Reinforcement learning-based decision-making algorithm for motion sickness mitigation in automated driving

Automated driving (AD) requires handling multiple coupled objectives, and comfort-oriented decision-making is especially challenging in curved and dynamic traffic environments where motion sickness (MS) is readily induced. Recent deep reinforcement learning (DRL) methods have demonstrated strong capability in AD planning. Therefore, we propose a hierarchical DRL framework for MS mitigation in AD. The proposed method combines a high-level Proximal Policy Optimization (PPO) decision policy with polynomial trajectory generation, model predictive control (MPC)-based trajectory refinement, and LQR/PID tracking control, enabling modular separation of decision-making and control execution. A 9-degree-of-freedom vehicle dynamics model is adopted to capture translational and rotational motion, and the reward function explicitly incorporates Motion Sickness Dose Value (MSDV) together with roll, pitch, and yaw angular rate penalties. The approach is first validated through real-vehicle tests on a drive-by-wire vehicle to establish practical feasibility, and subsequently evaluated in static and dynamic CARLA scenarios. The dynamic-scenario results show that the proposed PPO policy achieves the best comfort performance among the tested learning-based agents, reducing MSDV by 12.16% relative to the sampling-based Lattice baseline. In the dynamic scenario, PPO requires 3.18 ms per decision and achieves a 99.0% collision-free rate. These results confirm that PPO-based comfort trajectory optimization is viable for real-time deployment on autonomous vehicles, and it helps to address the issues of passengers’ acceptance of autonomous driving and driving comfort.