Research On Manipulator Motion Planning And Tracking Control Under Real-time Dynamic Tasks

The development of manipulators expands its task scenarios and types.A class of real-time dynamic tasks have the characteristics of high state dimension,dynamic and time-varying scenes,and strong uncertainty.This kind of task includes space manipulators capturing non-cooperative targets on-orbit,table tennis manipulators hitting table tennis balls,manipulators grabbing moving objects in the air,etc.Designing a feasible motion planning and control system and comprehensively considering the real-time,optimal and robustness of the algorithm is the focus and difficulty of current research.In recent years,data-driven machine learning and reinforcement learning methods have provided new ideas for intelligent planning and control of manipulators.In view of the characteristics and requirements of real-time dynamic tasks,this thesis proposes a motion planning and tracking control algorithm for the manipulator based on machine learning and reinforcement learning algorithms and utilizes the table tennis manipulator as an application scenario for testing.The essential contents and innovations of this thesis are as follows:Firstly,according to the requirements of accurate and fast solution of manipulator kinematics under real-time dynamic tasks,the technical selection of manipulator kinematics solving methods is completed.This part first performs standard DH modeling for KUKA iiwa,a 7-degree-of-freedom redundant manipulator.Then,the forward kinematics equation of the manipulator is given based on the standard DH model,and the inverse kinematics problem is addressed based on arm-angle parameterization.Then,the Jacobian matrix of KUKA iiwa is calculated based on the vector product to obtain the differential forward kinematics equation of the manipulator.By calculating the pseudoinverse of the Jacobian matrix,the differential inverse kinematic problem is addressed.Finally,the first order and differential kinematics solutions of the 7-degree-of-freedom redundant manipulator are verified experimentally.Secondly,for real-time dynamic task scenarios,a motion planning algorithm which can generate optimal reference trajectories satisfying multiple constraints online is proposed.In this part,aiming at the requirements of optimality of motion programming algorithm,the optimization model of the motion planning task of the manipulator is established,the optimal parameterized trajectory is obtained based on the Pontryagin’s minimum principle,and the optimal control problem is transformed into a nonlinear programming problem,and then the trajectory parameters are obtained based on the sequence quadratic programming algorithm.Then,aiming at the problem that the sequence quadratic programming algorithm is sensitive to the initial value,and it has a long optimization time and a certain probability to fall into local optimum,the environmental parameters and the trajectory parameters of the manipulator are recorded as a parameter dataset when several tasks are successful.The parameter regressor is then trained based on Gaussian process regression,then the initial value prediction of the optimization algorithm is realized,the optimization time of the algorithm is reduced,and the real-time performance of the planning algorithm is improved.Then,according to the requirements that the reference trajectory needs to meet complex constraints,a twostage trajectory generator with low computational cost is designed to generate a reference trajectory that meets multiple constraints.Thirdly,aiming at the problem that the modeling of the manipulator is difficult,and the conventional control method is increasingly unable to meet the accuracy requirements when the manipulator moves rapidly in real time,the model-free combination control algorithm with error compensation controller is studied.In this part,aiming at the requirements of high dynamic response performance of the manipulator under real-time tasks,the primary controller of the manipulator is designed based on the fractional PD,and the crow search algorithm is used to adjust the parameters to improve the dynamic response performance and basic tracking ability of each joint.Then,aiming at the problem that there is steady-state error in the non-integral primary controller,as well as the coupling between joints and other uncertainties affect the tracking performance,the dynamic compensation controller is designed based on the twin delayed deep deterministic policy gradient algorithm,and the joint compensation torque is generated in real time according to the state observation,which further improves the tracking performance of the manipulator.In addition,aiming at the problem that sparse reward leads to slow learning or even difficult convergence during controller model training,a hierarchical reward function of deep reinforcement learning,which is combined with sparse reward and formal reward,is designed for the trajectory tracking task of the manipulator.Finally,the offline training and online test of a 2-degree-of-freedom manipulator are realized based on MATLAB/Simulink,and by comparing the test results in different scenarios,it is found that the controller has high accuracy,robustness and dynamic response performance.Finally,on the basis of the research above,a table tennis manipulator simulation system is established and the technical verification is carried out.This part first introduces the composition of the table tennis manipulator system,and completes the simulation system construction based on MATLAB.Then,the application of motion planning and tracking control algorithms on the KUKA iiwa is analyzed,and the overall process of manipulator hitting table tennis is given.Finally,based on the table tennis manipulator simulation system,the result display and analysis are carried out,and the feasibility of the proposed research method is verified.