A Data-Based Approach to Path Following Controller Design for Autonomous Vehicles

Path following controller has played an important role in autonomous driving since it directly affects the driving quality as well as the comfort level of vehicle riders. Previous path following controller faces the difficulties of dynamics model building and parameter tuning, and is not able to mimic the driving style of human drivers. Therefore, this paper proposes a data-based trajectory following controller, which is a model-free controller, and the parameters are self-learned from naturalistic driving data. Moreover, the controller learns the driving style of the experimenters, which make the vehicle riders more comfortable. Naturalistic driving data is collected on a HAVAL H7 vehicle equipped with GPS-RTK system. Deep learning technique is applied to generalize the driving data and then generate the controller. To validate the performance of the path following controller, simulations are conducted using CarSim and MATLAB/Simulink. The results indicate that the controller has a strong ability of generalization, and is able to finish the test scenarios at a high driving quality.