With no well-specified rewards and state transitions that take place in a myriad of ways, training a robot via reinforcement learning represents perhaps the most complex arena for machine learning.