关键词: 自动驾驶, 转向控制, 路径跟踪, 强化学习, 深度Q学习神经网络

Abstract: To solve the problem of autonomous steering of autonomous vehicles, most researches are mainly based on the model predictive control（MPC） strategy, while the traditional MPC strategy requires an accurate mathematical model of the controlled object and a lot of real-time control calculations. To this end, a steering control strategy based on deep Q-Learning neural network（DQN）reinforcement learning is proposed, which enables autonomous vehicles to track paths accurately and effectively, and improves the accuracy and stability of path tracking. The strategy is based on DQN reinforcement learning to train the agent by selecting an appropriate learning rate, so that the trained agent can adaptively obtain the best front wheel turning angle according to different road conditions and vehicle speeds. The simulation comparison results show that compared with the unconstrained linear quadratic regulator（LQR） control strategy, the cumulative absolute lateral position deviation and cumulative absolute yaw angle deviation of the control strategy based on DQN reinforcement learning have increased significantly. But it is also within an acceptable range, which can effectively improve the accuracy of path tracking. The final real vehicle test results also show the effectiveness of the proposed control strategy.

Key words: automatic driving, steering control, path tracking, reinforcement learning, deep Q-learning network