Efficient Deep Reinforcement Learning With Expert Demonstrations for Human–Machine Shared Steering Control Under Emergency Obstacle Avoidance Conditions

作者：Feixiang Xu, Shiyong Feng, Yafei Wang, Junyan Chang, Chen Zhou · 发表于：IEEE Transactions on Vehicular Technology · 年份：2025 · DOI：10.1109/tvt.2025.3629701 · 被引用次数：8 · 研究领域：Vehicle Dynamics and Control Systems、Traffic control and management、Autonomous Vehicle Technology and Safety

Due to the suddenness and complexity of the environment under emergency obstacle avoidance conditions, fully manual driving carries a high safety risk, while fully autonomous driving is prone to decision-making errors. Therefore, it is necessary to develop a human-machine shared steering control (HMSSC) approach to enhance vehicle safety and stability under emergency obstacle avoidance conditions, through combining the human driver's cognitive and decision-making abilities with the control capabilities of automated controllers. Specifically, the reward function is acquired through maximum entropy inverse reinforcement learning (MIRL) method based on expert demonstrations. Subsequently, a deep deterministic policy gradient (DDPG) algorithm based on actor-critic structure is adopted to calculate the reference value of the driving authority, thereby achieving human-machine shared steering control under emergency obstacle avoidance conditions. The human-in-the-loop experiments results indicated that the proposed HMSSC approach significantly improved driving safety and stability of vehicle, and demonstrated strong generalization capability under different emergency avoidance conditions. For example, compared to the Fuzzy-based HMSSC method, the proposed DRL-based HMSSC approach improves vehicle safety and stability by 43.56% and 82.96% under a highway merging emergency avoidance condition.