Research And Realization Of Auto-driving Strategy In Virtual Environment

With the advent of 5G era and the development of artificial intelligence research,autonomous driving has become the focus of much attention in academia,industry and business circles.The traditional driving strategy model is to carry out mathematical modeling manually.For the complex traffic environment,it cannot be handled well and can not meet the needs of adaptive driving strategy.The autonomous learning ability of cars is the top priority in the research.It is unrealistic for self-driving cars to conduct adaptive training based on specific test sites and road sections,which is costly and inefficient.In order to solve the limitations and problems of traditional driving strategy highly dependent model and current field test,it is necessary to build a safe,efficient,learning and testing autonomous vehicle test platform.In view of the above problems,the work done in this paper is as follows: first,based on TORCS,the simulation system of autonomous driving in virtual environment is built.TORCS software is appropriately improved.In ubuntu16.04 system platform,TORCS is encapsulated by python language based on the Gym environment.The virtual car in the front end is realized to drive intelligently in the TORCS virtual environment.Secondly,the deep reinforcement learning algorithm is used to replace the traditional driving strategy model method,so that the autonomous driving strategy can be learned.This paper USES the DDPG driving strategy algorithm with dual agents,which does not need the vehicle dynamics model,but only needs to learn through interaction with the environment,and is more robust in complex environments.In the simulation system,the comparison of depth q-learning shows the efficiency and feasibility of the method used in this paper.Third,the use of MADDPG architecture to achieve multi-virtual car driving.The autonomous obstacle avoidance driving strategy of DDPG based on dual agents was extended and applied to multiple virtual cars.By concentrating multiple agents on learning and training,and performing them separately,the training time can be accelerated.Compared with the traditional DDPG,the MADDPG used in this paper is more stable in learning and has a shorter learning time,showing a stronger generalization ability.Through research and development,the learning system of autonomous driving in virtual environment is built,which can simulate the actual road condition realistically,and realize the learning effect of autonomous driving strategy in virtual environment.In this paper,the driving strategy learned by the deep reinforcement learning algorithm has achieved a better effect in the virtual environment and achieved the adaptive ability.The intelligence of the real self-driving car is expressed through the autonomous decision-making behavior of the virtual car,and the training process of the real self-driving car is assisted.