Active safety and ride comfort optimization for autonomous vehicles based on deep reinforcement learning and model predictive control

Abstract

To solve the problems of stiff control actions and insufficient ride comfort during static obstacle avoidance of autonomous vehicles, this paper proposes a hierarchical planning and control method that combines Twin Delayed Deep Deterministic Policy Gradient (TD3) and Model Predictive Control (MPC). The upper-level TD3 generates the target lateral offset and desired speed according to the environmental state. Then, a Bézier curve is used to form a smooth reference trajectory. The lower-level LTV-MPC performs receding optimization under vehicle kinematic constraints and actuator constraints, and outputs smooth and feasible control commands. The simulation results on the CARLA platform show that, compared with the control group without MPC constraints, the proposed method reduces the peak longitudinal deceleration from 4.50 m/s² to 2.92 m/s². The acceleration fluctuation variance decreases from 1.20 m²/s⁴ to 0.72 m²/s⁴, and the lateral trajectory RMSE decreases from 0.40 m to 0.30 m. The results indicate that the cooperation between TD3 and MPC can improve motion smoothness and trajectory execution stability during obstacle avoidance.