Autonomous Driving Systems Design And Implementation Based On Deep Reinforcement Learning

Autonomous driving technology has always been one of the hot spots in the field of artificial intelligence.The traditional modular method is limited by the complexity of the driving environment,and it is difficult to make a systematic design;the deep neural network method based on supervised learning learns human driving records,although it can imitate human driving actions,but is limited by the extensiveness of driving records,The generalization ability is not strong;The deep reinforcement learning method continuously learns through the interaction between the agent and the environment,and can explore various possible situations in the simulation environment without the support of the data set.Therefore,this article combines deep reinforcement learning and adopts an end-to-end method to study the autopilot system and realize the corresponding autopilot system.Aiming at the problem of low detection efficiency of traditional reinforcement learning in high-dimensional continuous space,we designed a learning method that first imitation phase and then reinforcement phase,simplifying the heterogeneous integration of states.In the imitation stage,driving data learning is used to reduce the dimensionality of high-dimensional image information to low-dimensional image features,together with key features describing the environment,to fully express the state of the vehicle,and use hierarchical integration and connection to heterogeneously fuse these features.In the reinforcement phase,the Deep Deterministic Policy Gradient Algorithm is used and the reward function is tailored for autonomous driving scenarios to guide the learning process.Experiments show that the system can effectively acquire driving skills,and the design based on heterogeneous fusion features can effectively accelerate the training process.The system can not only complete certain driving tasks,but also has the ability to respond to dynamic objects.In order to further optimize the autonomous driving method,we introduce a branch decision network to solve navigation problems.Through the global navigation information provided by the simulation platform,the branch decision-making network is designed into four states: going straight,turning left,turning right,and lane following.Each state is relatively independent for training.The reward function is further optimized for the branch navigation information,and the system performance is tested with reference to the Co RL2017 evaluation standard.The experimental results show that the method we use has certain advantages overall compared with the benchmark method,and the proportion of tasks successfully completed is higher,especially under different weather changes.Aiming at the problem of the lack of interpretability of the neural network,the latent state time series model is used to model the environment,the camera and radar information fusion method is used to further accurately describe the spatial environment information,and the soft actor critic reinforcement learning algorithm is used to further optimize the automatic The decision-making ability of the driving system introduces a two-dimensional semantic bird’s-eye view as a visual interpretation of the system’s understanding of the environment.Experiments show that this method can learn driving skills in a complex environment with a large number of dynamic objects,and the relevant performance indicators are excellent.The semantic bird’s eye view generated by the system can describe the surrounding environment information more accurately,and effectively express the system’s understanding of the current environment.This paper comprehensively researches the design of automatic driving system based on deep reinforcement learning from the aspects of reinforcement learning algorithm framework,input feature form and reward letter design.The experiments on the CARLA autopilot simulation platform have confirmed that the method used in this article can effectively learn autopilot technology and improve the performance of the reinforcement learning algorithm on autopilot tasks.