Robust Control Of Robotic Systems Based On Adaptive Dynamic Programming

Robotic system is one of the important components for intelligent manufacturing systems,and also an important equipment to support the transformation and upgrading of manufacturing industry.Since the robot always suffers external disturbances in the process of running,there are usually uncertainties in the system mode.Hence,it is very important to study robust control of robotic system to guarantee high performance control even in the presence of modeling error and external disturbances.Moreover,from the theoretical point of view,it is not easy to solve the robust control problem for nonlinear robotic systems.To improve the operation performance of the robotic system,this thesis will take the serial industrial robot as an example to explore the online solution of derived robust control problem,and carry out simulation verification and experimental validation based on typical robots.The purpose is to develop a new method robust control design for robotic systems and remedy the difficulties of robust control online solution.The main research contents are as follows:(1)Robust control of robotic systems based on adaptive dynamic programming.Firstly,taking the classic PUMA560 series robot as an example,we introduce a dynamic modeling method for serial robots,and then present a SCARA robot platform designed in our laboratory.We also analyze the potential modeling uncertainties in the robotic system models.Then,the problem description of robust control for robotic systems is given.Furthermore,the robust control problem of uncertain robot system is transformed into an optimal control problem of the nominal system,and a new idea to obtain the solution of robust control problem indirectly is given by solving the equivalent optimal control problem.Finally,the Adaptive Dynamic Programming(ADP)method is introduced to online obtain the solution of robust control equation,and a new adaptive law driven by the parameter estimation error is designed to online update the critic neural network(NN)weights,where the convergence of the optimal control solution and neural network weights can be guaranteed.Hence,the actor NN widely used in the other ADP structures is avoided.(2)Robust tracking control of robotic systems based on adaptive dynamic programming.Since the tracking accuracy of robotic system for a given trajectory is an important index to evaluate the robot motion,the robust tracking controller design and its online solution are studied.Firstly,the robust tracking control problem of uncertain robotic system is transformed into the optimal tracking control problem of nominal system.To realize the optimal tracking control,the traditional control design generally divides the original control into two parts(i.e.,steady-state control and transient control).Different to this method,this thesis combines the tracking error and the reference trajectory to construct an augmented system,and then suggest an one-step optimal tracking control design.To guarantee the boundedness of the performance index function,a discount factor is introduced.Finally,a single critic NN is introduced to approximate the optimal performance index function using the idea of ADP,and the corresponding adaptive law is designed to realize the online update of NN weights and optimal control.Hence,the robust control solution can be obtained.(3)Output-feedback robust control of robotic systems based on input/output datadriven.Most of existing robust and optimal control algorithms need complete system states,while some states(such as acceleration)are unmeasurable in robotic system operations,which limits the practical application of some advanced control algorithms.To realize the output-feedback robust control(using the position of joint motion only),we first transform the output-feedback robust control problem of uncertain system into an output-feedback optimal control problem of nominal system,thus avoiding the observer design.Then based on input/output data,a modified algebraic Riccati equation(MARE)can be constructed and two operations(i.e.,Kronecker product and vectorization)are used to reformulate MARE as a linear parametric form.This is adopted to design a novel adaptive law to obtain the solution of this MARE.Considering the fact that the dimension of unknown parameters is too high for multijoint robotics in the online learning process,a dimension reduction operation is introduced to ensure the feasibility of online learning.(4)Output-feedback robust tracking control of robotic system based on input/output data-driven.To realize the output-feedback robust tracking control,an augmented system is constructed by fusing the reference trajectory signal and the system state.Actually,the system output is the error between the reference trajectory and the actual output trajectory,which can be used for online learning.To address this output-feedback robust tracking problem,the output-feedback robust tracking control problem of uncertain augmented system is first transformed into the output-feedback optimal tracking control of nominal augmented system,and then a discount factor is introduced to construct a bounded performance index function.Hence,a new modified tracking algebraic Riccati equation(MTARE)can be obtained.To obtain the solution of MTARE,an adaptive law based on the parameter estimation error is designed to guarantee the convergence of estimated parameters,and the output-feedback robust tracking control problem can be online solved.To verify the effectiveness of the proposed methods,a classical PUMA560 industrial robot model and a SCARA robot experimental platform designed in our laboratory are used for simulation and experiments.Both simulation and experimental results show that the proposed control algorithms have advantages of fast convergence speed,low energy consumption,even in the presence of the modeling error.The experimental results also verify the superiority and potential engineering applicability of the proposed control methods.It is also indicated that the robust control designs and online learning algorithms proposed in this thesis can be further applied to other serial industrial robots and even other intelligent robots.