Learning highly diverse robot throwing movements through quality diversity search

The behavior of a robot depends on its environment. Evolutionary approaches can be adapted to look for controllers that are robust to changing conditions [4, 10], but it is hard to guarantee that a single behavior will adapt to any new situation. An alternative approach consists in building a repertoire of behaviors in which to look for behaviors adapted to the context [1, 2]. Having a large set of behaviors instead of a single one further allows for extracting information about the robot, the environment and the task, thus opening the way to a self-built understanding of what makes sense for the robot [3] and to the acquisition of higher level representations that can solve new tasks faster and more efficiently [14].