1. Introduction
A series robotic arm is an open loop motion chain composed of multiple links, with a wide range of motion, multiple functions, and a wide range of applications. As the production scenarios in which tandem robotic arms participate become more complex, offline programming methods are increasingly used in industry. This approach requires a high level of absolute accuracy of the tandem robotic arm to ensure that the tandem robotic arm will accurately reproduce the programmed end effector (EE) pose [
1]. However, the absolute accuracy of the current tandem robotic arm is poor (only 2~3 mm). Among the many factors that affect the absolute motion accuracy of the tandem robotic arm, the geometric error, that is, the EE pose error caused by the kinematic parameter error, accounts for about 80% of total errors. Geometric error refers to the deviation between the actual kinematic parameters and theoretical kinematic parameters of the tandem robotic arm caused by the errors in the manufacturing and assembly process of the parts of the tandem robotic arm, movement under continuous high load, collision, etc. [
2]. Therefore, identifying and compensating the kinematic parameter errors of the tandem robotic arm through kinematic calibration is of great significance to improve the absolute accuracy of the tandem robotic arm.
The main process of kinematics calibration is as follows: first, establish the kinematic model of the tandem robotic arm, including the kinematic forward solution model and the kinematic inverse solution model, then analyze the kinematic parameter error in the model and establish the functional relationship between the kinematic parameter error and the EE pose error of the tandem robotic arm; second, the appropriate measurement scheme is determined according to the function model to obtain the EE pose error of the tandem robotic arm; third, the EE pose error is brought into the function model to solve the kinematic parameter error; finally, a joint driving angle solution strategy is established for the positive kinematics model with errors. The above four steps are called kinematic modeling, EE pose measurement, kinematic parameter identification and error compensation [
3]. In the following section, the research status of each step will be introduced and an appropriate scheme will be proposed according to the characteristics of the tandem robotic arm used in this experiment.
Kinematic modeling: The essence of kinematic modeling is to establish the transformation matrix relationship between each coordinate system of the tandem robotic arm and the EE of the tandem robotic arm. It is the premise and basis for parameter identification. The model generally requires integrity, continuity, and parametric minima [
4]. The most basic and typical kinematic model is the Denavit–Hartenberg (D-H) model [
5]. The model uses four parameters to describe the kinematic parameters between adjacent joints, but when the adjacent joints are parallel or nearly parallel, the D-H model is discontinuous. Hayati [
6], Veitschegger [
7] et al. proposed a modified D-H model (MDH), adding a rotation around the y-axis between adjacent parallel axes in the D-H model to describe the non-parallelism of the two axes, solving the singularity problem of the D-H model. Stone proposed the S model of the three-translation and three-rotation 6-parameter model, which uses six parameters to describe the spatial transformation relationship between adjacent links. Compared with the D-H model, it is more intuitive, but it is not convenient for error compensation [
8]. Zhuang and Rot proposed a 4-parameter CPC model and a 6-parameter MCPC model with three translations and one rotation in 1992. The CPC model is named for its “Complete and Parametrically Continuous” properties; the MCPC model is the modified CPC model. The continuity of the model is ensured by increasing the kinematic parameters. Since its coordinate changes are relatively gentle, it will not cause the sudden change of the pose of the tandem robotic arm [
9]. Lv et al. established the forward kinematics model of the serial manipulator by using the screw method and verified its correctness using ADAMS software simulation. Compared to the D-H method, the screw method only needs to establish a tool coordinate system and a base coordinate system, which is simpler and more intuitive [
10]. Chen et al. proposed a POE exponential product model (Product of Exponentials) based on the screw theory for the modular tandem robotic arm [
11]. The POE model avoids the influence of the error caused by the transformation matrix on the calibration accuracy and makes the experimental process simpler. However, the calibration results of this method cannot be directly compensated to the tandem robotic arm control system and can only be converted into a D-H model first; in addition, the Jacobian matrix is difficult to solve, so the numerical solution can only be obtained by an iterative method.
EE pose measurement: There are various methods for measuring the EE pose of the tandem robotic arm, including laser interferometer, three-coordinate measuring instrument, telescopic ball bar, industrial camera, etc. Jian et al. [
12] used a laser tracker to measure the five-degree-of-freedom tandem robotic arm for parameter identification and correction. The experimental results show that the compensation error has a significant effect. Xie et al. [
13] used the linear structured light measurement system installed on the EE as a tool to obtain the three-dimensional coordinate information of the standard sphere center, established an error model according to the invariance principle of the spherical center coordinate in the basic coordinate system, and completed the calibration study of the tandem robotic arm. Wang [
14] obtained the visual features of the robotic arm operation trajectory through hand eye calibration and the least squares method, achieving end pose monitoring. The final experiment showed that this method can effectively reduce pose and height errors. Yang et al. proposed a dynamic angle correction scheme that uses Creaform’s C-Track780 optical coordinate measuring machine to measure angle information online as feedback to accurately guide the end link of the tandem robotic arm to reach the expected spatial coordinate point [
15]. Legnani et al. developed a line sensor-based measurement platform system and applied it to an anthropomorphic tandem robotic arm by obtaining isotropic precision and high sensitivity, then using the same measurement system to perform calibration experiments [
16]. Using three airborne cameras and relying on a coupled model combining kinematics and photogrammetry, Aitor proposed an analytical method to estimate the uncertainty of certain kinematic parameters, thereby enabling the transition from qualitative to observability to quantitative. Evaluation therefore becomes possible, and calibration of the tandem robotic arm is completed [
17]. Jiang et al. used deep learning methods to establish neural network structures related to measurement and theoretical pose, and the predicted accuracy after training is comparable to the ranging accuracy of laser trackers [
18]. Lu Yi and others completed the calibration of an Advantech LNC-S600 robot based on the linear structured light sensor; the average position error was reduced from 1.7256 mm to 0.3412 mm [
19]. Luo Zhenjun et al. completed the calibration of the stacking robot based on a wire displacement sensor [
20]. Balanji H. M. et al. calibrated the robot using a monocular camera; after calibration, the absolute position error and direction error of the robot converged to 2.5 mm and 0.2°, respectively [
21]. Rozlivek J. et al. used a binocular camera to identify the pose information of the marked object fixed on the end effector of the multi chain robot and obtained the pose information of the multi chain robot, completing the error identification of the multi chain robot [
22]. The above EE pose measurement methods have all been successfully applied in the kinematics calibration of tandem robotic arms, and each has its own advantages and disadvantages. For example, laser trackers and three-coordinate measuring instruments have high precision, but their cost is high, the measurement time is long, and this measurement method usually requires other auxiliary tools to be installed at the EE of the tandem robotic arm; despite this, the accuracy of the end auxiliary tools is often far superior. The measurement accuracy is lower than that of the laser tracker and the three-coordinate measuring instrument, resulting in redundant measurement accuracy. The visual measurement method needs to understand the principle of visual measurement, but the cost is cheap, and as a non-contact measurement tool, its measurement is more convenient and more efficient.
Parameter identification and error compensation: The process of parameter identification and error compensation is essential to solve the problem of optimal parameter error under constraint conditions and to compensate the optimal parameter error to the tandem robotic arm control system. Li et al. [
23] proposed a calibration method of the manipulator by combining a Kalman filter and quadratic interpolation beetle antenna search algorithm; this combination can search the optimal motion parameters. Experiments show that this method has high calibration accuracy. Wang et al. [
24] used the matrix method to analyze the calibration error model of kinematics parameters and added incremental error compensation; this can improve the accuracy of the manipulator. Urrea et al. [
25] used the least squares method to simulate the 3-degree-of-freedom selective flexible assembly tandem robotic arm SCARA, obtained the algorithm performance index capable of parameter identification, and obtained a higher-precision end pose through error compensation. To reduce the influence of measurement noise and improve the calibration performance, Chen et al. [
26] proposed an improved full-pose measurement and recognition optimization method. The method is based on the adaptive particle swarm optimization algorithm, has observable indicators and identification accuracy indicators, and considers external constraints, namely structural interference, and angle constraints; the simulation results show that the proposed error model has 36 error parameters, and has good stability. Dolinsky [
27] introduced an inverse static kinematics calibration technique based on genetic programming, which is used to establish and identify model structure and parameters; this avoids the problems of traditional calibration methods and has the potential to identify real calibration models. Lattanzi et al. [
28] developed a new geometric identification program for two kinds of industrial manipulators. The program allows the modification of the theoretical values of robot kinematics parameters. Experiments have proved the effectiveness of the geometric calibration method. Shi [
29] et al. used a binocular vision system to measure the end position of the manipulator, established theoretical and actual distance error function, and used particle swarm optimization to solve the kinematics parameter errors and compensate for them. Finally, the absolute accuracy of the manipulator was significantly improved. Xing et al. [
30] combined spin theory and the exponential product method to establish a distance error model for a UR10 robotic arm. The structural parameters were calibrated using a laser tracker on an experimental platform, and the local accuracy error was less than 1 mm. The Levenberg Marquardt (LM) algorithm [
31] is a widely used improved least squares method that combines the gradient method and the Gaussian Newton method and has high robustness. Zhao [
32] used an LM algorithm to complete parameter identification for the Eft ER50 robot; after calibration, its average position error was reduced from 2.038 mm to 0.136 mm. Chen [
33] used a monocular camera and a standard sphere as tools to solve the identification equation using the first-order difference quotient method and completed the robot calibration based on point and distance constraints. Nguyen, H. N. et al. [
34] used artificial neural networks to compensate for robot position errors caused by non-geometric errors and applied this method to the experimental calibration of the HH880 robot. The calibrated robot position accuracy has been effectively improved. Yan [
35] used vector matrix analysis to obtain the mapping relationship between attitude error and structural parameter error and used a genetic algorithm to optimize the attitude accuracy of parallel robots. Finally, the simulation results show that, when compared with the least square method, the error identification algorithm of robot kinematics parameters based on genetic algorithm has obvious advantages. Kosmatopoulos et al. [
36] solved the identification problem of manipulators using dynamic neural networks, and proposed a dynamic back-propagation scheme that can learn and identify nonlinear systems without requiring prior knowledge of the system for identification. This method can handle the sudden change of input data, the error convergence progress is fast, and, even when the input waveform does not appear, the network can still operate effectively after training.
However, there are still problems with the current research on calibration methods for robotic arms. At present, most of the robot kinematics parameter calibration methods use the joint length parameter and angle parameter identification method at the same time. Although a calibration method that only introduces position constraints can improve the positioning accuracy of the robot to a certain extent, it does not include angle accuracy as an evaluation indicator of the calibration results, and therefore cannot guarantee high angle accuracy. The angle error is usually smaller than the position error, and when measuring the pose error, the angle error is easily drowned out, resulting in poor calibration results. In addition, when using the least squares method for parameter identification, the identification equation is prone to pathological problems. Therefore, error identification will be affected, and compensation accuracy will also decrease.
In this paper, in view of the excellent characteristics of binocular vision that can obtain the full pose of the EE of the tandem robotic arm, and its low measurement cost, simple measurement and high efficiency, binocular vision is used to measure the pose of the EE of the tandem manipulator. Three ceramic standard spheres were used as measurement markers, and the EE pose of the tandem robotic arm was obtained in combination with the self-designed supporting measurement aids. At the same time, the MDH model (Modify-DH) solves the discontinuity problem of the D-H model and allows the obtained kinematic parameter errors to be easily compensated into the control system of the tandem robotic arm. The kinematics model in this paper selects the MDH model for modeling; based on the kinematic parameters of the tandem robotic arm corner control method, an error compensation strategy is established. Through the analysis of the error mapping model, it is found that the length parameter error has no effect on the end angle error, and the angle joint parameters and the length joint parameters are separated and calibrated; in addition, this article cites an ill-conditioned optimization method based on spectral correction iterative method, which completes parameter identification.
Compared to previous research, this article proposes new methods to solve existing problems. First, aiming at the problem of poor absolute accuracy of tandem robotic arms, a separation calibration scheme for kinematic parameter errors is formulated based on binocular vision, which improves the absolute position accuracy and absolute angle accuracy of tandem robotic arms. Second, aiming at the problem that the identification equation is prone to be ill-conditioned, an ill-conditioned optimization method based on the spectral correction method is adopted to make the error identification and compensation more accurate.
The structure of this paper is as follows. The introduction describes the research background and research status of this paper. In the first section, the kinematics modeling of the 6R series manipulator is carried out, including the positive and inverse kinematics solutions and the mapping relationship between the pose error of the end-effector of the serial manipulator. Next, according to the established error mapping relationship, the joint angle parameter error and the length parameter error are separated, and the pose constraint scheme and distance constraint scheme are established. Finally, the identification equation is optimized. The second section discusses the measurement principle of binocular stereo vision and the setting of measurement markers. The third section carries out the simulation experiment and the real machine experiment of the serial manipulator and discusses the experimental results. The conclusion part summarizes the research content and significance of this paper and proposes future research directions.