Model-free reinforcement learning for robust locomotion using demonstrations from trajectory optimization.