Learning in Growing Robots: Knowledge Transfer from Tadpole to Frog Robot

Inspired by natural growing processes, we investigate how morphological changes can potentially help to lead and facilitate the task of learning to control a robot. We use the model of a tadpole that grows in four discrete stages into a frog. The control task to learn is to locomote to food positions that occur at random positions. We employ reinforcement learning, which is able to find a tail-driven swimming strategy for the tadpole stage that transitions into a leg-driven strategy for the frog. Furthermore, by using knowledge transferred from one growing stage to the next one, we were able to show that growing can benefit from guiding the controller optimization through morphological changes. The results suggest that learning time can be reduced compared to the cases when learning each stage individually from scratch.