera miniature mobile robot can be used to build its own simulator with a semi-autonomous process, how to evolve neural network control systems for the Khepera robot in the robot's own simulator, and how to transfer the neural network control systems from the simulated to the real environment. By using this kind of simulator an expected gap in performance when transferring a robot control system from the simulator to the real environment is avoided.