Dynamic Locomotion for Humanoid Robots Via Deep Reinforcement Learning