An example-based approach for evolving robot controllers

One method to cope with the difficulty of dealing with the complex robot-environment interaction in building behavior-based robots is to automate the design of robot behaviors (control modules). To achieve this, the evolutionary approach has been proposed and has shown its efficiency in synthesizing controllers. In this paper, we present a new approach, called an example-based approach, for evolving robot behaviors. Instead of using lower level quantities, such as sensor or motor activities, to quantitatively describe the desired robot behaviors in defining fitness functions as other evolutionary robotics systems do, we give behavior examples to the evolutionary system to train populations of controllers. Experimental results show that our approach not only can construct robot behaviors automatically but also overcome the difficulty of defining fitness functions by lower level terms as in traditional work for evolving robot controllers