Improving Local POE Kinematics Calibration Utilizing Genetic Algorithm Ahmad Suryo Arifin 1 , Marcelo H Ang Jr 2 , Ma Chongyu 3 , Lim Chee Wang 4 1,2 National University Singapore, 3,4 Singapore Institute of Manufacturing Technology 1 g0800239@nus.edu.sg , 2 mpeangh@nus.edu.sg , 3 cwlim@SIMTech.a-star.edu.sg , 4 cyma@SIMTech.a-star.edu.sg Abtract: This paper presents genetic algorithm to optimize poses selection to improve the performance of kinematics calibration. Genetic algorithm is used to optimize the poses selection while local POE (Product of Exponential) formula is used to represent the joint kinematics. Kinematics calibration using local POE will be updated and calibrated using an iterative least square algorithm. This research focus on how to find an optimal set of poses so that we can reduce the number of poses used for the calibration without affecting the calibration performance. Observability index are used to evaluate the optimality of the set of poses. In order to find the optimal set of poses, observability index is used in the genetic algorithm fitness function. In our experiment, we use the 7-DOF Mitsubishi PA-10 manipulator as the platform and LEICA laser tracker for measurement. The experiment demonstrates that genetic algorithm can reduce the number of poses without affecting the calibration performance. Keyword: Local POE, Kinematics Calibration, Genetic Algorithm, Pose Optimization 1. Introduction An accurate kinematics model of the manipulator could improve the control performance of the motion at the end effector. However, the accuracy of the kinematics model might reduce due to the manufacturing error, link misalignment, and assembly error at the manipulator. Kinematics calibration gives an effective solution to improve the performance of the motion control of the manipulator. There are several works which are related to the kinematics calibration of robot manipulators. Wang, C,B et al [1] made use of a forward calibration method. The forward calibration identified the actual parameter of the manipulator based on the measurement at the workspace of the manipulator. Doria, A et al [2] introduced inverse kinematics using B-splines and multivariate parametric approximating splines functions as tools to do the calibration. Chen, I.M, et al [3] proposed a least square method to calibrate the manipulator. Local product of exponential is used to model the kinematics. Several works related to this method are also proposed in [4],[5], and [6]. Although the local POE calibration method provides significant improvement, this technique requires a lot of poses. In general, more poses could improve the kinematic calibration of the manipulator. To optimize the number of poses for calibration, several techniques are introduced. Chernoff, H [7] minimized the trace of non-zero singular values of the Jacobian matrices of the manipulator. This method is called A- Optimality. Wald [8] maximized the determinant of the non-zero singular values of the Jacobian matrices of the manipulator. This method is introduced as D-Optimality. Smith [9] proposed G-Optimality, it minimizes the maximum prediction variance of the non-zero singular values of the Jacobian matrices. Ehrenfeld [10] introduced E-Optimality. E-Optimality maximizes the minimum value of the non-zero singular value of the Jacobian matrices. Sun, Y, et al [11] compared the observability index above. It is shown that G-Optimality and E- Optimality are the best formula to minimize the uncertainty at the end effector position of the manipulator, and D-Optimality is the best formula to minimize the variance of the parameter. The objective of this paper is to optimize the poses selection so that it can improve the kinematics calibration performance. Genetic algorithm is selected as the method to optimize the poses selection while the local product of