A Dirichlet Process Mixture of Robust Task Models for Scalable Lifelong Reinforcement Learning.