Comparative studies of evolutionary methods and RL for learning behavior of virtual creatures

Simulation-to-reality is receiving more attention, as demonstrated by the evolution of virtual creatures such as Xenobot. Until now, researchers have conducted experiments on virtual creatures in their own environments. However, Evolution Gym, a benchmark for 2d soft robot simulations, has recently been proposed. Meanwhile, evolutionary methods are used to evolve the morphology of virtual creatures, and reinforcement learning (RL) is frequently used to learn their behavior. Despite the high performance of RL, there are cases where learning does not proceed well, and the evolutionary methods are expected to be effective in such tasks. In this study, we compare the evolutionary methods and RL method to learn the behavior, of the various robot structures and Evolution Gym tasks.