Modular Soft Robots (MSRs) are ensembles of elastic modules which achieve movement through their synergy of contractions and expansions. The inherent complexity of the dynamics of MSRs poses challenges in hand-crafting effective controllers for task execution. While Artificial Neural Networks (ANNs) have often been employed to this end, they lack transparency, preventing understanding the agent’s inner functionality and problem-solving strategy. To tackle this issue, we propose to adopt interpretable controllers, in the form of graphs, to be optimized with Graph-based Genetic Programming (GGP). This methodology enables the optimization of effective controllers, which can also facilitate gaining insights into information flow and decision-making processes within MSRs. From preliminary experiments, we find our approach feasible and promising.
