Research On High-precision Motion Calibration Of Industrial Manipulator

In real-world applications of the robot,machine and manufacturing flaws and structure deformation can cause characteristics such as arm's length deviation,angular offset,and linkage center of gravity to differ from theoretical values.The flaws are increasingly exacerbated from the robot base to the end-effector,resulting in robot position discrepancies.The robot fails to achieve the work mission due to a mismatch between the kinematic model parameters and the actual parameters;the control algorithm's failure to track the trajectory adequately.The main research work in this study is as follows to solve the above issues:The PSQP algorithm(Preprocessing Sequential Quadratic Programming)compensates for geometric errors.To address the issue of the LM(Levenberg Marquardt algorithm)method having poor positioning absolute accuracy after compensation,Firstly,using the Preprocessing Algorithm to reduce the rotation deviation of robot base coordinates,reduce the compensation dimension,and convert the high-dimensional solution space into a lower-dimensional subspace with lower non-convexity,improving solution efficiency and accuracy.The problem is then divided into QP subproblems,and geometric error compensation is performed using the SQP(Sequential Quadratic Programming)algorithm.The mean error decreased from 2.6368 mm before compensation to 0.3400 mm after the accurate machine testing,a reduction of 87.10 %The GPR(Gaussian Process Regression)algorithm provides non-geometric error calibration.For the problem of neural networks requiring a large amount of data,the structure needs to be carefully tuned,and the difficulty in selecting the SVM nonlinear fitting kernel function.The GPR algorithm is introduced to train the small-batch data set,and the GPR hyperparameters are tuned by adaptive hyperparameter adjustment to overcome the problem of challenging parameter setting and improve the non-geometric error compensation.After realworld experiments,the mean value of the absolute positioning accuracy was reduced from2.5510 mm to 0.1304 mm,the non-geometric error compensation using GPR decreases 81.34%relative to neural networks and 89.84% relative to SVM.The TIDP algorithm is used to minimize the tracking error.To address the slow convergence of linear combination error and poor interference immunity of PID control,dynamics feedforward control is first used to accelerate the error convergence,improve the dynamic response characteristics,and increase the robot trajectory tracking accuracy.Then,reinforcement learning TD3(Twin Delayed DDPG)is used to further compensate the unmodeled errors and improve the control performance.Finally,the TIDP algorithm combining TD3,dynamics feedforward,and dual-loop PID is obtained to achieve high-precision compensation of industrial robot trajectory errors at the moment loop level.To address the problem of insufficient compensation accuracy caused by the least-squares method for kinetic feedforward model parameter identification and inverse solution matrix pathology,the artificial bee colony algorithm is used to compensate for the kinetic model deviation and improve the kinetic feedforward control accuracy.The experimental results show that the kinetic trajectory tracking error decreases 29.33%compared to the dual-loop PID.