Enhancing Vehicle Lateral Stability: A DDPG-Based Active Anti-Roll Bar Control Strategy

Abstract

In recent years, Active Anti-Roll Bars (AARB) have become an important research direction aimed at improving vehicle comfort and lateral stability. Previous studies have indicated that asymmetric issues can occur when using reinforcement learning (RL) algorithms for active AARB, particularly when dealing with positive and negative lateral accelerations. This study explores an AARB control strategy that utilizes a 14-degree-of-freedom (14-DOF) vehicle model and an optimal reward function within the Deep Deterministic Policy Gradient (DDPG) algorithm to ensure symmetric responses to positive and negative accelerations. The simulation employs the DDPG algorithm, focusing primarily on the effects of different reward functions on training and output performance. Through training and analysis with seven distinct reward functions, the final results indicate that the actual output of the Reward_200 function outperforms the other reward functions. The generalization of the DDPG algorithm was validated through simulations using four different sets of parameters for step steer angle tests and sine wave input steer angle tests. In this simulation, DDPG effectively reduced the vehicle roll angle by approximately 43.73% to 82.24%, with no significant asymmetric issues observed in positive and negative lateral accelerations.

Keywords

• Active anti-roll bar
• reinforcement learning
• deep deterministic policy gradient
• proportional-integral-derivative
• vehicle dynamic model