Research On Autopilot Control Algorithm Based On Imitation Learning

Autonomous driving has broad application prospects in transportation,industry,agriculture and other fields.Aiming at the problems of poor real-time performance of traditional automatic driving control algorithm and weak robustness of existing end-to-end automatic driving control algorithm,this paper studies the automatic driving control algorithm based on imitation learning in the background of small commercial four-wheel vehicle under urban road.First,a Carla simulation environment for collecting data is built,and data is collected in the simulation environment to establish a data set required for research work.At the same time,in order to enhance the generalization ability of the data,the image is preprocessed,and the data set is divided into two parts,a training set and a verification set,to provide sufficient data guarantee for subsequent algorithm training and testing.Subsequently,combining the advantages of the automatic driving control algorithm with RGB images as the input predictive control parameters and the automatic driving control algorithm with the bird’s eye view as the input predictive control parameters,an automatic driving vehicle control algorithm based on two-stage imitation learning is designed.The algorithm includes a supervisory network and a perception control network.The supervisory network predicts control parameters by inputting a bird’s-eye view and navigation instructions,and the perceptual control network predicts control parameters by inputting RGB images and navigation instructions.In the first stage,the training set is used to train the perception control network.In the second stage,the trained supervision network is used as a supervisor to further train the perception control network.The trained perception control network is used as the vehicle control model for the final predictive control parameters.The model is tested,and the results showed that the vehicle control model obtained by the algorithm can predict the control parameters more accurately,while taking into account the real-time nature of the model.Finally,in view of the large fluctuations in the value of the predictive control parameters of the vehicle control model obtained by the two-stage imitation learning of the automatic driving vehicle control algorithm and the slow convergence speed during network training,the algorithm is improved,considering the difference between consecutive pictures time information and vehicle dynamics information,an automatic driving control algorithm that integrates temporal and spatial characteristics and vehicle dynamics information is designed.Simultaneously,simulation experiments are performed on the vehicle control model obtained by the algorithm to test the robustness and real-time performance of the model.The experimental results show that the vehicle control model obtained by the automatic driving control algorithm that integrates temporal and spatial characteristics and vehicle dynamics information proposed in this paper does not matter in different weather or in environments with different speeds or different levels of traffic intensity,the completion rate is above 80%,indicating that the model has good robustness.At the same time,the model takes an average of 23.75 ms to predict the control parameters of each frame of data,which meets the real-time requirements of automatic driving tasks.