Speeding Up Reinforcement Learning with a New Physics Simulation Engine