This paper proposes a scheme of trajectory tracking control for the hovercraft.Since the model of the hovercraft is under-actuated,nonlinear,and strongly coupled,it is a great challenge for the controller design.To so...This paper proposes a scheme of trajectory tracking control for the hovercraft.Since the model of the hovercraft is under-actuated,nonlinear,and strongly coupled,it is a great challenge for the controller design.To solve this problem,the control scheme is divided into two parts.Firstly,we employ differential flatness method to find a set of flat outputs and consider part of the nonlinear terms as uncertainties.Consequently,we convert the under-actuated system into a full-actuated one.Secondly,a reinforcement learning-based active disturbance rejection controller(RL-ADRC)is designed.In this method,an extended state observer(ESO)is designed to estimate the uncertainties of the system,and an actorcritic-based reinforcement learning(RL)algorithm is used to approximate the optimal control strategy.Based on the output of the ESO,the RL-ADRC compensates for the total uncertainties in real-time,and simultaneously,generates the optimal control strategy by RL algorithm.Simulation results show that,compared with the traditional ADRC method,RL-ADRC does not need to manually tune the controller parameters,and the control strategy is more robust.展开更多
基金This paper was supported by the National Natural Science Foundation of China under Grant No.61720106010.
文摘This paper proposes a scheme of trajectory tracking control for the hovercraft.Since the model of the hovercraft is under-actuated,nonlinear,and strongly coupled,it is a great challenge for the controller design.To solve this problem,the control scheme is divided into two parts.Firstly,we employ differential flatness method to find a set of flat outputs and consider part of the nonlinear terms as uncertainties.Consequently,we convert the under-actuated system into a full-actuated one.Secondly,a reinforcement learning-based active disturbance rejection controller(RL-ADRC)is designed.In this method,an extended state observer(ESO)is designed to estimate the uncertainties of the system,and an actorcritic-based reinforcement learning(RL)algorithm is used to approximate the optimal control strategy.Based on the output of the ESO,the RL-ADRC compensates for the total uncertainties in real-time,and simultaneously,generates the optimal control strategy by RL algorithm.Simulation results show that,compared with the traditional ADRC method,RL-ADRC does not need to manually tune the controller parameters,and the control strategy is more robust.
