Learning From Adaptive Neural Control Of Manipulator With Prescribed Performance

Manipulator,the most important actuator of robot,is commonly used in grabbing,handling and other mechanical work tasks.In these tasks,the end-effectors are required to follow some desired trajectories as close as possible.However,due to the fact of highly nonlinearity and dynamic coupling in robotic system,it is a great challenge to achieve perfect tracking performance.With the development of nonlinear control theory,intelligent algorithms such as neural network control and fuzzy control can deal with the unknown nonlinearity of the system,and also provides new possibilities for obtaining a great control performance of robot manipulator.Although the neural network has the ability to approximate the unknown nonlinearity,when the number of neural network nodes is very large,it is necessary to consume a lot of online adjustment time and resources.So it is important to study the learning ability of neural network which can obtain and store the dynamic information knowledge from the stable control process.Generally,the classical control schemes only focus on a qualitative study for the improvement of control performance.By analyzing the relationship between some controller parameters(or some related items)and the control performance,it concluded that the performance can be improved by modifying the relevant parameters,but there is no quantitative indicator to describe the degree of performance improvement.Therefore,we hope to find a solution that can handle the control issue with prescribed performance.In this solution,the performance constraints can be fused into the control scheme design as a prerequisite.In this paper,we study the trajectory tracking control of manipulator system with the unknown nonlinearity,and mainly focus on the following two points:(1)to achieve the tracking control task under the premise of satisfying the specified tracking error constraints;(2)to study the neural network learning ability of acquiring system unknown dynamic information from the stable control process.Firstly,the modeling of rigid manipulator is carried out,and an adaptive neural control method is proposed based on input state stability and small gain theorem,which can effectively solve the singular problem of controller caused by unknown affine term.Then,adaptive neural control of rigid manipulator with prescribed performance is studied.By introducing an error transformation method,the original restricted tracking control problem is transformed into an equivalent unrestricted stability problem.The proposed controller can guarantee that the tracking error converges to a small neighborhood of zero with satisfying the specified performance constraints.Subsequently,based on the deterministic learning theory,the convergence of the network weights can be obtained from the adaptive control by verifying persistent excitation condition of the regression vector.And the unknown dynamic information is expressed as experienced knowledge in the form of constant neural network weights,so the static learning controller can be constructed with the stored experienced knowledge.For the same or similar control task,a well performance can be obtained due to avoid the repeated weight adjustment process.In addition,we also study the learning control problem of the complex flexible joint robot system with prescribed performance.Finally,the simulation results show the effectiveness of the proposed scheme.