Delay-resilient robust control of automobile active suspensions via deep reinforcement learning

• A delay-resilient DRL method is developed for full-vehicle active suspension. • The method maintains robust vibration suppression under different driving conditions. • The training process setting subject to varying speed and random road excitation. In this paper, we tackle the control problem of the active suspension systems with consideration of vehicle dynamics and parametric uncertainties, and actuator delays. By embedding the long short-term memory (LSTM) into the twin delayed deep deterministic policy gradient (TD3) framework, a delay-resilient deep reinforcement learning (DRL) method is designed for enhancing vehicle ride comfort and stability. To capture the active suspension dynamics, a 14-degree-of-freedom (DOF) vehicle model is employed to establish the training environment subject to uncertain vehicle body mass and time-lags of damping force actuators, and driving conditions. By leveraging the temporal modeling capability of LSTM, the dynamic patterns from historical state–action trajectories are extracted to mitigate the adverse effects of damping forces with time-lags. Moreover, to cope with the temporal misalignment caused by the actuator delays, vehicle dynamics, and parametric uncertainties, we extends the classical TD3 framework with LSTM in generalization tests, maintaining stability and accuracy even under challenging asynchronous delay conditions. Additionally, the simulation results illustrate that the proposed LSTM-TD3 achieves significantly lower RMS values of body vertical acceleration, roll rate, and pitch rate, demonstrating superior vibration control performance in comparison with the baseline methods of passive suspension, DDPG, and TD3.