A Quality-Diversity Approach to Evolving a Repertoire of Diverse Behaviour-Trees in Robot Swarms.

Designing controllers for a swarm of robots such that collaborative behaviour emerges at the swarm level is known to be challenging. Evolutionary approaches have proved promising, with attention turning more recently to evolving repertoires of diverse behaviours that can be used to compose heterogeneous swarms or mitigate against faults. Here we extend existing work by combining a Quality-Diversity algorithm (MAP-Elites) with a Genetic-Programming (GP) algorithm to evolve repertoires of behaviour-trees that define the robot controllers. We compare this approach with two variants of GP, one of which uses an implicit diversity method. Our results show that the QD approach results in larger and more diverse repertoires than the other methods with no loss in quality with respect to the best solutions found. Given that behaviour-trees have the added advantage of being human-readable compared to neural controllers that are typically evolved, the results provide a solid platform for future work in composing heterogeneous swarms.