DQN-based Path Tracking Control for Intelligent Agricultural Machinery
Aiming at problems of low control precision and difficulty of controller parameter tuning for unmanned agricul-tural vehicles under complex agricultural unstructured field conditions,a deep reinforcement learning based path tracking control algorithm was designed.The DQN(Deep Q-Network)based path tracking controller was established based on five-layer BP neural networks,which realized lightweight and high portability of the network.As input states of the con-troller network,average path curvature of path segment in front of the vehicle was introduced along with steering angle and lateral distance deviation between the vehicle and path control point,which benefit to improve tracking accuracy on curved path.Simulation and field experiments were carried out to verify the convergence of the designed path tracking al-gorithm,and further compared the path tracking performance of two types of networks with path curvature input and with-out path curvature input.For the simulation experiment,the two types of path tracking controller were trained based on a sinusoidal curve,and average distance tracking error of the two type of controller were 0.008 4 m and 0.017 7 m,re-spectively.Field experiment was conducted on U-shaped paths with interval of 6 m,and the results showed that,the tracking accuracy of the controller with path curvature input was higher than that of the model without path curvature in-put.The average distance tracking errors for the two types of path tracking controller were 0.038 9 m and 0.068 4 m,respectively.The effectiveness of the path tracking control method proposed in this paper was verified,and satisfied the requirements of agricultural application.
国家科技期刊平台
NSTL
万方数据
