Autonomous vehicle technology has developed rapidly. However, traditional path planning algorithms have problems in dealing with complex and changeable traffic scenarios. Rules based and static maps can not meet the actual needs, and lack the ability to learn from driving experience. On this basis, a new unmanned vehicle path planning method based on neural network is proposed. This project takes Deep Reinforcement Learning (DRL) as the research object to improve the robot’s intelligent decision-making ability. This method uses the method of deep learning to simulate the road condition and vehicle running effectively. On this basis, a path planning method based on adaptive behavior space is proposed, and an intelligent algorithm based on reward mechanism is proposed to make it respond to different traffic conditions. In the first set of tests, the vehicle traveled 55 km, took 45 min and was rated safe, showing that the system was fast and safe in a variety of road conditions.