DERL: dynamic estimation reinforcement learning for vehicle path following with road features

This paper addresses path following on curved roads with superelevation and slope while considering vehicle stability. An efficient and safe path following control method is proposed. Based on vehicle parameters, a simple two stage Kalman filter (TSKF) is uilt to estimate the curvature of the vehicle trajectory, enabling accurate recognition of the driving path and the vehicle state with high measurement accuracy and adaptability to different driving environments. Because the road friction coefficient and load affect tire characteristics, a square root cubature Kalman filter (SRCKF) is used to estimate the tire lateral force and obtain a correction factor to design an adaptive adjustment rule for tire cornering stiffness. Vehicle control is described as reinforcement learning (RL) in a Markov decision process (MDP), which determines the steering angle by observing the lateral deviation and the path curvature between the actual and desired paths. The proposed algorithm introduces the path following and road models into the environment to interact with the RL and support learning of vehicle steering. Simulation results show accurate path following and improved driver safety and comfort.