General controllers evolved through grammatical evolution with a divergent search.

In this work, we analyse the performance of Novelty Search (NS) in a set of generalization experiments in a navigation task with Grammatical Evolution. Agents are trained on a single, simple environment, and tested on a selection of related, increasingly more difficult environments. We show that agents discovered with NS, although using a tiny number (six) of training samples, successfully generalise to these more difficult environments.