Research On End-to-End Automatic Driving Technology Based On Deep Reinforcement Learning

The task of automatic driving is that the vehicle senses the road environment through various sensors,and changes the driving behavior in real time without human intervention,including steering,acceleration and braking.Automatic driving can reduce the occurrence of traffic accidents,and improve the utilization rate of road traffic resources and save travel costs.Therefore,the research on automatic driving technology is of great significance.Because the end-to-end automatic driving does not require human-specified rules and directly learns driving behaviors,the research on the end-to-end method is one of the important research directions in the field of automatic driving.Deep reinforcement learning methods are similar to the way humans learn to drive through interactive learning policy with the environment,and are widely used in end-to-end driving tasks.This thesis uses deep reinforcement learning algorithms to study the autonomous driving of vehicles in a virtual environment.This algorithm is based on the improvement of the Deep Deterministic Policy Gradient algorithm.For the problem of low training sample utilization,we combine the prioritized experience replay method with the Deep Deterministic Policy Gradient algorithm.The original sensor input is obtained from the simulation environment,the model outputs continuous acceleration,steering,and braking behaviors,and the training data is stored in the buffer,and the training speed is accelerated by the efficient sampling method of prioritized experience replay.Because deep reinforcement learning requires multiple interactions between the vehicle and the environment.During the training process,wrong driving behavior will occur.In the real environment,training automatic vehicles will cause great damage to the vehicle and the surrounding environment,so this experiment is implemented in the virtual environment of the simulation platform.However,the ultimate goal of end-to-end driving is that real vehicles make driving decisions automatically in the real environment.But,there are huge differences between a structured simulation environment and a complex real environment.This thesis uses the method of image translation to connect the gap between virtual and reality.The CycleGAN network is used to transform the virtual image into a visual appearance similar to the real image.The generated image is used as the input of deep reinforcement learning,and the policy learned in the virtual environment are directly applied to the real world to improve the learning efficiency of the real world.In the deep reinforcement learning experiment,this thesis uses the simulation platform to verify the improved deep reinforcement learning algorithm.It proves that the algorithm can achieve end-to-end automatic driving and speed up the training speed.And this thesis verifies the virtual-to-real image translation model using real data sets,and proves the effectiveness of the image translation migration method.