Research On Automatic Driving Method Based On Environment Modeling And Reinforcement Learning

Intelligent driving is a hot research direction at present.It is difficult to realize large-scale application of L4 level automatic driving vehicles.One of the main reasons is that the current decision-making systems based on scene driven and simple data driven can not cope with the complex and changeable driving environment.In order to study the intelligent decision-making and control technology of automobile,the traditional artificial potential field is improved and extended,and a "unified field model" which includes three vector fields and a scalar field and is suitable for traffic environment is proposed.Based on this model,an automatic driving method combining reinforcement learning and unified field model is proposed.First,the traditional artificial potential field method and its shortcomings is introduced,and the characteristics,composition and demand of traffic environment are analyzed.By deconstructing the advantages of artificial potential field method and the essence of traffic environment,the similarities between field and traffic environment is analyze,and further a targeted subdivision modeling according to the components of traffic environment is put forward,so as to reflect all kinds of attributes of traffic environment as accurately and completely as possible.On this basis,the multi-objective optimization method is used to solve the traffic problem The basic idea of the demand problem in the environment.Next,based on the field theory,a unified field model for traffic environment with the main purpose of reducing the scene dependence is proposed.The model includes the "dynamic field" model which provides the forward expectation of the main vehicle,the "environmental field" model which describes various environmental factors in the traffic environment,and the "intelligent body field" model which describes each agent in the traffic environment.In the process of establishing the model,the factors that affect the traffic environment are analyzed one by one,and expressed through the model and model parameters.At the same time,the calibration method of the model is proposed,and some parameters are calibrated according to the actual traffic data.Finally,the application of unified field model is explored.Aiming at the application of complex scenes which can not be perfectly solved by relying on scene Library in traffic environment,a dynamic planning method combining unified field model and reinforcement learning is proposed.As an example,the test platform and co simulation platform are built.The effectiveness of the algorithm is verified by algorithm comparison and co simulation.Compared with the existing potential field model based on field theory extension,this paper contributes to the establishment of a unified field model with more accurate model and stronger expansibility,and combines and improves the reinforcement learning algorithm for application according to the actual needs,which provides a new idea and method for L4 level automatic driving scheme in complex environment.