Smooth Exploration for Robotic Reinforcement Learning