GA-based Co-Design Algorithm with Successive Halving

Co-design is the problem of creating high-performing modular robots by simultaneously optimizing their structure and control. One approach to composing such modular robots is referred to as soft-body robotics. However, designing soft modular robots is a very difficult task, due to the difficulties in controlling them, as well as high degree of freedom when designing them. However, thanks to the recent advancements in machine learning, it is possible to create increasingly more sophisticated controllers for modular robots. As for structural optimization, Genetic Algorithm has been the prevailing method used to automatically generate complex designs. Combining these two approaches makes it possible to solve a variety of tasks that were impossible to solve with conventional methods. Unfortunately, co-design is a complex task for which the optimization process takes a very long time. Therefore, in this paper, we develop a more efficient co-design algorithm, Genetic Algorithm with Successive Halving (GASH), which focuses available computational resources on promising robot designs. We conduct experiments in Evolution Gym benchmark environment. The results show that GASH can find good solutions more quickly in more than half of the tasks. Additionally, we found that the advantages of GASH increase as the number of robot modules increases.