Vehicle steering control is crucial to autonomous vehicles. However, unknown parameters and uncertainties of vehicle steering systems bring a great challenge to its control performance, which needs to be tackled urgently. Therefore, this paper proposes a novel model free controller based on reinforcement learning for active steering system with unknown parameters. The model of the active steering system and the Brushless Direct Current (BLDC) motor is built to construct a virtual object in simulations. The agent based on Deep Deterministic Policy Gradient (DDPG) algorithm is built, including actor network and critic network. The rewards from environment are designed to improve the effectiveness of agent. Simulations and testbench experiments are implemented to train the agent and verify the effectiveness of the controller. Results show that the proposed algorithm can acquire the network parameters and achieve effective control performance without any prior knowledges or models. The proposed agent can adapt to different vehicles or active steering systems easily and effectively with only retraining of the network parameters.

车辆转向控制对于自动驾驶车辆至关重要。然而，车辆转向系统的未知参数和不确定性给其控制性能带来了巨大挑战，这一问题亟待解决。因此，本文针对参数未知的主动转向系统提出了一种基于强化学习的新型无模型控制器。建立主动转向系统和无刷直流（BLDC）电机的模型以在仿真中构造虚拟对象。建立了基于深度确定性策略梯度（DDPG）算法的Agent，包括参与者网络和评论者网络。来自环境的奖励旨在提高代理的有效性。仿真和测试平台实验被实施以训练代理并验证控制器的有效性。结果表明，该算法无需任何先验知识或模型，即可获取网络参数并实现有效的控制性能。所提出的代理仅需重新训练网络参数即可轻松有效地适应不同的车辆或主动转向系统。