Data-driven vehicle modeling for path tracking based on the Combination of a Neural Network and Kinematics Model

Abstract

Autonomous driving systems must safely navigate in increasingly diverse and challenging conditions, which necessitate the incorporation of vehicle dynamic models capable of accurately capturing a vehicle's behavior in diverse conditions. Moreover, these models need to be easily and rapidly developed to meet the needs of rapid autonomous driving software updates. Currently used models have limited accuracy, require extensive parameter tuning, and cannot meet these demands. This paper introduces the Combination of Neural Network and Kinematics Model (CNKM). A neural network is utilized to model the nonlinear characteristics of vehicle subsystems (powertrain, braking, steering, tires) and various unknown factors. It ultimately outputs accelerations that are fed into a planar kinematics model to derive the vehicle states. The neural network is trained using a dataset collected from natural driving. A weighting formula suitable for natural driving data is proposed to mitigate the impact of an uneven dataset distribution. This model is compared with commonly used models under typical and high lateral acceleration scenarios, and the position and heading errors of CNKM are 15.34% and 14.71% of those of the nonlinear dynamic model, respectively.