Training Physics-Based Controllers For Articulated Characters With Deep Reinforcement Learning