Novel Recurrent Neural-Network Strategy For Motion Planning Of Redundant Robot Manipulators

With the development of robot industry,robot technology has attracted more and more attention.As the executive mechanism of most robots,the research of robot manipulator is significant in robot control technology.Redundant robot manipulator refers to the manipulator that has more degrees of freedom(DOF)than the number of which is required to perform the main task of the end-effector.With more DOFs,redundant robot manipulators have greater flexibility,and can perform more complex tasks.The inverse kinematics of the manipulator maps the endeffector task space to the joint angle space.For redundant manipulators,since the dimension of the joint angle space is usually larger than that of the end-effector task space,there exist infinite solutions to the inverse kinematics.Repetitive motion is a typical application of redundant robot manipulators in industrial production process.It requires all joints of the manipulator to return to their initial states after completing one cycle motion with closed end-effector path.If the joints of the manipulator don't return to their initial states after one cyclic motion,this is called joint drift phenomenon in the repetitive motion.Based on quadratic programming(QP),the study on repetitive motion QP scheme of redundant robot manipulators and the corresponding recurrent neural network(RNN)solvers have developed and investigated in this thesis.In the aspect of QP scheme,a velocity-level repetitive motion QP scheme of redundant robot manipulators is firstly constructed,which can effectively solve the joint drift problem.Then,we combine the repetitive motion QP scheme at velocity level with the minimizing torque norm QP scheme of redundant robot manipulators to construct a hybrid-level repetitive motion QP scheme,so that the joint torque can be optimized while the repetitive motion is realized,and the oscillation and divergence of joint angular velocity,acceleration and torque can be avoided.In terms of QP solver,we have proposed a varying parameter recurrent neural network(VPRNN)solver.Compared with the traditional approach to QP scheme,the proposed VPRNN solver has super exponential convergence rate for error,more accurate computation results and stronger robustness.Moreover,an adaptive fuzzy recurrent neural network(AFRNN)is proposed based on the fuzzy control technology and neural dynamics.The proposed AFRNN solver can adjust the convergence coefficient adaptively according to the error,which can further improve the stability of the model while guaranteeing the calculation accuracy.Finally,the theoretical proof,computer simulations and physical experiments demonstrate that the effectiveness of the proposed velocity-level repetitive motion QP scheme and hybridlevel repetitive motion QP scheme.Meanwhile the advantages of the proposed VPRNN and AFRNN solvers are analyzed in terms of end-effector tracking accuracy,joint drift,convergence of error,the vibration and divergence of joint angle velocity,acceleration and torque compared with the traditional approach.