Research On The Control Method Of Autonomous Vehicle Based On The Characteristics Of Human Drivers

The essence of autonomous vehicles is a wheeled mobile robot,which is a comprehensive system integrating environmental perception,pattern recognition,planning and decision-making,and intelligent control.Autonomous driving systems generally have a hierarchical structure,in which the function of the control layer is to receive instructions from the decision-making layer and control each actuator to complete the corresponding actions,to accurately track the path and reasonably control the speed and distance.The traditional closed-loop control system based on error feedback is complicated to adjust the parameters,and the control effect is blunt.The control algorithm represented by model predictive control requires a mathematical analytical model for accurate numerical solution,and the model with higher accuracy often has a higher computational cost,and it is difficult to ensure real-time performance.The main research content of this paper is the lateral and longitudinal control methods of autonomous vehicles based on reinforcement learning.The driving data of skilled drivers are collected in real traffic environments to analyze their driving characteristics.The reward function is designed based on the reverse reinforcement learning method to guide the agent training to behave more like a human driver.The adversarial reinforcement learning is introduced to improve the longitudinal safety of autonomous vehicles.The model-based method is introduced in vehicle lateral control to improve real-time performance while ensuring tracking accuracy.The specific research content and innovation points include the following aspects:(1)Aiming at the key problems of rigid control and poor comfort,a longitudinal and lateral control model of autonomous vehicles based on human driver’s characteristics is proposed.The overall structure and the functions of various modules of "Jinglong VI",an autonomous driving experimental platform,is introduced.The manual driving data collection and experiments in real traffic environment are just run on this experimental platform.(2)An autonomous longitudinal control model is proposed base on the deep deterministic policy gradient(DDPG)algorithm.Aiming at the task of autonomous driving longitudinal control,an apprenticeship learning method combined with IRL is proposed to obtain the real reward function.The reward function for training longitudinal control model is designed based on the analysis of human’s driving data.The networks are designed and trained with a two-car hybrid training method based on the idea of adversarial learning to improve the robustness of the longitudinal control model.Through the analysis of human driving data,the relationship between longitudinal speed and road curvature is fitted,and a curvature adaptive speed adjustment method is realized.Aiming at the shortcomings of unsmooth control of the traditional sub-condition control method at the critical point of the working condition,a virtual particle method is proposed.This method uses a unified longitudinal control model in different working conditions.All longitudinal experiments are run in the Panosim simulation environment,and the results show that the longitudinal control model can ensure safety and comfort,and can conform to the driving habit of human drivers.(3)A model-based reinforcement learning method is proposed to realize the lateral control of autonomous vehicles.Aiming at the current low efficiency of online solution for predictive motion control,and the decline in tracking accuracy caused by model errors with linear model approximation,a non-affine nonlinear model for vehicle motion control is first established.Under the framework,predictive motion control is transformed into an optimization problem.Based on the motion control reference model and cost function,a model-based policy gradient solution method is proposed.The lateral control data of human drivers is analyzed to established the reward function of the lateral control reinforcement learning model.Finally,the policy network training and testing environment is built.The simulation results show that the vehicle lateral control model training converges quickly,and can track accurately even with large curvature trajectory.The simulation results also shows that the lateral control model conforms to the driving habits of human driver when tracking a preset trajectory.(4)The validity and reliability of the lateral and longitudinal control model proposed in the article are verified based on the "Jinglong Ⅵ" in real traffic environment,.The experiments include braking experiment when encountering obstacles,car following experiment on highway,and comprehensive experiment on urban and rural roads.The experimental results show that the lateral and longitudinal control model can realize safe and stable automatic driving,and the control is smooth,which is in line with human driving habits.The safety and comfort of the autonomous control system have been tested in domestic competitions such as the China Unmanned Vehicle Challenge,the World Intelligent Driving Challenge,and the i-VISTA Autonomous Driving Challenge.