Solving deceptive tasks in robot body-brain co-evolution by searching for behavioral novelty

Evolutionary algorithms are a frequently used technique for designing morphology and controller of a robot. However, a significant challenge for evolutionary algorithms is premature convergence to local optima. Recently proposed Novelty Search algorithm introduces a radical idea that premature convergence to local optima can be avoided by ignoring the original objective and searching for any novel behaviors instead. In this paper, we apply novelty search to the problem of body-brain co-evolution. We demonstrate that novelty search significantly outperforms fitness-based search in a deceiving barrier avoidance task but does not provide an advantage in the swimming task where a large unconstrained behavior space inhibits its efficiency. Thus, we show that the advantages of novelty search previously demonstrated in other domains can also be utilized in the more complex domain of body-brain co-evolution, provided that the task is deceiving and behavior space is constrained.