Research Of Uncalibrated Visual Servoing Robot Control System

Visual servoing control has been the hotspot in the field of robotics since the late 1990's. Most control methods require accurate calibration of the system model, however in practice, a variety of reasons limit the implementation process. Therefore, the uncalibrated visual servoing control has become the focus of academic research, which does not need to calibrate the system model. That is, there is nothing to do with the robot and camera types.Based on the summary of the development of uncalibrated visual servoing, this paper using the image-based visual servoing control structure establishes the nonlinear mapping relationship between the image feature space and the robot joint angle space from two aspects of dynamic estimation and neural network learning, the function of the compound Jacobian matrix is achieved, and so the calibration of the robot and camera model is avoided. The main work in this paper can be summarized as follows:First, this dissertation adopts the dynamic quasi-Newton approach based on recursive least squares minimization to estimate the compound Jacobian matrix. To overcome the problem that recursive least squares algorithm has poor tracking performance on sudden signals and time-varying signals, variable forgetting factor is introduced, and is adjusted according to changes of estimated errors to improve the rate of convergence and the tracking performance. The simulation experiments are carried out in a 3DOFs robot. Results show that both for static or moving target tracking, the proposed algorithm has higher precision and faster response than the scheme of fixed forgetting factor.Second, BP neural network, due to the great nonlinear mapping ability, is developed to learn the nonlinear relationship between image features and robot angles. Considering the initial weights and thresholds have crucial impact on the network performance and the training efficiency, the genetic algorithm is utilized to optimize the network initial values. This paper constructs image-based visual servoing stable system to obtain training data, and bring the trained network into the control system. The simulation experiments are conducted in the Puma560 robot. Results show that the method can not only guarantees the accuracy but also improves the real-time performance of the control system.Third, the Robotics Toolbox 8.0 Simulink blocksets are used to construct image-based visual servoing simulation system, by which the robot dynamic locating process can be seen clearly. The Genetic Algorithm Toolbox GUI is adopted to optimize initial values of BP neural network, with the help of which the current results are viewed and the spare data can be exported conveniently. The use of Matlab toolboxes can reduce the workload of researchers and has universality to a great degree.