Reinforcement Learning Control for Target Curvature Tracking in Steer-by-Wire Vehicles
This study addresses target curvature tracking for lateral vehicle motion under low-friction driving conditions. To this end, a reinforcement learning-based steering control framework was developed for steer-by-wire vehicles using a curvature-based steering target. The target curvature was generated from the driver’s steering input and constrained according to the allowable lateral acceleration, thereby accounting for both the driver’s steering demand and lateral dynamic feasibility limits. The control policy was trained using the soft actor-critic algorithm with a progressive curriculum learning strategy, in which the agent gradually experienced increasingly difficult low- and high-friction environments. This training process improved convergence and enhanced the policy’s adaptability to various road surface conditions. Stage-wise mean rewards and reward variances over multiple training runs were analyzed to indicate that the agent adapted to progressively challenging environments. Simulation-based comparisons with conventional controllers suggest scenario-dependent improvements in target curvature tracking performance. The results indicate that the proposed framework provides an interpretable simulation-based reinforcement learning strategy for target curvature tracking under varying road-friction conditions.