AutoMoDe-Cedrata: Automatic Design of Behavior Trees for Controlling a Swarm of Robots with Communication Capabilities

Behavior trees are a control architecture that has gained recent attention in AI and robotics. Previous research on the use of behavior trees in swarm robotics has shown the necessity for the behaviors to have proper return values, instead of running indefinitely. This work extends our previous work in which we defined AutoMoDe-Cedrata, an automatic modular design that makes use of modules that have been explicitly defined for behavior trees. While the search space is sufficiently large to include well-performing solutions, Cedrata had problems discovering communication-based strategies. In this work, we extend Cedrata by introducing Cedrata-GP and Cedrata-GE which are based on genetic programming and grammatical evolution, respectively. We test these design methods on two missions and compare the performance of the automatic design methods against the performance of solutions created by human designers. The results show that the structure of Cedrata allows for well-performing solutions that are reliably found by human designers. However, the automatic design methods fail to discover the same communication strategies as the human designers.