Combining Model-Based Control And Reinforcement Learning For Versatile And Robust Legged Locomotion