Intelligent suspension control using DDPG in 14-DOF vehicle dynamics framework

This study proposes an active suspension control strategy based on deep reinforcement learning, where a DDPG algorithm is incorporated into a comprehensive 14-DOF full-vehicle dynamic framework. The model accounts for both linear and rotational motions of the vehicle body, including longitudinal, lateral, and vertical movements, as well as roll, pitch, and yaw dynamics, in addition to the dynamics of the wheels. The interaction between the tire and road surface is modeled with the Magic Formula, which makes a more accurate representation of the nonlinear contact mechanics involved in vehicle dynamics possible. The DDPG agent is trained in a simulation environment with randomized inputs, such as road displacement excitation, double-lane change maneuver, and emergency acceleration or braking scenarios. Its observation space consists of 34 dimensions, which simulate the data inputs from various onboard sensors, including vehicle attitude sensors, wheel displacement sensors, and road-surface laser sensors. The action space was defined by four continuous suspension control signals that directly influenced the dynamic response of the vehicle. The reward function was designed to improve ride comfort, roll stability, and pitch stability by reducing key dynamic parameters, such as vertical velocity, vertical acceleration, roll angle, pitch angle, and their corresponding angular velocities. By comparing the performance of the proposed using the traditional control methods—including passive suspension, PID control, and a semi-active suspension based on a skyhook-type algorithm—across 1×10^6 randomized simulation episodes, the results show that the DDPG-based controller performs much better than these methods, with an improvement of up to 20.6% in metrics of comfort and stability. These findings demonstrate that the DDPG agent has strong adaptability under complex and highly dynamic road conditions.