A case study for learning behaviors in mobile robotics by evolutionary fuzzy systems

Service robots will play an increasing and more important role in the society in the next years. One of the main challenges is to endow robots with enough autonomy to operate on real environments. To reach that goal, the design of controllers to solve simple tasks must be automatized. Engineers look for learning algorithms that are general, robust, require low expertise knowledge, and generate controllers that can run on the real robot without any tuning stage. In this paper, a framework to learn behaviors (controllers) in mobile robotics, fulfilling the previous requirements, has been used. The framework is based on two modules: dataset generation and a data-driven evolutionary-based learning algorithm to obtain fuzzy controllers. Nevertheless, the design of a fuzzy controller still requires the selection of the type of learning algorithm, and also to choose the value of some design parameters. In this paper we present an exhaustive study on a set of evolutionary-based data-driven learning algorithms, for learning fuzzy controllers in mobile robotics, that cover a wide range of the accuracy/interpretability trade-off. The study has also evaluated the influence of the values of all the design parameters over accuracy and interpretability. The objective is to analyze the performance of the different algorithms for the design of behaviors in mobile robotics, and to extract some general rules that can help in the process to design new behaviors. The analysis comprises two different behaviors (wall-following and moving object following) and more than 450 tests, both in simulation and on a Pioneer II AT robot. Results have shown very good performances in complex and realistic conditions for the different combinations of algorithms and parameters

keywords: Service robots, Data-driven learning algorithms, Evolutionary algorithms, Fuzzy controllers, Design of behaviors