A hormone-based controller for evolutionary multi-modular robotics: From single modules to gait learning

For any embodied, mobile, autonomous agent it is essential to control its actuators appropriately for the faced task. This holds for natural organisms as well as for robots. If several such agents have to cooperate, the coordination of actions becomes important. We present an artificial homeostatic hormone system which is a bio-inspired control paradigm. It allows to control both, a single robot as well a set of cooperating modules in multi-modular reconfigurable robotics. Our approach is inspired by chemical signal-processing and hormone control in animals. Evolutionary computation is used to adapt controllers for two distinct morphological robot configurations (uni-and multi-modular), different environmental conditions, and tasks. This approach is compared to artificial neural networks. Our results indicate, that the proposed control paradigm is well adaptable to different robot morphologies and to different environmental situations. It is able to generate behaviors for several robotic tasks and outperforms neural networks in terms of evolvability in the tested multi-modular robotic setting tested.