Research On The Design Of Nonlinear Model Prediction Based Multi-task Controller For Robotics Systems In Remote Handling

It is necessary to maintain the components inside tokamak by remote handling since they will get ruined due to thermal load,electromagnetic load and neutron irradiation during operation.Currently,remote handling robotics systems are controlled manually,with low efficiency,poor accuracy,large variability in operation quality,and difficulty in standardization,which cannot meet the maintenance needs of future commercial fusion reactors.To develop automated and intelligent remote handling robotics systems for future fusion reactors,a series of fundamental problems need to be addressed,one of which is to establish the generic and efficient design theory and implementation tools on multi-task controller for complex robotics systems and diverse maintenance operations.This thesis conducts multi-task controller design research based on nonlinear model predictive control(NMPC),providing a feasible solution for the controller design of remote handling robotics systems,the main contents of which are as follows.To improve the efficiency of solving NMPC problems,a new Newton-type method—the two-step recursive primal-dual interior point method is proposed.In the discretization process of continuous-time NMPC problem,a backward-time implicit discretization scheme is adopted for the system dynamics,transforming the NMPC problem into a nonlinear programming problem(NLP).By analyzing the structure of the KKT conditions corresponding to the NLP,a new costate estimation formula is proposed to convert the coefficient matrix inversion calculation of the original NLP into a two-step recursive operation of first updating the costate backward and then updating all optimization variables forward along the prediction horizon,reducing the computational complexity from cubic power of the number of prediction steps to first power of that.The equivalence of the two-step recursive method and Newton’s method is rigorously proved,with a second-order convergence rate.Furthermore,by reasonably approximating the sensitivity matrix in the costate estimation formula,the parallelization of the two-step recursive algorithm is achieved to further reduce the computational efficiency.The convergence and superlinear convergence rate of the parallel algorithm are also rigorously proved.To ensure the high execution efficiency of the two-step recursive primal-dual interior point method,an NMPC controller design toolbox—NMPCToolkit,is developed.By integrating nonlinear optimization techniques such as regularization,densification,Gaussian-Newton Hessian matrix approximation,merit function and filters-based line search,as well as warm-start and barrier strategies,and selecting algorithm parameters reasonably,the numerical stability and robustness of the code implementation of the two-step recursive primal-dual interior point method are ensured.In addition,to ensure the real-time performance of the controller,a code generation interface tailored for the real-time linux kernel with PREEMPT-RT patch is developed.Performance comparisons with the qpOASES and qpDUNES solvers provided by the mainstream open-source NMPC controller design toolbox-ACADO on the quadcopter’s point stabilization control problem demonstrate that the controller generated by NMPCToolkit have highest execution efficiency with the same computation results.To develop an effective multi-task organization framework,a discussion on the choice of multi-task organizing strategies is conducted.First,the mathematical formulations of priority and weight are presented,which are shifted from simultaneous optimal control to NMPC.Then,the performances of them are investigated based on the mobile manipulator.It is found that when the degree number of multi-tasks does not exceed that of the system,the two strategies have almost the same behavior;when the degree number of multi-tasks exceeds that of the system,the priority strategy sacrifices the completion quality of lower-priority tasks to ensure the completion quality of higher-priority tasks,while the weight strategy balances the completion quality of each task according to their relative weights.Finally,a series of remote operation and maintenance examples,such as the inspection planning of the EAMA robot in the EAST vacuum vessel,the unmanned aerial vehicle cruising in the CFETR vacuum vessel,and the visual servoing of the CMOR heavy-load manipulator in the CFETR vacuum vessel,are presented to demonstrate the usability,efficiency,and versatility of the NMPC-based multi-task controller in the face of complex remote handling robotics systems and various maintenance operations.