Nonholonomic Narrow Dead-End Escape with Deep Reinforcement Learning