You are on page 1of 14

Control Engineering Practice 141 (2023) 105691

Contents lists available at ScienceDirect

Control Engineering Practice


journal homepage: www.elsevier.com/locate/conengprac

Coordination control of the automated fiber placement system using


photogrammetry-based leader–follower approach✩
Ronghua Zhang a,d , Yaonan Wang a,d , Wenfang Xie b ,∗, Tingting Shu b , Haoran Tan a,d ,
Yiming Jiang c,d
a
College of Electrical and Information Engineering, Hunan University, Changsha, 410082, China
b Department of Mechanical, Industrial and Aerospace, Concordia University, Montreal, H3G2W1, Canada
c School of Robotics, Hunan University, Changsha, 410082, China
d National Engineering Laboratory of Robot Visual Perception and Control Technology, Changsha, 410082, China

ARTICLE INFO ABSTRACT

Keywords: The use of multiple robots in an automated fiber placement system (AFPS) makes it possible to manufacture
Automated fiber placement system composite structures with complex shapes. The robots in such system usually adopt independent trajectory
Leader-follower planning and control methods. This makes the AFPS lack coordination, and rely heavily on the high-
Coordination Control
precision calibration of the system when completing complex fiber placement tasks. In this article, a novel
Multi-robot collaborative manufacturing
photogrammetry-based leader–follower (PBLF) approach is proposed for a particular AFPS which consists of a
serial manipulator and a parallel robot to improve its accuracy and coordination. The approach is composed
of an offline trajectory planning module, a leader module, and a follower module. The trajectory planning
algorithm in the offline trajectory planning module solves the problem of robot workspace constraints and
mandrel occlusion in the fiber placement tasks for complex structure. Besides, a novel photogrammetry-based
pose tracking method is proposed to make the follower robot track its desired relative trajectory with respect
to the moving reference frame and keep the coordination of the two robots. The theoretical stability of the
proposed approach is proved. The results of experiments indicate that the proposed approach can reduce the
average mean absolute errors of the relative pose in position and orientation direction by 62.15% and 46.42%,
respectively. While the corresponding average root mean square errors are reduced by 61.20% and 45.26%,
respectively. Therefore, compared with Non-Coordination approach, the PBLF approach can effectively improve
the coordination of the two robots.

1. Introduction two tremendous obstacles still prevent its application. One obstacle
is the path positioning accuracy of the robot, which affects course
Compared to metallic materials, composite materials have some at- overlaps, course gaps, and tow-end placement errors (Jeffries, 2013).
tractive features such as high mechanical performance and low density, Another obstacle is the relatively low coordination of the two robots
which have made these materials widely used in many fields such as in the system, which directly affects the performance of the fiber
automobiles, aerospace, precision instruments, etc. Aized and Shirin- placement tasks.
zadeh (2011). With the aim of enhancing the quality and manufacturing In practice, the accuracy of robots can be improved in the follow-
efficiency of composite structures, robots were integrated into the fiber ing three methods: (1) installing extra high-accuracy sensors at each
placement system in various ways (Oromiehie, Prusty, Compston, & robot joint (Devlieg, 2010); (2) calibrating robot with the assistance of
Rajan, 2019; Xie, 2016; Zhang et al., 2022). Among these systems, the advanced algorithms and high-accuracy metrology instrument (Filion,
AFPS (Xie, 2016), consisting of two robots with different mechanical Joubair, Tahan, & Bonev, 2018); (3) correcting/tracking the pose of
structures, makes it more favorable for manufacturing components with robot in real-time (Stadelmann, Sandy, Thoma, & Buchli, 2019). The
complex shapes. Despite the obvious potential advantages of the AFPS,

✩ This document is the results of the research project funded by the Natural sciences and engineering research council of Canada under Grants N00892, the
Major project of Xiangjiang Laboratory under Grant 22xj01006, the Special funding support for the construction of innovative provinces in Hunan Province under
Grant 2021GK1010, and the General project of Xiangjiang Laboratory under Grant 22xj03002.
∗ Corresponding author.
E-mail addresses: zrh796@hnu.edu.cn (R. Zhang), yaonan@hnu.edu.cn (Y. Wang), wfxie@encs.concordia.ca (W. Xie), t_shu@encs.concordia.ca (T. Shu),
tanhaoran@hnu.edu.cn (H. Tan), ymjiang@hnu.edu.cn (Y. Jiang).

https://doi.org/10.1016/j.conengprac.2023.105691
Received 9 March 2023; Received in revised form 27 August 2023; Accepted 31 August 2023
Available online 9 September 2023
0967-0661/© 2023 Elsevier Ltd. All rights reserved.
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Table 1 the problems of joint limits and singularity. However, the strategies
The comparison of methods for improving robot accuracy.
mentioned above usually require accurate system calibration and have
Advantages Disadvantages yet to consider directly controlling the relative positioning among the
1st method Low computation Hardware constraint or robots. Leader–follower based approach avoids those disadvantages and
high hardware cost (Yang
has been widely applied to many multi-robot tasks (Huang, Zheng,
et al., 2017)
2nd method Low hardware cost Poor performance in
Wang, Ota, & Zhang, 2020; Luh & Zheng, 1987; Meng et al., 2014; Shen,
non-static conditions Pan, Ahmad, & He, 2021). The main idea of leader–follower based
(Stadelmann et al., 2019) approach is to determine one or more robots as the leader robot and
3rd method Suitable for robot in High real-time the other robots as the follower robot when executing multi-robot coor-
motion state requirements
dination tasks. The leader robot is the reference object of the follower
robot, and the follower robot ensures the coordination between robots
by keeping a certain distance and direction value with the leader robot.
advantages and disadvantages of these methods are shown in Table 1. In (Luh & Zheng, 1987) and Huang et al. (2020), leader-follower based
Since the AFPS used in our research contains 13 degree of freedoms approach was proposed for robot motion control and redundant dual-
(DOFs) and the fiber placement processes are under dynamic condi- arm assembly task, respectively. As elaborated in (Liang, Ge, Liu, Ling,
tions, applying the 1st and the 2nd methods to improve the system & Liu, 2021), the formation tracking problem of networked marine
accuracy will inevitably lead to excessive costs and poor performance surface vehicles was investigated. The successful application of leader-
in non-static conditions, respectively. Fortunately, the disadvantages of follower concept in the coordination control of multi-robot systems
the 3rd method can be compensated by employing external sensors indicates that applying this concept to the coordination control of the
with high real-time performance (Li, Li, Tian, & Liao, 2023; Wu & AFPS is feasible and has significant potential applications.
Kuhlenkoetter, 2022; Yang, Wang, Fan, Dong, & Zhou, 2017). For these Inspired by the aforementioned work, a novel photogrammetry-
reasons, the third method is a suitable solution to enhance the accuracy based leader–follower approach for the AFPS is proposed to improve
of the AFPS. the relative accuracy and coordination of the fiber placement system.
The pose correcting/tracking issue of robots has recently been stud- This strategy combines the concepts of leader–follower based coordi-
ied in Cong and Hanh (2023), Jia, Liu, and Liu (2015), Li, Ding, nation and pose tracking. The main contributions of this paper are
Zhao, and Liu (2023), Li, Ding, et al. (2023), Li, Zhang, Li, Tian, summarized as follows:
and Wang (2022), Shu, Gharaaty, Xie, Joubair, and Bonev (2018). (1) A photogrammetry-based leader–follower control structure for a
For instance, a dynamic path tracking (DPT) scheme was proposed particular AFPS is proposed. Within the structure, a photogra-
in Shu et al. (2018) to realize the three-dimensional dynamic path mmetry-based PD controller is designed for the leader robot to
tracking of a robot. With the DPT scheme, the accuracy of line-tracking track the offline generated trajectories, and a novel photogra-
and circle-tracking were improved. In (Li et al., 2022), a feedforward mmetry-based pose tracking controller for the follower robot is
compensation strategy and a joint closed-loop feedback controller were developed to enhance the coordination of the AFPS.
designed for an industrial robot. With the purpose of achieving high
(2) An offline trajectory planning (OTP) algorithm for the rotary
position accuracy of industrial robots in precision tasks, a continuous
stage and the leader robot based on relative manufacturing tra-
dynamic time warping-based real-time error calculation method was
jectory is proposed. The algorithm can ensure that the structure
developed in Li, Ding, et al. (2023). With the improvement of the com-
is in a manufacturing-friendly pose.
puting performance of controller hardware, many advanced methods
(3) Compared to most coordination strategies for dual-robot sys-
were proposed for pose correcting/tracking (Bilal, Unel, Tunc, & Gonul,
tems (Bai et al., 2023; Zhang et al., 2018), only the relative
2022; Cong & Hanh, 2023; Jia et al., 2015; Tang, Zhou, Zakeri, Shu,
manufacturing trajectories are required, which can be derived
& Xie, 2023). For example, in Jia et al. (2015), a model-free iterative
easily from third-party commercial software, and the coordina-
learning control algorithm is studied for a 6-DOF robotic manipulator
tion of the system is free from strict calibration of the relative
with a monocular camera mounted on its end-effector. After five iter-
pose of two robots.
ations, the trajectory tracking performance is significantly enhanced.
However, as a consequence of the computational load, the maximum The structure of this paper is organized as follows. In Section 2,
frequency of the controller is 10 HZ, which is far from the requirement the preliminaries and problem formulation are introduced. The PBLF
for high-performance tasks. On the basis of the research (Shu et al., approach for the AFPS is discussed in Section 3. Then, in Section 4,
2018), an adaptive Neuro-PID Control algorithm and a robust Kalman the simulation and experimental results of the proposed approach
filter were developed to improve robot accuracy and robustness of are presented. Finally, the conclusion and future work are given in
the system (Tang et al., 2023). It is worth pointing out that pose Section 5.
correcting/tracking controllers were designed for single-robot systems,
and the desired trajectories were generated with respect to a static 2. Preliminaries and problem formulation
coordinate frame.
Currently, a lot of approaches to improve the coordination of mul- 2.1. The model of the AFPS
tiple robots have been proposed (Bai et al., 2023; Carnevale, Camisa,
& Notarstefano, 2023; Dong Sun & Mills, 2002; Feng, Hu, Sun, & An overview of the AFPS is shown in Fig. 1. The AFPS is composed
Soon, 2020; Meng, Dimarogonas, & Johansson, 2014; Sun, Dong, Qin, of three subsystems, the manufacturing subsystem, the pose adjust-
& Wang, 2020; Xiao & Philip Chen, 2021; Xiao, Yu, & Chen, 2022). ment subsystem and the photogrammetry measurement subsystem.
According to the existence of agents in the multi-robot system, robot co- The manufacturing subsystem includes a serial manipulator (Fanuc
ordination methods can be classified into decentralized based approach M20-iA), and a fiber placement head. Its main function is to lay
and leader–follower based approach (Feng et al., 2020). Decentralized composite materials onto the mandrel. The pose adjustment subsystem
based approach is an effective way to improve the coordination of is dedicated to adjust mandrel pose, which includes a parallel robot
multiple robots (Carnevale et al., 2023; Sun et al., 2020). For instance, (710LP-6-500-220), and a rotary stage. Due to the existence of this
vision guided dynamic synchronous control of dual manipulators was subsystem, the system is more flexible than others and can manufacture
explored by a cross-coupled sliding mode control protocol (Bai et al., the structure with complex shapes. The third subsystem is the pho-
2023). In Zhang, Xie, and Hoa (2018), a semi-offline synchronization togrammetry measurement subsystem, which includes a fixed bracket
algorithm was presented for the AFPS. The proposed algorithm solves and a photogrammetry sensor (C-track 780) providing pose signal of

2
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Fig. 1. The structure and coordinate frames of AFPS.

the robots. Besides, two computers are used to implement the controller Fig. 2. In this paper, the differential kinematics model is used to design
and image processing in this system. During the fiber placement tasks, the controller of the serial manipulator, while the kinematics model is
the parallel robot holds the mandrel through three-jaw-chuck on rotary used to design the controller of the parallel robot.
stage. By controlling the parallel robot, the pose of the mandrel is The serial manipulator is an industrial robot with 6-DOFs. Its differ-
adjusted in real-time. At the same time, the fiber placement head held ential kinematics model is expressed as
by the serial manipulator is controlled to follow the planned trajectory
𝒗𝑠 = 𝑱 𝑔,𝑠 𝒒̇ 𝑠 (1)
so that the tow from the fiber placement head is laid onto the mandrel.
For this AFPS, the accuracy of each robot and the relative motion where 𝒗𝑠 is the end-effector velocity of the serial manipulator, 𝑱 𝑔,𝑠 is
accuracy of the two robots are important factors affecting the composite the geometric Jacobian matrix, which is a function of the joint vector
component’s quality. 𝒒 𝑠 , and 𝒒̇ 𝑠 is the joint velocity.
The parallel robot is composed of an upper platform, a base platform
2.2. Coordinate frames of the AFPS and six kinematic chains. More precisely, those chains have the same
structure, and the two links in each chain are connected by a spherical
The coordinate frames of the AFPS are shown in Fig. 1. All coor- joint. One end of a chain is connected to one corner of the base platform
dinate frames in this system obey the right-hand rule. The coordinate by a revolute joint, and another end of the chain is connected to the
frame of the C-track is named photogrammetry sensor frame 𝑭 𝑝𝑠 , and corresponding corner of the upper platform by spherical joint. The
its origin is located in the center of the two cameras. The base frames inverse kinematics solution of the parallel robot is represented as
of the parallel robot and the serial manipulator are defined as 𝑭 𝑝𝑏 and 𝒒 𝑝 = 𝐼𝑛𝑣𝐾(𝒑𝑝 ) (2)
𝑭 𝑠𝑏 , respectively. The homogeneous transformation matrix 𝑯 𝑝𝑠 𝑝𝑏
from
the base frame of the parallel robot 𝑭 𝑝𝑏 to the photogrammetry sensor where 𝒒 𝑝 is the joint vector of the parallel robot, and 𝒑𝑝 is the end-
frame 𝑭 𝑝𝑠 is fixed and known. Similarly, the homogeneous matrix from effector’s pose of the parallel robot.
the base frame of the serial manipulator 𝑭 𝑝𝑏 to the photogrammetry
sensor frame 𝑭 𝑝𝑠 is defined as 𝑯 𝑣𝑠𝑏 . The end-effector frame of the 2.4. Photogrammetry measurement model
parallel robot 𝑭 𝑝𝑒 is defined at the center of the parallel robot’s upper
In the proposed approach, the real-time feedbacks of the pose
platform, and its homogeneous transformation matrix 𝑯 𝑝𝑠 𝑝𝑒 from the
information of the two end-effectors are used to calculate poses of the
end-effector frame of the parallel robot 𝑭 𝑝𝑒 to the photogrammetry
two end-effectors with respect to the photogrammetry sensor frame
sensor frame 𝑭 𝑝𝑠 can be computed by using the measurements of
𝑭 𝑝𝑠 . All the pose information are detected by C-Track 780 which has
the C-track. The end-effector frame of the serial manipulator 𝑭 𝑠𝑒 is
two cameras. C-Track 780 is capable of tracking many point-shaped
defined at the top of the fiber placement head, and its homogeneous
reflectors at the same time. The photogrammetry measurement model
transformation matrix 𝑯 𝑝𝑠 𝑠𝑒 from the end-effector frame of the serial is shown in Fig. 3 . If Assumption 1 is held, the perspective projection
manipulator 𝑭 𝑠𝑒 to the photogrammetry sensor frame 𝑭 𝑝𝑠 can also
of the reflectors on the camera coordinate is expressed as
be obtained through the measurements of the C-track. The reference
𝑐 𝑐
frame 𝑭 𝑟𝑒𝑓 is attached to the rotary stage, and its origin and 𝑧 axis are 𝒅 𝑟,𝑖𝑗 = 𝑪 𝑗 𝑯 𝑝𝑠𝑖 𝒅 𝑝𝑠
𝑟,𝑖 (3)
identical with the origin and the 𝑧 axis of 𝑭 𝑝𝑒 , respectively. Its rotation 𝑐
angle with respect to 𝑭 𝑝𝑒 can be obtained through a high-precision where 𝒅 𝑟,𝑖𝑗
is the projection coordinates of the 𝑖th reflector on the
encoder. 𝑗th camera frame 𝑭 𝑐𝑗 , 𝒅 𝑝𝑠𝑟,𝑖 is the homogeneous coordinates of the
corresponding reflector with respect to the photogrammetry sensor
𝑐
frame. The 𝑪 𝑗 and 𝑯 𝑝𝑠𝑗 are the camera matrices and the homogeneous
2.3. Kinematic model of the AFPS
transformation matrices from the photogrammetry sensor frame 𝑭 𝑝𝑠 to
𝑐
the 𝑗th camera frame 𝑭 𝑐𝑗 . The 𝑪 𝑗 and 𝑯 𝑝𝑠𝑖 are known after the C-Track
The robot kinematics and differential kinematics models are widely
780 is calibrated. Then, the homogeneous coordinates of 𝑖th reflector
used to design robot controllers. Since two robots are relatively inde-
is shown as
pendent, the models of the AFPS can be decomposed into two parts.
The mechanism of the serial robot and the parallel robot is shown in 𝒅 𝑝𝑠 𝑝𝑠 𝑜
𝑟,𝑖 = 𝑯 𝑜 𝒅 𝑟,𝑖 (4)

3
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Fig. 2. The mechanical structure of two robots.

[ ]
𝑅𝑜𝑡(𝒑𝑝𝑠
𝑜 ) 𝑇 𝑟𝑎𝑛𝑠(𝒑𝑝𝑠
𝑜 )
𝑯 𝑝𝑠
𝑜 = (5)
𝑂1×3 1
where 𝒑𝑝𝑠
𝑜 is the pose of object frame 𝑭 𝑜 (the end-effector frame)
with respect to the photogrammetry sensor frame 𝑭 𝑝𝑠 , 𝒅 𝑜𝑟,𝑖 is the
homogeneous coordinate of the 𝑖th reflector with respect to the object
frame, and 𝑅𝑜𝑡(∙) as well as 𝑇 𝑟𝑎𝑛𝑠(∙) are the rotation function and the
transform function (Corke, 2017), respectively.

Assumption 1. The objects are rigid, and the reflectors are also rigidly
stuck on each object. During the fiber placement tasks, the reflectors are
in the field of view of the photogrammetry sensor.

The pose of the object contains six independent variables. Thus,


at least six independent equations are needed to solve the variables.
The conclusion of the (Yuan, 1989) points out that the solution of the
equations is unique under the condition that there are at least four co-
planar points. In this paper, more than four reflectors are employed
to improve the robustness of pose detection. The coordinates of these
reflectors 𝒅 𝑜𝑟,𝑖 with respect to object frame 𝑭 𝑜 are known. Thus, the pose
of the object can be obtained by solving a series of above-mentioned Fig. 3. The Photogrammetry measurement model.
equations. It is worth noting that in our experiments, we use C-track
and probe to locate the coordinates of reflectors with respect to the
object frame by means of special points with distinctive geometric
𝜙(𝑘) is the angle between the 𝑧 axis of the end-effector frame 𝑭 𝑠𝑒 and
features in the object (Zhang et al., 2018).
the reference frame 𝑭 𝑟𝑒𝑓 .
(ii) With the help of the photogrammetry sensor and the encoder in
2.5. Problem statement
rotary stage, designing a coordination control algorithm to make the
pose errors of the parallel robot 𝒆𝑝 (𝑘) → 0, and the relative pose errors
The manufacturing trajectories are usually planned in a reference
of the serial manipulator with respect to the reference frame 𝒆𝑟 (𝑘) → 0.
frame for convenience and generalization for the fiber placement tasks.
These trajectories are also named desired relative manufacturing tra-
jectories. Our research aims to design an approach that can be used for 3. Photogrammetry-based leader–follower approach for the AFPS
keeping the components in a manufacturing-friendly pose and ensure
that the motion of the serial manipulator with respect to the reference As far as this paper is concerned, the use of two robots improves the
frame is consistent with the desired trajectories. Specifically: ability of the AFPS to manufacture structures with complex geometric
(i) Given the desired relative trajectories 𝒑𝑑𝑟,𝑠 , planning the desired shapes, but at the same time, the coordination between the two robots
trajectories of the parallel robot and rotary stage (𝒑𝑑,𝑝 and 𝜃𝑑 ) so that must be considered. In this section, the PBLF approach for the AFPS
the components are in a manufacturing-friendly pose which satisfies: is described. Specifically, the control structure of the PBLF approach
(1) the pose of each robot can be easily detected by the photogram- is presented in Section 3.1. Estimation methed for end-effector’s pose
metry sensor and the manufacturing points can be easily reached by is given in Section 3.2. In Section 3.3, The offline trajectory planning
the serial manipulator. algorithm is proposed for the leader robot and rotary stage to make the
(2) The height of the manufacturing point and the 𝜙(𝑘) − 𝜋2 are components in manufacturing-friendly poses during the fiber placement
within or close to a certain range as far as possible, which is tested tasks. Two robot controller are designed in Section 3.4 and Section 3.5
to be an important condition for guaranteeing the quality of tasks. The to accurately track the trajectories generated by the OTP algorithm

4
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

and enhance the coordination of the system, respectively. Finally, the it possible that the reflectors at the robot end-effectors will not be
stability of the system is proved in Section 3.6. obscured by the robot and components. Thus, the sensor can monitor
the robot’s motion throughout the fiber placement tasks.
3.1. The control structure of photogrammetry-based leader-follower ap-
Although, the pose of the object can be obtained by the above
proach
method, the presence of detection noise which affects the pose accuracy
is inevitable (Janabi-Sharifi & Marey, 2010). In order to reduce the
The kinematic-based control technique performs well when robots influence of noise and disturbance, the Kalman filter is employed to
do not require too fast motions or rapid accelerations (Di Lillo, estimate the object pose. The current state vector at time instant 𝑘 is
Pierri, Antonelli, Caccavale, & Ollero, 2021; Lauer, Lerke, Blagojevic, denoted as
Schwieger, & Sawodny, 2023; Siciliano, Sciavicco, Villani, & Oriolo,
𝒎𝑘 = [𝑥(𝑘), 𝑥(𝑘),
̇ 𝑦(𝑘), 𝑦(𝑘),
̇ 𝑧(𝑘), 𝑧(𝑘),
̇
2009). This technique has also been proved to be effective in improv- 𝑇 (6)
𝛾(𝑘), 𝛾(𝑘),
̇ ̇
𝛽(𝑘), 𝛽(𝑘), 𝛼(𝑘), 𝛼(𝑘)]
̇
ing the pose accuracy of AFPS (Jeffries, 2013). The kinematic-based
controllers of the two robots are designed for the following factors: The current observation vector obtained from photogrammetry sensor
(1) When the AFPS works, the maximum speed of the two robots are is represented as
around 0.02 m/s which is lower than those (0.1 m/s) of most robot
𝒏𝑘 = [𝑥𝑚 (𝑘), 𝑦𝑚 (𝑘), 𝑧𝑚 (𝑘), 𝛾𝑚 (𝑘), 𝛽𝑚 (𝑘), 𝛼𝑚 (𝑘)]𝑇 (7)
tasks; (2) the contact force between the fiber placement head and the
mandrel surface is around 20 𝑁 which is lower than those around 150 Then the dynamic and output model of discrete time system are written
𝑁 in some robot handling tasks; (3) Fanuc M20-iA does not provide as
torque control access. Since the manufacturing trajectories used for the
𝒎𝑘 = 𝑨𝒎𝑘−1 + 𝝃 𝑘 (8)
tasks are usually planned under constraints such as the limit of velocity
and acceleration, and the initial pose errors of each robot are not set
too large, the physical limits of the robot’s input of the robot such 𝒏𝑘 = 𝑾 𝒎𝑘 + 𝜼𝑘 (9)
as velocity and acceleration do not need to be specifically considered where 𝑨 ∈ R12×12
is the state R6×12
( transition
) matrix, 𝑾 ∈ is the
when designing the controller. observation matrix, 𝝃 𝑘 ∼  0, 𝑸𝑘 is the disturbance noise vector,
( )
The PBLF approach provides a novel robot cooperative mode for 𝜼𝑘 ∼  0, 𝑹𝑘 is the observation noise vector. Then the Kalman filter
AFPS. The control structure of PBLF approach is shown in Fig. 4. algorithm for pose estimation is summarized as following steps.
In AFPS, the parallel robot acts as leader, and tracks the trajectory
generated by the OPT algorithm. The main function of the leader is Prediction equations
to adjust the pose of the component to facilitate manufacturing. The
̂ 𝑘|𝑘−1 = 𝑨𝒎
𝒎 ̂ 𝑘−1 (10)
serial robot acts as the follower to maintain the coordination of the two
robots. To reduce the effects of calibration errors of AFPS and external
disturbances, the poses of two robots calculated from the signals of 𝑷 𝑘|𝑘−1 = 𝑨𝑷 𝑘−1 𝑨𝑇 + 𝑸𝑘−1 (11)
the sensors are used to control the serial robot. More precisely, in Kalman gain
the offline trajectory planing module, the offline trajectory planning ( )−1
(OTP) algorithm is proposed to provide the desired trajectory of the 𝑲 𝑘 = 𝑷 𝑘|𝑘−1 𝑾 𝑇 𝑾 𝑷 𝑘|𝑘−1 𝑾 𝑇 + 𝑹𝑘 (12)
leader robot for complex fiber placement tasks. In the leader module,
Estimate update
the leader robot is controlled independently by an inverse kinematics
( )
based proportional-derivative controller to follow the desired trajectory ̂𝑘 =𝒎
𝒎 ̂ 𝑘|𝑘−1 + 𝑲 𝑘 𝒏𝑘 − 𝑯 𝒎
̂ 𝑘|𝑘−1 (13)
generated by the OTP algorithm. In the follower module, the motion of
follower robot will be corrected in real-time according to the motion 𝑷 𝑘 = 𝑷 𝑘|𝑘−1 − 𝑲 𝑘 𝑾 𝑷 𝑘|𝑘−1 (14)
of the reference frame 𝐹𝑣 . To track the desired relative manufacturing
trajectory of follower robot, a novel photogrammetry-based pose track- where 𝒎 ̂ 𝑘|𝑘−1 is a priori estimate of 𝒎𝑘 based on all priori measurement
ing controller is designed. Moreover, The Kalman filter is employed to except the one at time instant 𝑘, 𝒎 ̂ 𝑘 is a posteriori estimate of 𝒎𝑘
estimate the pose of the objects in different frames. In a task, the leader based on all available measurements at time instant 𝑘, 𝑷 𝑘|𝑘−1 is a priori
robot should have an advantage in load capacity, while the follower covariance matrix, 𝑷 𝑘 is a posteriori covariance matrix.
robot should have strengths in precision, flexibility, and workspace. As mentioned in Section 2.2, the base frame of the parallel robot
Therefore. We choose the parallel robot as the leader robot and the 𝑭 𝑝𝑏 with respect to 𝑭 𝑝𝑠 is fixed and known. Thus, the pose 𝒑𝑝 can be
serial robot as the follower robot. calculated from the corresponding homogeneous transformation matrix
𝑯 𝑝𝑏
𝑝𝑒 as
3.2. Estimation of the end-effectors’ pose ( )−1
𝑝𝑠
𝑯 𝑝𝑏
𝑝𝑒 = 𝑯 𝑝𝑏 𝑯 𝑝𝑠
𝑝𝑒 (15)
The pose of the end-effector of the leader robot with respect to the
where 𝑯 𝑝𝑠 𝑝𝑒 can be computed using the filtered pose data of the end-
base frame of the parallel robot 𝒑𝑝 and the pose of the end-effector
effector frame 𝑭 𝑝𝑒 .
of the follower robot with respect to the moving reference frame 𝒑𝑟,𝑠
After the transformation matrices between end-effector’s frames and
are used to design the leader robot controller and the follower robot
the photogrammetry sensor frame 𝑭 𝑝𝑠 were estimated, the relative pose
controller, respectively. The pose 𝒑𝑝 can be detected directly by C-
of the two end-effector is calculated as
track. However, the pose 𝒑𝑟,𝑠 can only be obtained indirectly, for the ( )−1
reason that the reflectors in the reference frame are not in the field of 𝑯 𝑝𝑒
𝑠𝑒 = 𝑯 𝑝𝑒
𝑝𝑠
𝑯 𝑝𝑠
𝑠𝑒 (16)
view. In order to obtain the pose 𝒑𝑟,𝑠 during the entire fiber placement
where 𝑯 𝑝𝑠 𝑠𝑒 is also computed using above-mentioned method.
process, the C-track and high-precision encoder are used to detect the
The high-precision encoder in the rotary stage can measure the ro-
corresponding information.
tation relationship between the end-effector frame of the parallel robot
𝑭 𝑝𝑒 and the reference frame 𝑭 𝑟𝑒𝑓 , so the homogeneous transformation
Remark 1. The maximal measuring volume of C-track is 7.8 𝑚3 . By
matrix 𝑯 𝑝𝑒 𝑟𝑒𝑓
from 𝑭 𝑟𝑒𝑓 to 𝑭 𝑝𝑒 is expressed as
placing the photogrammetry sensor in the right pose (as shown in
[ ]
Fig. 1), the robot can be guaranteed to move within its field of view. 𝑇
Rot 𝑧 (𝜃𝑓 ) 𝑂1×3
The proposed offline trajectory planning algorithm in Section 3.3 makes 𝑯 𝑝𝑒
𝑟𝑒𝑓
= (17)
𝑂1×3 1

5
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Fig. 4. The structure of PBLF approach for the AFPS.

where 𝜃𝑓 is the filtered rotation angle of the rotary stage, by using function, and 𝑚𝑜𝑑(𝑎, 𝑏) returns the remainder after diphotogrammetry
the Kalman filter. Combining Eqs. (16) and (17), the homogeneous of 𝑎 by 𝑏.
transformation matrix 𝑯 𝑟𝑒𝑓
𝑠𝑒 from 𝑭 𝑠𝑒 to 𝑭 𝑟𝑒𝑓 is computed as The planned trajectory of the parallel robot should guarantee that
( )−1 the components is in a manufacturing-friendly pose. Besides, for the
𝑝𝑒
𝑯 𝑟𝑒𝑓
𝑠𝑒 = 𝑯 𝑟𝑒𝑓 𝑯 𝑝𝑒
𝑠𝑒 (18) purpose of improving the stability of the system, the trajectory of the
leader robot should not have drastic change. The trajectories of the
the pose 𝒑𝑟,𝑠 is also calculated from the corresponding homogeneous
parallel robot’s end-effector along 𝛽 and 𝑧 directions are designed as
transformation matrix 𝑯 𝑟𝑒𝑓
𝑠𝑒 . {
𝛽𝑝 (𝑘 − 1) + 𝑘𝑡1 𝛽𝛥 (𝑘) COND One
3.3. Offline trajectory planning for the leader robot 𝛽𝑝 (𝑘) = (20)
𝛽𝑝 (𝑘 − 1) otherwise
{
During the fiber placement process, the limitation of robot 𝑧𝑝 (𝑘 − 1) − 𝑘𝑡2 (𝛥𝑧(𝑘)) + 𝑘𝑡3 𝛽𝛥 (𝑘) 𝑧min < ||𝑧𝑟 (𝑘)|| < 𝑧max
𝑧𝑝 (𝑘) =
workspace and the shape of the structures create difficulties in the 𝑧𝑝 (𝑘 − 1) otherwise
process, such as collision and singularity. To solve these problems, the
(21)
parallel robot and rotary stage in the AFPS are used to adjust the pose of
the structures. However, for some specific mandrel, it is hard to decide where
when and how to adjust the parallel robot and the rotary stage. In the ( )
similar scene of multi-robots coordination tasks, the researchers and 𝛽𝛥 (𝑘) = sgn 𝜙𝑚 (𝑘) (𝜙(𝑘) − 𝜙(𝑘 − 1)) (22)
engineers usually generate multiple robot trajectories through manual
teaching which relies heavily on operation experience. 𝜋
𝜙𝑚 (𝑘) = 𝜙(𝑘) − (23)
In this research, the trajectories are generated offline, which can sig- 2
nificantly reduce computational complexity during the process. The rel-
( )
ative manufacturing trajectory 𝒑𝑑𝑟,𝑠 = 𝑥𝑟 (𝑡), 𝑦𝑟 (𝑡), 𝑧𝑟 (𝑡), 𝛾𝑟 (𝑡), 𝛽𝑟 (𝑡), 𝛼𝑟 (𝑡) 𝛥𝑧(𝑘) = 𝑧𝑟 (𝑘) − 𝑧𝑟 (𝑘 − 1) (24)
between two robots with 6-DOFs is pre-planned according to the
method proposed by Hély, Birglen, and Xie (2017). During the fiber
placement process, the trajectory of each robot is constrained by the COND One: |𝛽(𝑘)| < 𝛽max and 𝜙min < ||𝜙𝑚 (𝑘)|| (25)
relative manufacturing trajectory. When the relative manufacturing
In Eqs. (21) and (25), 𝑧min , 𝑧max , 𝛽max , 𝜙min are the constraints in
trajectory and the leader robot trajectory are known, the trajectory of
trajectory planning. In Eq. (22), 𝜙(𝑘) is the angle between the 𝑧 axis of
the follower robot is determined. The generated trajectories are defined
serial robot end-effector frame and the 𝑧 axis of the reference frame.
in the reference frame. In order to simplify the fiber placement process,
The proposed trajectory planning for the rotary stage and the parallel
the rotary stage and four parameters of the parallel robot pose are
robot is summarized as Algorithm 1.
adjusted to make the manufacturing point in front of the end-effector of
the follower robot. Hence, the values of the parallel robot trajectory in
𝑥, 𝑦, 𝛾 and 𝛼 direction are zero. More precisely, the trajectory of rotary 3.4. Proportional derivative control in leader module
stage is expressed as
The desired trajectory of leader robot at time instant 𝑘 is denoted
⎧𝜃𝑏 (𝑘) + 𝜋2 sign(𝑦𝑟 (𝑘)) 𝑥𝑟 (𝑘) = 0, and 𝑚𝑜𝑑(𝜃𝑏 (𝑘), 2) = 0 as 𝒑𝑑,𝑝 (𝑘) which is generated in the offline trajectory planning module,
⎪ then the corresponding desired joint angle 𝒒 𝑑,𝑝 (𝑘) can be computed as
⎪𝜃𝑏 (𝑘) 𝑥𝑟 (𝑘) = 0, and 𝑚𝑜𝑑(𝜃𝑏 (𝑘), 2) = 1
𝜃(𝑘) = ⎨ ( )
𝑦𝑟 (𝑘)
⎪ 𝜃𝑏 (𝑘) + arctan 𝑥𝑟 (𝑘) ≠ 0, and 𝑚𝑜𝑑(𝜃𝑏 (𝑘), 2) = 0 𝒒 𝑑,𝑝 (𝑘) = 𝐼𝑛𝑣𝐾(𝒑𝑑,𝑝 (𝑘)) (26)
( 𝑥𝑟 (𝑘) )
⎪𝜃 (𝑘) + arctan 𝑦𝑟 (𝑘) − 𝜋
sign(𝑦𝑟 (𝑘)) otherwise
⎩ 𝑏 𝑥𝑟 (𝑘) 2 the actual pose of leader robot 𝒑𝑝 (𝑘) at time instant 𝑘 can be computed
(19) using the method mentioned in Section 3.2. Then, the actual joint
( ) vector 𝒒 𝑝 is computed as
where 𝜃𝑏 (𝑘) = 𝜋2 𝑓 𝑖𝑥 2𝜃(𝑘−1)
𝜋
, 𝑓 𝑖𝑥(∙) is a function that rounds each
element of 𝑋 to the nearest integer toward zero, sign(∙) is the sign 𝒒 𝑝 (𝑘) = 𝐼𝑛𝑣𝐾(𝒑𝑝 (𝑘)) (27)

6
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

To track its desired trajectory, a PD controller is designed as

𝒖𝑝 (𝑘) = 𝒌𝑝1 𝒆𝑝 (𝑘) + 𝒌𝑝2 (𝒆𝑝 (𝑘) − 𝒆𝑝 (𝑘 − 1)) (28)


Assumption 2. The relative manufacturing trajectory is well pre-
where 𝒆𝑝 (𝑘) = 𝒒 𝑑,𝑝 (𝑘) − 𝒒 𝑝 (𝑘), 𝒌𝑝1 and 𝒌𝑝2 are control gain matrices, planned, so that it does not contain the singular pose of the robots,
𝒖𝑝 (𝑘) is joint velocity vector of the leader robot. and it can guarantee that matrix 𝑇𝑟𝑎 is of full rank.
Algorithm 1 Offline trajectory planning (OTP) algorithm for the rotary
stage and the parallel robot Remark 2. The singularity of the working Jacobian matrix is closely
Input: related to the singularity of the robot Jacobian matrix and matrix 𝑇𝑟𝑎 .
Relative manufacturing trajectory for AFPS, If Assumption 2 is satisfied, the work Jacobian matrix is reversible.
(𝑥𝑟 (𝑡), 𝑦𝑟 (𝑡), 𝑧𝑟 (𝑡), 𝛾𝑟 (𝑡), 𝛽𝑟 (𝑡), 𝛼𝑟 (𝑡));
To implement relative pose tracking task, a photogrammetry-based
Output:
pose tracking (PBPT) control law is designed as
Desired trajectory of the rotary stage 𝜃𝑑 (𝑘);
( )
Desired trajectory of the parallel robot 𝒑𝑑,𝑝 (𝑘); 𝒖𝑠 (𝑘) = 𝒌𝑠1 𝑠(𝑘) + 𝒌𝑠2 𝒆𝑟 (𝑘) + 𝑱 −1 (37)
𝑤 𝒑𝑑𝑟,𝑠 (𝑘 + 1) − 𝒑𝑑𝑟,𝑠 (𝑘)
1: Planning the desired trajectory of the rotary stage using Eq. (19) ;
2: Planning the four parameters of the parallel robot so that make the where 𝒖𝑠 (𝑘) is the motor speed vector of the serial manipulator, 𝒌𝑠1 , 𝒌𝑠2
manufacturing point at the front of the end-effector of the follower and 𝝀 are control gain matrices which are tuned by trial and error in
robot during the process; the simulation and experiment. The term 𝒌𝑠1 𝑠(𝑘) is used to improve the
robustness of the system, and term 𝒌𝑠2 𝒆𝑟 (𝑘) is used to quickly reduce
𝑥𝑝 (𝑘) = 0; 𝑦𝑝 (𝑘) = 0; 𝛾𝑝 (𝑘) = 0; 𝛼𝑝 (𝑘) = 0 (29) the relative pose error of the follower robot with respect to the moving
3: Computing the angle between the 𝑧 axis of the frame 𝑭 𝑠𝑒 and the
reference frame. The rest of the controller is a feedforward term which
𝑧 axis of reference frame; is based on the desired relative manufacturing trajectory, and its main
function is to improve the response speed of the system.
( )
𝑅(∶, 3) ⋅ 𝑒⃖⃖⃖𝑧⃗
𝜙(𝑘) = arccos (30) 3.6. Stability analysis
|𝑅(∶, 3)| ||𝑒⃖⃖⃖𝑧⃗||
where 𝑅(∶, 3) is the third column of the rotation matrix 𝑅𝑟 (𝑘), The leader robot is controlled by the PD control algorithm, which
𝑅𝑟 (𝑘) = 𝑅𝑜𝑡𝑧 (𝛾𝑟 (𝑘))𝑅𝑜𝑡𝑦 (𝛽𝑟 (𝑘)𝑅𝑜𝑡𝑥 (𝛼𝑟 (𝑘)) and 𝑒⃖⃖⃖𝑧⃗ is unit vector of 𝑧 has been widely verified. Therefore, we only analyzes the stability of
axis of the reference frame. follower robot controller in this part.
4: Initializing the initial pose of the parallel robot end-effector with
Combining Eqs. (34)–(37), we have
respect to its base frame 𝑭 𝑝𝑏 , and constraints 𝛽𝑚𝑎𝑥 , 𝜙min , 𝑧min and
𝑧max 𝒆𝑟 (𝑘 + 1) − 𝒆𝑟 (𝑘)
( )
= 𝒑𝑑𝑟,𝑠 (𝑘 + 1) − 𝒑𝑟,𝑠 (𝑘 + 1) − 𝒑𝑑𝑟,𝑠 (𝑘) − 𝒑𝑟,𝑠 (𝑘)
𝑥𝑝 (0) = 0; 𝑦𝑝 (0) = 0; 𝑧𝑝 (0) = 𝑧0 ; = 𝒑𝑑𝑟,𝑠 (𝑘 + 1) − 𝒑𝑑𝑟,𝑠 (𝑘) − 𝑱 𝑤 𝒖𝑠 (𝑘) (38)
(31) ( )
𝛽𝑝 (0) = 0 𝛾𝑝 (0) = 0; 𝛼𝑝 (0) = 0 = −𝑱 𝑤 𝒌𝑠1 𝒔(𝑘) + 𝒌𝑠2 𝒆𝑟 (𝑘)
(( ) )
5: Planning the trajectory of the parallel robot end-effector along 𝛽 = −𝑱 𝑤 𝒌𝑠1 𝑱 −1 −1
𝑤 𝝀 + 𝑱 𝑤 + 𝒌𝑠2 𝒆𝑟 (𝑘) − 𝒆𝑟 (𝑘 − 1)
and 𝑧 directions using Eqs. (20) – (25). Then, Eq. (38) can be rewritten as

𝒆𝑟 (𝑘 + 1) = 𝝁1 + 𝝁2 𝒆𝑟 (𝑘) (39)
3.5. Photogrammetry-based pose tracking control in the follower module
where
To facilitate the subsequent derivation and stability analysis, an
novel work Jacobian matrix is defined as 𝝁1 = −𝒆𝑟 (𝑘 − 1) (40)
[ ][ ]
𝑰3 𝑹𝑟𝑒𝑓 (( )
𝑱𝑤 = ( )−1 𝑠𝑒 𝑱 𝑔,𝑠 (32)
𝑻 𝑟𝑎 (𝑘) 𝑟𝑒𝑓
𝑹𝑠𝑒 𝝁2 = −𝑱 𝑤 𝒌𝑠1 𝑱 −1 −1
𝑤 𝝀 + 2𝑱 𝑤 + 𝒌𝑠2 (41)

where 𝑰 3 ∈ R3×3
is an identity matrix, 𝑹𝑟𝑒𝑓
is the rotation matrix of
𝑠𝑒
Combining Eqs. (36) and (38), we have
𝑭 𝑠𝑒 with respect to the frame 𝑭 𝑟𝑒𝑓 at time instant 𝑘, and the 𝑹𝑟𝑒𝑓
𝑠𝑒 is 𝒔(𝑘 + 1) = 𝝁3 + 𝝁4 𝒆𝑟 (𝑘) (42)
computed using the method mentioned in Section 3.2, 𝑻 𝑟𝑎 (𝑘) is denoted
as where
⎡ 0 −𝑆𝑖𝑛(𝛾(𝑘)) 𝐶𝑜𝑠(𝛾(𝑘))𝐶𝑜𝑠(𝛽(𝑘)) ⎤ ( )
𝝁3 = 𝑱 −1
𝑤 𝝀 + 𝑰 6 𝝁1 (43)
𝑻 𝑟𝑎 (𝑘) = ⎢ 0 𝐶𝑜𝑠(𝛾(𝑘)) 𝑆𝑖𝑛(𝛾(𝑘))𝐶𝑜𝑠(𝛽(𝑘)) ⎥ (33)
⎢ ⎥
⎣ 1 0 −𝑆𝑖𝑛(𝛾(𝑘)) ⎦
(( ) )
Combining Eqs. (1), (32) and (33), a novel motion model of serial 𝝁4 = 𝑱 −1
𝑤 𝝀 + 𝑰 6 𝝁2 − 𝑰 6 (44)
manipulator can be deduced as Assuming that at time instant 𝑘 − 1, the 𝒆𝑟 (𝑘 − 1) is bounded, and
𝒑̇ 𝑠 = 𝑱 𝑤 𝒖𝑠 (34) the parameters 𝝁1 and 𝝁3 are also bounded. 𝝁2 and 𝝁4 are related to
control parameters 𝒌𝑠1 , 𝒌𝑠2 , and 𝝀. The selected the Lyapunov function
where 𝒑̇ 𝑠 is the time derivative of pose of end-effector of follower robot in this paper is
with respect to the reference frame. The tracking error can be defined
as 𝑽 (𝑘) = 𝑽 1 (𝑘) + 𝑽 2 (𝑘) (45)
𝒆𝑟 (𝑘) = 𝒑𝑑𝑟,𝑠 (𝑘) − 𝒑𝑟,𝑠 (𝑘) (35) where
where 𝒑𝑟,𝑠 (𝑘) is the actual relative manufacturing trajectory provided 𝑽 1 (𝑘) = 𝒆𝑟 (𝑘)𝑇 𝒆𝑟 (𝑘) (46)
by Section 3.2. With Assumption 2 held, an error signal in joint space
is introduced, which is given by
( ) 𝑽 2 (𝑘) = 𝒔(𝑘)𝑇 𝒔(𝑘) (47)
𝒔(𝑘) = 𝑱 −1
𝑤 𝝀𝒆𝑟 (𝑘) + 𝒆𝑟 (𝑘) − 𝒆𝑟 (𝑘 − 1) (36)

7
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Fig. 5. The Simscape model of the AFPS.

is satisfied, an inequality 𝛥𝑽 1 (𝑘 + 1) < 0 holds.

2𝝁𝑇1 𝝁1 < 𝒆𝑇𝑟 (𝑘)𝑴𝒆𝑟 (𝑘) < ‖𝑴‖ ‖𝒆𝑟 (𝑘)‖2 (49)

where 𝑴 = 𝑰 6 − 2𝝁𝑇2 𝝁2 .
Similarly, the term 2𝝁𝑇4 𝝁4 − 𝑰 6 can also be negative definite by
selecting the control parameters 𝒌𝑠1 , 𝒌𝑠2 , and 𝝀. Besides, an inequality
𝛥𝑽 2 (𝑘 + 1) < 0 holds, under following conditions

2𝝁𝑇3 𝝁3 < 𝒆𝑇𝑟 (𝑘)𝑵𝒆𝑟 (𝑘) < ‖𝑵‖ ‖𝒆𝑟 (𝑘)‖2 (50)

where 𝑵 = 𝑰 6 − 2𝝁𝑇4 𝝁4 .
Based on the above analysis, choosing suitable control parameters
𝒌𝑠1 , 𝒌𝑠2 , and 𝝀 can make matrices 𝑀 and 𝑁 be negative, and the
inequalities (51) are satisfied.
{
2𝝁𝑇1 𝝁1 < 𝒆𝑇𝑟 (𝑘)𝑴𝒆𝑟 (𝑘) < ‖𝑴‖ ‖𝒆𝑟 (𝑘)‖2
(51)
2𝝁𝑇3 𝝁3 < 𝒆𝑇𝑟 (𝑘)𝑵𝒆𝑟 (𝑘) < ‖𝑵‖ ‖𝒆𝑟 (𝑘)‖2

Therefore, the inequality 𝛥𝑽 (𝑘 + 1) < 0 is held, which means 𝒔(𝑘)


and 𝒆𝑟 (𝑘) can be convergent to zero.
Fig. 6. The partial desired trajectories of the rotary stage and leader robot generated
by the OTP algorithm in simulation. In practice, the control parameters 𝒌𝑠1 and 𝒌𝑠2 are selected by trial
and error method. Then, we run the follower robot i.e. serial robot
under the PBPT controller by tracking the planned desired trajectories
and record the error signals 𝑒𝑟 (𝑘) and 𝑠(𝑘) whose norm values determine
the bounds of 𝝁1 and 𝝁3 . With the above information, the parameter 𝝀
Substituting Eq. (39) into Eq. (46), the derivate of 𝑽 1 (𝑘) with is chosen to make both matrices 2𝝁𝑇2 𝝁2 − 𝑰 6 and 2𝝁𝑇4 𝝁4 − 𝑰 6 negative
respect to time is computed as definite. It is worth pointing out that 𝝀 is not unique and is kept as
unchanged as possible while satisfying the two inequalities.
𝛥𝑽 1 (𝑘 + 1) = 𝑽 1 (𝑘 + 1) − 𝑽 1 (𝑘)
= 𝒆𝑟 (𝑘 + 1)𝑇 𝒆𝑟 (𝑘 + 1) − 𝒆𝑟 (𝑘)𝑇 𝒆𝑟 (𝑘)
( )𝑇 ( ) 4. Simulation and experimental results
= 𝝁1 + 𝝁2 𝒆𝑟 (𝑘) 𝝁1 + 𝝁2 𝒆𝑟 (𝑘) − 𝒆𝑟 (𝑘)𝑇 𝒆𝑟 (𝑘) (48)
( )𝑇 ( )
≤ 2𝝁𝑇1 𝝁1 + 2 𝝁2 𝒆𝑟 (𝑘) 𝝁2 𝒆𝑟 (𝑘) − 𝒆𝑠 (𝑘)𝑇 𝒆𝑟 (𝑘) The comparison results of three approaches are presented in this
( )
= 2𝝁𝑇1 𝝁1 + 𝒆𝑟 (𝑘)𝑇 2𝝁𝑇2 𝝁2 − 𝑰 6 𝒆𝑟 (𝑘) section. The first comparison algorithm is Non-Coordination (Non-C)
approach. In Non-C approach, there is no connection between the two
where the term 2𝝁𝑇2 𝝁2 − 𝑰 6 can be negative definite by selecting the robots, and two robots track their desired trajectory with the help of
control parameters 𝒌𝑠1 , 𝒌𝑠2 , and 𝝀. Therefore, if the following inequality two different PD controllers, separately. The controller of the parallel

8
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Fig. 7. Desired and actual relative manufacturing trajectories in simulation using the proposed approach.

Fig. 8. The data communication structure of the AFPS.

robot is the same as that in the proposed approach. The controller of 4.1. Simulations of fiber placement process with semicircle relative manu-
the serial manipulator is designed as follows facturing trajectory
(
𝒖𝑠2 = 𝑱 −1
𝑟𝑜𝑏 𝒌𝑠3 𝒆𝑠 (𝑘) + 𝒌𝑠4 (𝒆𝑠 (𝑘) − 𝒆(𝑘 − 1)) (52)
To verify the effectiveness and calibration sensitivity of the pro-
where 𝒆𝑠 (𝑘) = 𝒑𝑑,𝑠 (𝑘) − 𝒑𝑠 (𝑘) is the pose error, 𝒑𝑑,𝑠 (𝑘) is the de- posed algorithm, the comparison simulations in Matlab are designed.
sired trajectory of the serial manipulator with respect to 𝑭 𝑠𝑏 . The The system model shown in Fig. 5 is established using Simulink/
desired trajectory 𝒑𝑑,𝑠 (𝑘) is calculated using the theoretical data of Simscape. In this model, the actuation signals of all revolute joints are
relative manufacturing trajectory and coordinate frame relationship of the motion commands provided by input. Any two above-mentioned
the AFPS. The second comparison approach is Semi-offline trajectory coordinate frames have a direct or indirect pose relationship. For
synchronized approach (SOTS) (Zhang et al., 2018). Since the trajectory example, the pose of 𝑭 𝑠𝑏 with respect to 𝑭 𝑝𝑠 is known, and the pose of
used for the simulations and experiments does not suffer from singular- 𝑭 𝑠𝑒 with respect to 𝑭 𝑣 can be calculated using Eq. (53)
ity problems, it is compensation pose calculation part does not work, 𝑝𝑠 𝑠𝑏
𝑯 𝑝𝑠
𝑠𝑒 = 𝑯 𝑠𝑏 𝑯 𝑠𝑒 (53)
that is, its desired trajectories are the same as that of the first approach.
The simulations of fiber placement process with closed semicir- where 𝑯 𝑠𝑏𝑠𝑒 can be calculated using forward kinematics. Therefore, a
cle relative manufacturing trajectory are presented in Section 4.1. In user-defined Matlab function based on the principle of homogeneous
Section 4.2, the experimental results of fiber placement process with transformation is designed to simulate the function of the photogram-
Y-shape mandrel are presented. In the simulations and experiments, metry sensor. In order to simulate the inaccurate calibration of AFPS,
the fiducial trajectory of the AFPS is relative manufacturing trajectory. the actual pose of the serial manipulator base frame with respect to the
The pose tracking errors of the parallel robot are employed to evalu- parallel robot base frame𝑭 𝑝𝑏 is different from the theoretical values.
ate the motion accuracy of the leader robot, while the relative pose The pose errors along 𝑥 and 𝑦 direction with respect to the parallel
tracking errors are used to evaluate the accuracy of the coordination robot base frame 𝑭 𝑝𝑏 are 𝛥𝑥 = 0.02 m and 𝛥𝑦 = 0.02 m. This pose error
(synchronization) movement of the system. can be regarded as calibration error. In practice, the pose calibration

9
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Table 2
The steady errors of the relative pose of the serial manipulator with
respect to the reference frame in simulation.
Direction PBLF Non-C SOTS
𝑥 [mm] 0.0673 17.516 −8.1370
𝑦 [mm] −0.0397 18.352 −6.2445
𝑧 [mm] 0.0435 0.4610 −0.1394
𝛾 [deg] 0.0239 0.0250 0.03410
𝛽 [deg] −0.0245 −0.0261 −0.0546
𝛼 [deg] −0.0312 0.0293 0.02197

the control parameters are set as


𝒌𝑝1 = 𝑑𝑖𝑎𝑔(0.35, 0.35, 0.35, 0.35, 0.35, 0.35)
𝒌𝑝2 = 𝑑𝑖𝑎𝑔(0.07, 0.07, 0.07, 0.07, 0.07, 0.07)
𝒌𝑠1 = 𝑑𝑖𝑎𝑔(0.23, 0.23, 0.23, 0.23, 0.23, 0.23)
𝒌𝑠2 = 𝑑𝑖𝑎𝑔(0.05, 0.05, 0.05, 0.05, 0.05, 0.05)
𝝀 = 𝑑𝑖𝑎𝑔(0.10, 0.10, 0.10, 0.10, 0.10, 0.10)
In the Non-Coordination approach, the control parameters of the leader
robot are set as the same as those in the proposed approach, while the
control parameters of follower robot are set as
Fig. 9. The relative manufacturing trajectory of Y-shape mandrel. 𝒌𝑠3 = 𝑑𝑖𝑎𝑔(0.30, 0.30, 0.30, 0.30, 0.30, 0.30)
𝒌𝑠4 = 𝑑𝑖𝑎𝑔(0.05, 0.05, 0.05, 0.05, 0.05, 0.05)
In the SOTS approach, its parameters are consistent with the Zhang
et al. (2018).
Fig. 6 shows the desired trajectory of the follower robot generated
by the OTP algorithm. The trajectory is smooth and will not cause
the parallel robot vibration during the process. Besides, the motion
of the rotary stage and the parallel robot solve the occlusion problem
caused by mandrel while completing straight line parts. Therefore, the
proposed algorithm is suitable for the fiber placement process.
Fig. 7 shows the desired relative manufacturing trajectory and the
actual one in the proposed approach. It can be observed that the two
trajectories are almost identical. Consequently, the follower robot can
effectively track the relative manufacturing trajectory with respect to
the moving reference frame 𝑭 𝑟𝑒𝑓 . The desired trajectory and control
method for the parallel robot in two approaches are set as the same
in this simulation. So only the coordination of the two robots needs
to be analyzed. Table 2 shows the steady error of the relative tracking
errors of the follower robot with respect to the reference frame 𝑭 𝑟𝑒𝑓 .
The relative position error of follower robot along 𝑥 and 𝑦 directions
Fig. 10. The partial desired trajectories of the rotary stage and leader robot generated
in the proposed approach is smaller than that in the Non-C and SOTS
by OTP algorithm in experiment. approaches, while the relative rotation errors of the follower robot in
the three approaches are the same, The reason for this phenomenon is
that actual values of the relative pose between the robots are different
errors will not exceed the above values. The purpose of setting such from the theoretical values. Besides, in order to guarantee the accuracy
large errors is to verify whether the proposed approach has strong of parallel robots, the SOTS algorithm makes compromises in terms
robustness to calibration error. of the system coordination. During the fiber placement process, if the
Straight lines and circles are common trajectories in fiber placement system calibration is not accurate, the above-mentioned phenomenon
process. The candidate relative manufacturing trajectory in simulation will occur. The results of the simulations demonstrate that the proposed
is closed semicircle. More precisely, the trajectory is composed of approach does not rely on precise calibration of the system.
straight line and the half circle, and the starting point of the fiber
placement process is at the intersection of the straight line and semi- 4.2. Experiments of fiber placement process with a Y-shape mandrel
circle. If a mandrel is inside the trajectory and cannot be adjusted
during the process, the mandrel will cause occlusion. Therefore, the To illustrate the actual performance of the proposed approach for
trajectory is suitable for testing the OPT algorithm and the two ap-
the AFPS, a comparison experiments of manufacturing composite struc-
proaches. For safety considerations, the trajectories used in simulations
ture with Y-shape is presented in this subsection. There are several
and experiments are planned under the constraints on velocity and
differences between the actual experiments and simulations: (1) The
acceleration.
fiber placement head used in the experiments is different from the
The parameters in OTP algorithm are set as
one used in the simulations; (2) In the experiments, the pose of the
𝑘𝑡1 = 0.020; 𝑘𝑡2 = 0.015; 𝑘𝑡3 = 0.032; 𝛽max = 0.26 end-effectors of robots can be obtained through the measurement in-
𝜙min = 0.087, 𝑧min = 0.23, 𝑧max = 0.28 formation from the C-track; (3) The two robots are controlled by
The control parameters of the proposed approach and Non-coordi- different computers, which can transmit data in real time. The data
nation approach are tuned by trial and error. In the proposed approach, communication structure of the system is shown in Fig. 8.

10
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Fig. 11. The status of the system passing through the special points.

In order to simulate the fiber placement process of structure with


complex shape, we selected a Y shape mandrel and a relative man-
ufacturing trajectory composed of linear and curved trajectories in
experiments. As shown in Fig. 9, the relative manufacturing trajectory
for the Y-shape mandrel consists of three segments. In the first segment
(from 𝐴 to 𝐵), the fiber head is perpendicular to Y-shape surface, and
moves in a constant spiral along the 𝑦-axis of 𝑭 𝑟𝑒𝑓 . The paths in the
second segment (from 𝐵 to 𝐶) and third segment (from 𝐶 to 𝐷) are
curve and line, respectively. The parameters of the OTP algorithm are
the same as those in simulation, and the desired trajectory of leader
robot is shown in Fig. 10.
30 ms is the fastest sampling time of the C-Track 780. A lot of
experimental tests show that 20 ms-30 ms is the proper control interval
that can deliver satisfactory control performance (Shu et al., 2018;
Zhang et al., 2018). Therefore, the sampling interval is set as 30 ms.
In expriments, the control parameters in the proposed approach are set
as
𝒌𝑝1 = 𝑑𝑖𝑎𝑔(0.40, 0.40, 0.40, 0.40, 0.40, 0.40)
𝒌𝑝2 = 𝑑𝑖𝑎𝑔(0.06, 0.06, 0.06, 0.06, 0.06, 0.06)
𝒌𝑠1 = 𝑑𝑖𝑎𝑔(0.37, 0.37, 0.37, 0.37, 0.37, 0.37)
𝒌𝑠2 = 𝑑𝑖𝑎𝑔(0.05, 0.05, 0.05, 0.05, 0.05, 0.05) Fig. 12. The position errors of the leader robot in experiment.
𝜆 = 𝑑𝑖𝑎𝑔(0.12, 0.12, 0.12, 0.12, 0.12, 0.12)
In the Non-Coordination approach, the control parameters of the par-
allel robot are the same as those of the proposed approach, and the between the PBLF and the Non-C approach. While the results of the
control parameters of the serial manipulator are set as SOTS approach are a little worse compared the results of the other two
approach. Figs. 14 and 15 show that the PBLF approach provides the
𝒌𝑠3 = 𝑑𝑖𝑎𝑔(0.37, 0.37, 0.37, 0.37, 0.37, 0.37)
best coordination effect for the system, followed by the SOTS approach,
𝒌𝑠4 = 𝑑𝑖𝑎𝑔(0.06, 0.06, 0.06, 0.060, 0.06, 0.06)
while the Non-C approach exhibits the worst results. An interesting
In the SOTS approach, its parameters are consistent with the Zhang phenomenon which also appears in the experimental results is that the
et al. (2018). SOTS approach compromises the accuracy of the parallel robot for the
In the proposed approach, the status of AFPS passing through A, purpose of enhancing the coordination of the system. With the help of
B, C and D is shown in Fig. 11. The tracking errors of the leader the proposed approach, the relative errors can be maintained within
robot and the relative errors of the follower robot with respect to the 1.8 mm, that is, two robots can keep the coordination during the fiber
moving reference frame 𝑭 𝑟𝑒𝑓 are described in Figs. 12–15, respectively. placement process.
Similar to the simulation results, Figs. 12 and 13 demonstrate that there The mean absolute error (MAE) and root mean square error (RMSE)
is no discernible difference in the tracking errors of the leader robot of the leader robot and the relative pose of follower robot with respect

11
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Table 3
The MAE and RMSE of the leader robot in experiment.
Direction Characteristics PBLF Non-C SOTS
MAE (mm) 0.5290 0.5239 0.7347
𝑥
RMSE (mm) 0.3758 0.4167 0.5451
MAE (mm) 0.5348 0.7578 0.5973
𝑦
RMSE (mm) 0.4002 0.5560 0.4340
MAE (mm) 0.8319 0.6947 0.9554
𝑧
RMSE (mm) 0.5970 0.5409 0.5231
MAE (deg) 0.3825 0.2178 0.3345
𝛾
RMSE (deg) 0.2821 0.1813 0.3241
MAE (deg) 0.3390 0.3296 0.4023
𝛽
RMSE (deg) 0.2769 0.2521 0.3826
MAE (deg) 0.3687 0.3025 0.4016
𝛼
RMSE (deg) 0.2623 0.2427 0.2836

Table 4
The MAE and RMSE of the relative pose of follower robot with respect to the reference
frame in experiment.
Direction Characteristics PBLF Non-C SOTS
MAE (mm) 0.3692 0.7710 0.6173
Fig. 13. The rotation errors of the leader robot in experiment. 𝑥
RMSE (mm) 0.3121 0.5760 0.4522
MAE (mm) 0.2364 0.5954 0.2714
𝑦
RMSE (mm) 0.1675 0.4653 0.2832
MAE (mm) 0.2419 0.9319 0.3654
𝑧
RMSE (mm) 0.1911 0.7291 0.2367
MAE (deg) 0.1487 0.4251 0.2864
𝛾
RMSE (deg) 0.1096 0.3581 0.2131
MAE (deg) 0.2149 0.2336 0.1843
𝛽
RMSE (deg) 0.1682 0.1633 0.1155
MAE (deg) 0.1493 0.4423 0.2761
𝛼
RMSE (deg) 0.1132 0.3092 0.1358

to the reference frame 𝑭 𝑟𝑒𝑓 are shown in Tables 3 and 4. Obviously,


RMSE and MAE of relative pose in proposed approach are also smaller
than those in the Non-C and the SOTS approaches. Using the data in
Table 4, we find that, the PBLF approach can reduce the average MAEs
of the relative pose in position and orientation direction by 62.15%
and 46.42%, respectively. While the corresponding average RMSEs are
reduced by 61.20% and 45.26%, respectively. In fact, the accuracy and
coordination of the system mainly depend on control effect of the fol-
lower robot. In the PBLF approach, selecting an appropriate robot as the
Fig. 14. The relative position errors of the follower robot in experiment. follower robot can reduce the performance requirements of the leader
robot. The experimental results demonstrate that the PBLF approach
can improve the coordination of the AFPS. Although the experiments do
not use a real AFP head, it can be inferred from the experimental results
that the PBLF approach can improve the accuracy of fiber placement
process by reducing the tracking errors of the relative manufacturing
trajectory.

5. Conclusion

In this paper, a photogrammetry-based leader–follower approach


is proposed for an AFPS which consists of two robots with different
structures. An OPT algorithm is designed in this approach to generate
desired trajectories of the leader robot and rotary stage. With the help
of the OPT algorithm, the mandrel is in a manufacturing-friendly pose.
Two different robot controllers are designed to achieve the coordina-
tion of the two robots. One is PD controller which is used to make the
leader robot track its desired trajectory. Another is photogrammetry-
based pose tracking controller which uses the feedback signal of the
photogrammetry sensor to track its relative manufacturing trajectory
with respect to moving reference frame. The simulation result demon-
strate that the proposed approach has strong robustness to pose cali-
Fig. 15. The relative rotation errors of the follower robot in experiment. bration error among the robots. The experiment results indicate that

12
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Table A.5 Carnevale, G., Camisa, A., & Notarstefano, G. (2023). Distributed online aggregative
A list of crucial or frequently appeared notions in the article. optimization for dynamic multirobot coordination. IEEE Transactions on Automatic
Notions Descriptions Control, 68(6), 3736–3743.
Cong, V. D., & Hanh, L. D. (2023). A review and performance comparison of visual
𝑭 𝑝𝑏 , 𝑭 𝑠𝑏 Base frame of parallel robot and serial manipulator
servoing controls. International Journal of Intelligent Robotics and Applications, 7(1),
𝑭 𝑝𝑒 , 𝑭 𝑠𝑒 End-effector frame of parallel robot and serial manipulator
65–90.
𝑭 𝑝𝑠 , 𝑭 𝑟𝑒𝑓 Photogrammetry sensor frame and reference frame
Corke, P. (2017). Springer tracts in advanced robotics: vol. 118, Robotics, vision and control
𝑭 𝑜 , 𝑭 𝑐𝑗 Object frame and 𝑗th camera frame
- fundamental algorithms in MATLAB. Cham: Springer International Publishing.
𝑯 𝑝𝑠
𝑝𝑏
Homogeneous transformation matrix from 𝑭 𝑝𝑏 to 𝑭 𝑝𝑠 Devlieg, R. (2010). Expanding the use of robotics in airframe assembly via accurate
𝒗𝑠 End-effector velocity of serial manipulator robot technology. SAE International Journal of Aerospace, 3(1), 198–203.
𝒑𝑝 , 𝒑𝑠 End-effector poses of parallel robot and serial manipulator Di Lillo, P., Pierri, F., Antonelli, G., Caccavale, F., & Ollero, A. (2021). A framework
𝒑𝑟,𝑠 End-effector poses of serial manipulator with respect to 𝑭 𝑟𝑒𝑓 for set-based kinematic control of multi-robot systems. Control Engineering Practice,
𝜃 Position of rotary stage 106, Article 104669.
𝜃𝑓 Filtered position of rotary stage Dong Sun, & Mills, J. (2002). Adaptive synchronized control for coordination of
𝒒𝑝 , 𝒒𝑠 Joint vectors of parallel robot and serial manipulator multirobot assembly tasks. IEEE Transactions on Robotics and Automation, 18(4),
𝑱 𝑔,𝑠 Geometric Jacobian matrix of serial manipulator 498–510.
𝐼𝑛𝑣𝑘(∙) Inverse kinematics function of parallel robot Feng, Z., Hu, G., Sun, Y., & Soon, J. (2020). An overview of collaborative robotic
𝒑𝑑𝑟,𝑠 Desired relative manufacturing trajectory manipulation in multi-robot systems. Annual Reviews in Control.
𝒑𝑑,𝑝 Desired trajectory of parallel robot Filion, A., Joubair, A., Tahan, A. S., & Bonev, I. A. (2018). Robot calibration using a
𝒑𝑑,𝑠 Desired trajectory of serial manipulator portable photogrammetry system. Robotics and Computer-Integrated Manufacturing,
𝜃𝑑 Desired trajectory of rotary stage 49, 77–87.
𝒆𝑝 Pose errors of parallel robot (leader robot) Hély, C., Birglen, L., & Xie, W.-F. (2017). Feasibility study of robotic fibre placement
𝒆𝑠 Pose errors of serial manipulator (follower robot) on intersecting multi-axial revolution surfaces. Robotics and Computer-Integrated
𝒆𝑟 Relative pose errors of serial manipulator with respect to 𝑭 𝑟𝑒𝑓 Manufacturing, 48, 73–79.
𝑱𝑤 Work Jacobian matrix Huang, Y., Zheng, Y., Wang, N., Ota, J., & Zhang, X. (2020). Peg-in-hole assembly based
𝒔(𝑘) Auxiliary error signal of serial manipulator in joint space on master-slave coordination for a compliant dual-arm robot. Assembly Automation,
𝝀 Gain matrix in auxiliary error signal 𝑺(𝑘) 40(2), 189–198.
𝒌𝑠1 , 𝒌𝑠2 Control parameters of the PBPT algorithm Janabi-Sharifi, F., & Marey, M. (2010). A Kalman-filter-based method for pose
𝒌𝑝1 , 𝒌𝑝2 Control parameters of the PD controller estimation in visual servoing. IEEE Transactions on Robotics, 26(5), 939–947.
𝒌𝑠3 , 𝒌𝑠3 Control parameters of the Non-C algorithm Jeffries, K. A. (2013). Enhanced robotic automated fiber placement with accurate
𝑽 (𝑘) Lyapunov function robot technology and modular fiber placement head. SAE International Journal of
AFPS Automated fiber placement system Aerospace, 6(2), 774–779.
PBLF Photogrammetry-based leader–follower Jia, B., Liu, S., & Liu, Y. (2015). Visual trajectory tracking of industrial manipulator with
PBPT Photogrammetry-based pose tracking iterative learning control. Industrial Robot: An International Journal, 42(1), 54–63.
OPT Offline trajectory planning Lauer, A. P. R., Lerke, O., Blagojevic, B., Schwieger, V., & Sawodny, O. (2023). Tool
DOF Degree of freedom center point control of a large-scale manipulator using absolute position feedback.
MAE Mean absolute error Control Engineering Practice, 131, Article 105388.
RMSE Root mean square error Li, R., Ding, N., Zhao, Y., & Liu, H. (2023). Real-time trajectory position error
compensation technology of industrial robot. Measurement, 208, Article 112418.
Li, B., Li, Y., Tian, W., & Liao, W. (2023). Pose accuracy improvement in robotic
machining by visually-guided method and experimental investigation. Robotics and
the proposed approach improves the coordination of the two robots. Autonomous Systems, 164, Article 104416.
Li, B., Zhang, W., Li, Y., Tian, W., & Wang, C. (2022). Positional accuracy improvement
Specifically, compared with the Non-C and the SOTS approach, the of an industrial robot using feedforward compensation and feedback control.
PBLF approach significantly reduce the average MAEs and RMSEs of the Journal of Dynamic Systems, Measurement, and Control, 144(7), Article 071003.
relative pose. The use of PBLF approach makes the trajectory planning Liang, C.-D., Ge, M.-F., Liu, Z.-W., Ling, G., & Liu, F. (2021). Predefined-time formation
of the AFPS more convenient, and it also provides a necessary basis tracking control of networked marine surface vehicles. Control Engineering Practice,
107, Article 104682.
for the AFPS to manufacture high-precision composite structure. The
Luh, J. Y. S., & Zheng, Y. F. (1987). Constrained relations between two coordinated
future work includes the application of the proposed control strategy to industrial robots for motion control. The International journal of robotics research,
AFPS with operational fiber placement head to manufacture composite 6(3), 60–70.
structures. Meng, Z., Dimarogonas, D. V., & Johansson, K. H. (2014). Leader–follower coordinated
tracking of multiple heterogeneous Lagrange systems using continuous control. IEEE
Transactions on Robotics, 30(3), 739–745.
Declaration of competing interest Oromiehie, E., Prusty, B. G., Compston, P., & Rajan, G. (2019). Automated fibre place-
ment based composite structures: Review on the defects, impacts and inspections
The authors declare the following financial interests/personal rela- techniques. Composite Structures, 224, Article 110987.
Shen, H., Pan, Y.-J., Ahmad, U., & He, B. (2021). Pose synchronization of multiple
tionships which may be considered as potential competing interests:
networked manipulators using nonsingular terminal sliding mode control. IEEE
Dr. Wen-Fang Xie reports financial support was provided by Natural Transactions on Systems, Man, and Cybernetics: Systems, 51(12), 7497–7509.
Sciences and Engineering Research Council of Canada. Shu, T., Gharaaty, S., Xie, W., Joubair, A., & Bonev, I. A. (2018). Dynamic path tracking
of industrial robots with high accuracy using photogrammetry sensor. IEEE/ASME
Transactions on Mechatronics, 23(3), 1159–1170.
Appendix
Siciliano, B., Sciavicco, L., Villani, L., & Oriolo, G. (2009). Robotics - modelling,
planning and control. In M. J. Grimble, & M. A. Johnson (Eds.), Advanced textbooks
The crucial or frequently appeared notions are presented in Ta- in control and signal processing, London: Springer London.
ble A.5. Stadelmann, L., Sandy, T., Thoma, A., & Buchli, J. (2019). End-effector pose correction
for versatile large-scale multi-robotic systems. IEEE Robotics and Automation Letters,
4(2), 546–553.
References Sun, Y., Dong, D., Qin, H., & Wang, W. (2020). Distributed tracking control for multiple
Euler–Lagrange systems with communication delays and input saturation.
Aized, T., & Shirinzadeh, B. (2011). Robotic fiber placement process analysis and Tang, J., Zhou, T., Zakeri, E., Shu, T., & Xie, W.-F. (2023). Photogrammetry-based
optimization using response surface method. International Journal of Advanced dynamic path tracking of industrial robots using adaptive neuro-PID control method
Manufacturing Technology, 55(1–4), 393–404. and robust Kalman Filter. In 2023 9th international conference on automation, robotics
Bai, Q., Li, P., Tian, W., Shen, J., Li, B., & Hu, J. (2023). Vision guided dynamic and applications (pp. 91–96). Abu Dhabi, United Arab Emirates: IEEE.
synchronous path tracking control of dual manipulator cooperative system. Journal Wu, K., & Kuhlenkoetter, B. (2022). Dynamic behavior and path accuracy of an
of Manufacturing Science and Engineering, 145(12), Article 121003. industrial robot with a CNC controller. Advances in Mechanical Engineering, 14(2),
Bilal, D. K., Unel, M., Tunc, L. T., & Gonul, B. (2022). Development of a vision Article 168781322210828.
based pose estimation system for robotic machining and improving its accuracy Xiao, H., & Philip Chen, C. (2021). Time-varying nonholonomic robot consensus for-
using LSTM neural networks and sparse regression. Robotics and Computer-Integrated mation using model predictive based protocol with switching topology. Information
Manufacturing, 74, Article 102262. Sciences, 567, 201–215.

13
R. Zhang et al. Control Engineering Practice 141 (2023) 105691

Xiao, H., Yu, D., & Chen, C. P. (2022). Self-triggered-organized mecanum-wheeled Zhang, W., Liu, F., Jiang, T., Yi, M., Chen, W., & Ding, X. (2022). Overview of
robots consensus system using model predictive based protocol. Information sciences, current design and analysis of potential theories for automated fibre placement
590, 45–59. mechanisms. Chinese Journal of Aeronautics, 35(4), 1–13.
Xie, W. (2016). Design and analysis of collaborative automated fiber placement Zhang, X., Xie, W., & Hoa, S. V. (2018). Semi-offline trajectory synchronized
machine. International Journal of Advanced Robotics and Automation, 1(1), 01–14. algorithm of the cooperative automated fiber placement system. Robotics and
Yang, J., Wang, D., Fan, B., Dong, D., & Zhou, W. (2017). Online absolute pose Computer-Integrated Manufacturing, 51, 53–62.
compensation and steering control of industrial robot based on six degrees of
freedom laser measurement. Optimization and Engineering, 56(3), Article 034111.
Yuan, J.-C. (1989). A general photogrammetric method for determining object position
and orientation. IEEE Transactions on Robotics and Automation, 5(2), 129–142.

14

You might also like