Comparison between fuzzy and neural controllers to cross the reality gap in evolutionary robotics

摘要

This paper compares fuzzy and neural controllers when trying to cross the reality gap in evolutionary robotics. Reality gap is one of the most relevant open questions in evolutionary robotics for it restricts its use in practical and complex applications of robotics. Controllers are compared by navigation metrics for differential drive robots (a Pioneer 3-DX). Based on the metrics, similarity between the controllers could be verified. This similarity allows the conclusion that Fuzzy Logic can also be used in evolutionary robotics. Furthermore, when compared with a Neural Network controller, the fuzzy control strategy results in smoother trajectories. Another relevant result is that the use of a Fuzzy Logic approach makes the evolutionary process faster, only requiring 30 generations, while the Neural Network approach requires approximately 70 generations. Finally, the fuzzy controller allows the user to include known characteristics of the system by a human specialist. This cannot be achieved with the use of Neural Networks. However, the neural network can be indicated in complex systems or tasks, where the designer has very little or no information about the system.