Exploiting Physics to Learn an Optimal Swing-Up-Strategy for a Variable-Length Pendulum Using Deep Reinforcement Learning