“The current focus on supervised imitation learning for robotics is to build a foundation of prior knowledge, which is a necessary prerequisite for effective reinforcement learning.”