Learning And Control Of Redundant Robots With Unknown Structure

Redundant robots refer to robots that have more joints or degrees of freedom(DOFs)than necessary for a basic task.Due to extra DOFs,redundant robots have higher reliability and flexibility compared to ordinary robots and are often used in industrial production,space exploration,healthcare,and many other fields.However,the complex structure brought by the additional joints and the control uncertainty caused by redundant solutions also take challenges to the efficient control of redundant robots.Additionally,when the structure of the robot is unknown or inaccurate,the control algorithm may generate control signal errors due to the robot modeling error,which reduces the efficiency and accuracy of the robot.Therefore,the accuracy of the structure is crucial in the control of redundant robots,ensuring that the robot operates efficiently,safely,and precisely.Motion-force control of redundant robots is one of the core problems in robot control,especially in cutting,polishing,deburring,and other scenarios that need to maintain contact with the object.However,when the structure of the robot is unknown,it is difficult to achieve efficient motion-force control.Therefore,this paper proposes a data-driven motion-force control(DDMFC)approach.Firstly,the motion-force control problem is solved from the kinematics perspective.Then,the structure parameters involved in the kinematics are estimated efficiently under the condition that the structure of the robot is unknown.In addition,this approach also includes an orientationmaintaining method for the end-effector.Further,a recurrent neural network(RNN)is devised to solve the scheme,and the learning convergence and control convergence of the scheme are proved by theoretical analysis.Simulation and physical experiments running on a 7-DOF redundant robot intuitively illustrate the feasibility and practicability of the scheme.Secondly,a simultaneous learning and control scheme built on the joint velocity level with physical constraints on the decision variable and its derivative,i.e.,joint angle,joint velocity,and joint acceleration constraints,is proposed in this paper.The scheme includes both learning and control strategies and is suitable for the situation where the kinematic structure of the robot is unknown or inaccurate.Furthermore,an RNN is designed to solve the quadratic programming(QP)scheme,and the convergence of learning and control of the RNN is proved theoretically.Simulations and physical experiments on a 7-DOFs redundant robot show that,aided with the proposed scheme,the redundant robot with unknown structure parameters can perform a given inverse kinematics task with high accuracy while satisfying physical constraints on the decision variable and its derivative.Finally,a learning and control scheme for mobile robots based on admittance control is proposed.Firstly,based on the idea of admittance control,the motion-force control problem of maintaining a contact force between the robot and the contact object is established from the perspective of kinematics.Secondly,a control scheme for the mobile robot with unknown structure is constructed on the velocity level,in which the Jacobian matrix required for kinematics is estimated in real time by the gradient-based learning law.Furthermore,the scheme is formulated as a QP problem.On this basis,a dynamic neural network(DNN)is designed to solve the problem.Theoretical analysis proves the learning and control ability of the proposed controller.Finally,the feasibility of the proposed scheme is verified by simulation and physical experiments.