Evolutionary Learning of Basic Functionalities for Snake-Like Robots

Abstract

The objective of the work presented in this paper is to investigate the optimal learning strategy for snake-like modular robots using a (1+1) Evolutionary Algorithm. We take into account three different but correlated tasks: efficient locomotion, reaching a given point and obstacle avoidance.

Starting from earlier results on locomotion, we have performed three different sets of experiments. In the first, the snake must learn to go to a goal point, and we investigate how employing different fitness functions affect the learning of this task. In the second experiment the snake must learn how to avoid obstacles. In this experiment we test two possible behaviors, called mixed and switched strategies. Finally, in the third set of experiments, we introduce the concept of incremental learning and compare it with the “all-at-once” learning schemes of the first two experiments.

The results of the simulations indicates that the modular robots are able to learn both tasks with the (1+1) Evolutionary Strategy adopted, and that the fitness function that explicitly rewards each of the tasks perform better that the fitness function that takes into account the locomotion task only implicitly, rewarding only the reaching of the target point. We also demonstrate that the obstacles avoidance configuration with only one behaviour (mixed) is better than the configuration with two behaviours (switched) and that incremental learning provide a faster evolution towards good controllers.