Modes Switching Control Of Variable Structure Wheeled Robot Based On Reinforcement Learning

Two-wheeled robot has the characteristics of simple structure,small and flexible,and low energy consumption,which makes it widely studied.Two-wheeled robots can generally be divided into bicycle robots and Segway robots from the structure.These two kinds of robots have their own unique motion characteristics,and most of the existing research focuses on the single mode of the above two-wheeled robot,which greatly limits the application scenarios of the two-wheeled robot.This paper focuses on a variable structure wheeled robot that can switch between modes such as bicycle and Segway,and studies not only its motion control in different modes,but also its switching control between different modes,so as to improve the adaptability of the two-wheeled robot to the environment.Compared with the common two-wheeled robot,the variable structure wheeled robot has higher state dimension and more control input,which greatly increases the complexity of the system dynamics model and the difficulty of the control algorithm design.The system of establishing the"global" controller in the theoretical sense has the disadvantages of complex structure and large amount of computation of the control algorithm.Firstly,the local dynamic model of the variable structure wheeled robot system is established for different motion modes.Then,local controllers are designed for different motion modes.A combined control algorithm based on synchronous reinforcement learning is proposed for the variable structure wheeled robot bicycle mode.The algorithm adjusts the vehicle speed based on the driving speed condition that maintains the balance of the robot,and the actor network structure designed based on the feedback linearization control law can approximate the optimal control strategy by gradient descent.For Segway mode of variable structure wheeled robot,different controllers are designed using two methods.The model is first decoupled based on partial feedback linearization.In the first method,the cascade sliding mode method is used to complete the system controller design and stabilize the system with zero dynamics.Considering that the controller parameters are difficult to tune and the change of handlebar Angle in the mode switching process will cause the change of system dynamic performance,the second method introduces the TD3 algorithm to adaptively tune the controller parameters.The actor network structure of TD3 algorithm is designed based on PID algorithm,and the actor network parameters are the controller parameters of PID algorithm.After the local controllers of each motion mode were designed,the purpose of mode switching of variable structure wheeled robot was achieved by switching multiple local controllers based on the idea of tossing control.Finally,the design and program design of the measurement and control system of the physical prototype are completed.Different control algorithms are deployed on the physical prototype,and the effectiveness of the proposed algorithm is verified by experimental results.