Toward Deep Reinforcement Learning Without a Simulator: An Autonomous Steering Example