Download as pdf or txt
Download as pdf or txt
You are on page 1of 11

Robotics and Computer–Integrated Manufacturing 71 (2021) 102165

Contents lists available at ScienceDirect

Robotics and Computer-Integrated Manufacturing


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

A novel kinematic parameters calibration method for industrial robot based


on Levenberg-Marquardt and Differential Evolution hybrid algorithm
Guoyue Luo, Lai Zou *, Ziling Wang, Chong Lv, Jing Ou, Yun Huang
State Key Laboratory of Mechanical Transmission, Chongqing University, Chongqing 400044, China

A R T I C L E I N F O A B S T R A C T

Keywords: The poor absolute positioning accuracy of industrial robots is the main obstacle for its further application in
Industrial robot precision grinding of complex surfaces, such as blisk, blade, etc. Based on the established kinematic error model
Kinematic parameters calibration of a typical industrial robot FANUC M710ic/50, a novel kinematic parameters calibration method is proposed in
Positioning accuracy
this paper to improve the absolute positioning accuracy of robot. The pre-identification of the kinematic
Levenberg-Marquardt
Differential Evolution
parameter deviations of robot was achieved by using the Levenberg-Marquardt algorithm. Subsequently, these
identified suboptimal values of parameter deviations were defined as central values of the components of initial
individuals to complete accurate identification by using Differential Evolution algorithm. The above two steps,
which were regarded as the core of this Levenberg-Marquardt and Differential Evolution hybrid algorithm, were
used to obtain the preferable values for kinematic parameters of the robot. On this basis, the experimental in­
vestigations of kinematic parameters calibration were conducted by using a laser tracker and numerical simu­
lation method. The results revealed that the robot positioning error decreased from 0.994 mm, initial positioning
error measured by laser tracker, to 0.262 mm after calibration with this proposed hybrid algorithm. The absolute
positioning accuracy has increased by 40.86% than that of the Levenberg-Marquardt algorithm, increased by
40.31% than that of the Differential Evolution algorithm, and increased by 25.14% than that of the Simulated
Annealing algorithm. This work shows that the proposed kinematic parameters calibration method has a sig­
nificant improvement on the absolute positioning accuracy of industrial robot.

1. Introduction improve the absolute positioning accuracy of the robot.


In the past few decades, many researches have been conducted on
With the development of modern industrial technologies, the the calibration of robot kinematic parameters [3, 4]. Jiang et al. [5]
emerged approach of robotic grinding through offline programming can identified the kinematic parameters of Kawasaki RS10N robot with the
effectively improve the efficiency and quality of blade processing [1], Extend Kalman filter and particle filter algorithms, and the absolute
meeting the urgent need for automation and flexibility of precision positioning accuracy was improved after calibration with the posi­
grinding. And the requirement for the robot absolute positioning accu­ tioning error decreased from 3.1407 mm to 0.263 mm. Gao et al. [6]
racy is getting higher and higher. proposed a novel hybrid algorithm that employs BP neural network and
Due to factors such as the machining and assembly errors, structural particle swarm optimization algorithm for the kinematic parameter
deformation, the actual kinematic parameters of the robot are deviated identification, after compensating for the parameter errors, the posi­
from the nominal parameters. And the parameter errors are amplified tioning accuracy of ER20-C10 has improved to 0.4 mm. Wang et al. [7]
when they are transmitted to the end of the robot through the structure proposed a 6R robot closed-loop kinematic calibration method to
linked in series by the rotary joint, resulting in a low positioning accu­ improve the absolute positioning accuracy with point and distance
racy. To ensure the precision of robotic grinding, it is necessary to obtain constraints through machine vision. Chen et al. [8] established the
the actual structural parameters as accurate as possible through kine­ exponential product model (POE) [9], which can be well applied to
matic calibration methods [2], which have the characteristics of strong modular robots. Chen et al. [10] completed the non-kinematic calibra­
versatility and simple implementation. The way of kinematic calibration tion of industrial robots using a rigid-flexible coupling error model and a
can compensate for the parameter errors at the minimum cost and full measurement method. Wang et al. [11] put forward a universal

* Corresponding author.
E-mail address: zoulai@cqu.edu.cn (L. Zou).

https://doi.org/10.1016/j.rcim.2021.102165
Received 28 September 2020; Received in revised form 12 March 2021; Accepted 13 March 2021
Available online 21 March 2021
0736-5845/© 2021 Elsevier Ltd. All rights reserved.
G. Luo et al. Robotics and Computer-Integrated Manufacturing 71 (2021) 102165

observability index for optimal measurement configuration selection in


robot kinematic calibration experiments and proposed an improved
Particle Swarm Optimization approach to maximize the universal
observability index. Messay et al. [12] made use of a simplified version
of the original Denavit-Hartenberg kinematic model to calibrate the
kinematic parameters of Yaskawa Motoman HP20D manipulator with
the Cartesian coordinates of calibration points. The method used a novel
hybrid optimization scheme composed of the Simulated Annealing al­
gorithm (SAA) and the Trust Region algorithm. After calibration, the
positioning error was reduced to 0.2789 mm, 0.3039 mm and 0.2001
mm from 7.4553 mm, 0.77 mm and 1.7363 mm along the x, y and z,
respectively. Nguyen et al. [13] combined a robot-model-based identi­
fication approach and an artificial-neural-network-based error
compensation technique to enhance the positioning accuracy of an in­
dustrial Hyundai HH800 robot. As a result, after error compensation
using the artificial neural network, the average robot positioning accu­
racy was enhanced up to 0.3264 mm from 4.0654 mm. Cho et al. [14]
proposed a new calibration method that takes into consideration the Fig. 1. Link coordinate system.
effect of joint compliance for robot manipulators using the circular point
analysis. Gan et al. [15] proposed a calibration method for robot kine­
Table 1.
matic parameters based on the drawstring displacement sensor, and the
Nominal values of kinematic parameters.
parameter deviations were identified by the least-squares method, at
last, the average positioning error was reduced from 2.533 mm to i αi /◦ di /mm θi /◦ ai /mm
0.3575 mm. Gao et al. [16] proposed a novel kinematic model and a 1 − 90 0 θ1 150
calibration method using articulated arm coordinate machines to 2 − 180 0 θ2 -90 870
improve the positioning accuracy of industrial robot. After calibration 3 − 90 0 θ3 170
the average absolute errors reached 0.213 mm, 0.238 mm and 0.203 mm 4 90 − 1016 θ4 0
from 5.294 mm, 3.106 mm, and 2.843 mm in the direction of x, y and z, 5 − 90 0 θ5 0
respectively. Zhu et al. [17] proposed a kinematic self-calibration 6 − 180 − 175 θ6 0
method for dual-manipulators and used the Levenberg-Marquardt al­
gorithm [18-20] to identify kinematic parameter errors of robot. Tian
et al. [21] investigated the kinematic calibration technology of a the effectiveness and correctness of the proposed method, taking the
new-type 5-degree of freedom hybrid kinematic machine by considering FANUC M710ic/50 robot as the experimental object.
the ill-posed identification problem using regularization method.
Although all these methods can calibrate parameters of robot and 2. Robot kinematic error model
improve the positioning accuracy to varying degrees, from the
comprehensive perspective of calibration accuracy and efficiency, there Fig. 1 shows that each link coordinate system is established for the
is currently no recognized most effective method. It is necessary to find FANUC M710ic/50 robot. The values of nominal D-H parameters of
new method, which could obtain the higher accuracy, to better ensure robot are listed in Table 1. Based on the convention of the Modified D-H
the machining precision of robotic grinding of blade and improve the model [27, 28], known as MDH, the kinematic model of robot is
processing quality. established. The link transformation matrix of two consecutive link
The robot kinematic model is a highly nonlinear system, it is often frames can be described as following equation.
linearized by dropping second-order and above error terms and then the i
least-square algorithm [22] is used to quickly identify the kinematic ⎡ = Rot(z, θi )⋅Trans(z, di )⋅Trans(x, ai )⋅Rot(x, αi )⋅Rot(y, β⎤i )
Ti+1
cθi cβi − sαi sθi sβi − sθi cαi cθi sβi + sαi sθi cβi ai cθi
parameters. Meanwhile, a kind of linear truncation error has been made, ⎢ sθi cβi + sαi cθi sβi cθi cαi sθi sβi − sαi cθi cβi ai sθi ⎥ (1)
which makes Levenberg-Marquardt (L-M) algorithm work not well on =⎢⎣ − cαi sβi sαi cαi cβi di ⎦

highly nonlinear system. The Levenberg-Marquardt algorithm will fall 0 0 0 1
into the dilemma of stagnation when the search is close to the optimal
solution. Differential Evolution (DE) algorithm [23-26] can solve the Wherecθi = cosθi ,sθi = sinθi , and so on.θi ,di ,ai ,αi are the joint angle,
nonlinear problem well. It is a population-based global search optimi­ the link offset, the link length and the link twist angle, respectively.βi is a
zation algorithm, which has the characteristics of high reliability, strong small rotation angle of the joint axiszi around the axisyi . When the
robustness, good optimization performance. However, there are many adjacent joint axeszi andzi+1 are parallel,βi ∕
= 0, otherwise,βi = 0.
kinematic parameters to be identified and the computation is compli­ The robot kinematic model represented as Eq. (2) can be obtained by
cated, which will cause the problem of slow convergence of Differential multiplying the link transformation matrices.
Evolution algorithm. 0
T7 = 0 T1 1 T2 2 T3 3 T4 4 T5 5 T6 6 T7 (2)
Considering the above problems, this paper proposes a new kine­
matic parameters calibration method based on the established robot The deviations of D-H parameters cause the link transformation
kinematic error model. The Levenberg-Marquardt algorithm is used to matrix error which can be approximately calculated by the following
quickly calibrate the kinematic parameters of robot, with approximate equation.
values gained, which were used as the central values of the components
∂i Ti+1 ∂i Ti+1 ∂i Ti+1 ∂i Ti+1 ∂i Ti+1
of initial individuals of Differential Evolution algorithm. Then the di Ti+1 = Δθi + Δdi + Δai + Δαi + Δβi
∂θi ∂di ∂ai ∂αi ∂βi
Levenberg-Marquardt and Differential Evolution hybrid (LM-DEH) al­
gorithm with a better initial population and smaller search range was = δi Ti+1 ⋅i Ti+1
used to obtain more accurate value of each parameter. Differential (3)
Evolution algorithm was used to find the offset angles in the process of
WhereΔθi ,Δdi ,Δai ,Δαi ,Δβi are the parameter deviations of link i,
error compensation. And experiments have been conducted to validate

2
G. Luo et al. Robotics and Computer-Integrated Manufacturing 71 (2021) 102165

Fig. 2. The identification procedure of kinematic parameters.

anddi Ti+1 is the link transformation matrix error.δi Ti+1 is differential



6
(i )
transformation matrix. 0
T7 + d0 T7 = Ti+1 + di Ti+1 (4)
Taking the D-H parameter deviations of all links into consideration, i=0
the transformation matrix from robot base coordinate to robot end co­
Eq. (5) can be got by expanding Eq. (4) and ignoring the high-order
ordinate can be expressed as Eq. (4) based on the link transformation
differential terms.
matrix error.

Fig. 3. The joint space error compensation method.

3
G. Luo et al. Robotics and Computer-Integrated Manufacturing 71 (2021) 102165

( )

6 {[ ] }− 1
d 0 T7 = 0
Ti ⋅δi Ti+1 ⋅0 Ti − 1
⋅0 T7 (5) ΔXk = − J(Xk )T J(Xk ) + λk I J(Xk )T f (Xk ) (9)
i=0
The estimation value ofXat the(k + 1)thiteration is updated by the
And the mapping relationship between position errors and param­ following equation.
eter errors could be described as Eq. (6) [5, 29].
⎡ ⎤ Xk+1 = Xk + ΔXk (10)
⎡ ⎤ Δθ
dx ⎢ ⎥ When the iteration meets the stopping criteria or the maximum
⎢ Δd ⎥
ΔP = ⎣ dy ⎦ = [ J1 J2 J3 J4 J5 ]⎢ ⎢ Δa ⎥ = J⋅ΔX
⎥ (6) number of iterations is reached, the loop ends and the suboptimal values
dz ⎣ Δα ⎦ of kinematic parameter deviationsΔXare obtained by subtracting the
Δβ nominal kinematic parametersXN of robot from the identified kinematic
parametersX.
In Eq. (6),ΔPis the position errors vector of robot end in x, y, and z, its
elements are the first three lines of the last column of the matrixd0 T7 .Jis ΔX = X − XN (11)
the extended Jacobian matrix affected only by the nominal parameters
of the robot.ΔXis the parameter errors vector composed
ofΔθ,Δd,Δa,Δα,Δβ, each of them contains six parameters exceptΔβ, 3.2. DE algorithm based accurate identification of kinematic parameters
which has only one parameterΔβ2 . Because only the axesz2 andz3 are
parallel,Δβ2 is non-zero, the others(Δβ1 ,Δβ3 ,Δβ4 ,Δβ5 ,Δβ6 )are defined as The suboptimal values are taken as the central values of components
zero. Therefore, onlyΔβ2 needs to be identified. of individuals (xi,g ). Smaller upper and lower deviations are set for each
The robot kinematic error model reflects the mapping relationship component, and the initial population is randomly generated from the
between the position errors and the parameter deviations, it is the basis allowed domain. The difference vectors of individuals in the population
for the kinematic parameter identification based on the difference be­ are scaled and added to other different individuals (called target vectors)
tween the theoretical position and the actual position of robot end. to obtain the mutant vectors (vi,g ). The trial vectors (μi,g ) are generated
from mutant vectors by crossover operation. Individuals that survived
3. Identification of kinematic parameters and error the competition between the target vectors and the trial vectors are
compensation selected as the next-generation population individuals (xi,g+1 ). The
mutation, crossover, and selection operations are iteratively carried out
The procedure of identifying kinematic parameters is shown in until the algorithm reaches the maximum number of iterations or any
Fig. 2. In order to obtain the actual values of D-H parameters, the least other stopping criteria, with a best individual found as the optimal
square objective function is constructed based on the established robot values of kinematic parameter deviations.
kinematic error model and the L-M algorithm is used to find suboptimal
values of kinematic parameter deviations. L-M algorithm realizes the 1) Smaller upper and lower deviations can be set for each component of
combination of the advantages of the Gauss-Newton algorithm and the individual. Individuals (xi,g ,i = 1,2,...,NP) are represented as vectors
gradient descent method. It can obtain the suboptimal values at a faster of size25 × 1, each of them represents a candidate solution for the
speed with its fast convergence characteristics. LM-DEH algorithm could optimization problem.NPis the size of the initial population. Each
overcome the slow convergence problem of DE algorithm caused by the initial value of vector is generated by summing the corresponding
poor initial population and large search range, and obtain higher ac­ suboptimal value of kinematic parameter deviationsΔXobtained by
curacy of parameter identification. LM-DEH could find optimal values of L-M algorithm and a random number within the set upper and lower
parameter errors, taking the suboptimal values as central values of deviations. The initial population individuals are generated through
components of the initial population individuals. Eq. (12). Better initial population and smaller upper and lower limits
can greatly accelerate the convergence speed and obtain more ac­
3.1. Pre-identification of kinematic parameters based on L-M algorithm curate values of kinematic parameter deviations.

L-M algorithm applies the initial position errors vectorΔP(calculated xi,g = ΔX + V (12)
by subtracting the initial theoretical position vectorPt from the actual
position vectorPr of robot end in x, y, and z) to find present parameter
errors of robot. The kinematic model is updated based on the parameter
errors, and the updated theoretical position vectorPte is updated with the Vis a25 × 1vector, and each value of the vector is a random number
corrected kinematic model. Thus, the updated position errors vec­ within[tolL, tolU],tolLis the set lower deviation andtolUis the upper de­
torΔPe can be provided. In this loop, the suboptimal values of kinematic viation. tolLandtolUare used in a large sense to represent the lower and
parameter deviations that make the improvement of robot positioning upper deviations of each parameter. The values of lower and upper
accuracy are found through continuous iteration. deviations of each parameter are different. The boundary values of each
The least square objective function is constructed as Eq. (7). parameter are set separately. Under normal circumstances, the bound­
ary value of the angle parameter will be set smaller than the boundary
1
F(X) = f (X)T f (X) (7) value of the length parameter. In addition, the boundary values of pa­
2 rameters of LM-DEH can be the size of 1/3 of that of DE thanks to
suboptimal values from LM. 1/3 is an empirical value.
f (X) = [ΔP1 ΔP2 ⋯ ΔPn ]T (8)
In Eq. (8),Xis the D-H parameters vector,ΔPi (i = 1, 2, ..., n)is the 1) Mutant vector is obtained by adding the target vector to the differ­
position errors vector. ence vector, which is scaled by multiplying the scale factor F. The
Identifying the kinematic parameters of robot is to obtainXwhich used mutation strategy is listed as Eq. (13).
makesF(X)get the optimal value. With the extended Jacobian ( )
vi,g = xr5,g + F xr1,g − xr2,g + xr3,g − xr4,g (13)
matrixJ(Xk )and position error matrixf(Xk ), parameter errors vec­
torΔXk can be computed.Xk represents the D-H parameters vector at
thekthiteration.λk is the damping factor.
xr1,g , xr2,g , xr3,g , xr4,g , xr5,g (r1 ∕
= r2 ∕
= r3 ∕
= r4 ∕
= r5 ∕
= i) are randomly

4
G. Luo et al. Robotics and Computer-Integrated Manufacturing 71 (2021) 102165

Fig. 4. Robot kinematic parameters calibration procedure.

selected individuals in the population. Scale factor is used to amplify the y, and z. The actual position of robot end is measured by a laser tracker.
difference. If the number of iterations reaches the maximum value or any other
stopping criteria have met, the iteration is over. The individual obtained
1) A trial vector is generated with the crossover strategy described as at the end of the iteration with the smallest evaluation function value in
Eq. (14), enhancing the diversity of the population. the population is the optimal values of kinematic parameter deviations.
{ The real kinematic parameters can be calculated by Eq. (22).
vi,j,g , if rand < CRorrandi = j
ui,j,g = i = 1, 2, ..., NP; j = 1, 2, ..., 25
xi,j,g , if rand > CRorrandi ∕
=j
(14) 3.3. Error compensation

Due to the authority restriction of the industrial robot control system,


it is impossible to directly modify the D-H parameters in the control
In Eq. (14),vi,j,g andxi,j,g are thejthcomponents ofvi,g andxi,g respectively.
system. The user can only control theθparameter, so this article proposes
rand ∈ [0,1]is a random number.CR ∈ [0, 1] is the crossover rate. randi ∈
a kind of joint space error compensation method to compensate for
[1, 25] is a randomly selected integer.
parameter errors.
1) Mutation operation may cause the variables of trial vector to exceed
1) Calculate the joint angles (ϕi ,i = 1,2,...,6) of the target position (Tn )
the specified range. The variables that are out of range will be
of robot end with inverse kinematics.Tn is the position reached by the
reinitialized with Eq. (15), wherexi,j,min /xi,j,max can be calculated by
robot expectedly, computed by forward kinematics with nominal
adding theΔxj to the lower/upper deviation.
parameters.

⎨ xi,j,min , if ui,j,g < xi,j,min 2) Identified parameter errors (δdi ,δai ,δαi ,δβi ) are used to calculate the
ui,j,g = xi,j,max , if ui,j,g > xi,j,max i = 1, 2, ..., NP; j = 1, 2, ..., 25 (15) approximate position (Ts ) and the partial differential Eq. (23) is

ui ,j,g , else constructed.

6
xi,j,min = Δxj + tolL (16) T n − Ts =
∂Ts
Δθi (i = 1, 2, ..., 6) (23)
i=1
∂θi
xi,j,max = Δxj + tolU (17)

3) Since there are only six variables, the DE algorithm can quickly and
2) Based on the greedy competition mechanism, the survival in­ correctly obtain the offset anglesΔθi (i = 1, 2, ..., 6), taking the
dividuals are selected as the next-generation population individuals vector[Δθ1 , Δθ2 , Δθ3 , Δθ4 , Δθ5 , Δθ6 ]T as the individual of DE algo­
between the target vectors and the trial vectors. rithm’s population.
{ ( ) ( )
ui,g , if func ui,g < func xi,g 4) Update the joint angles with Eq. (24), whereδθi is the identified joint
xi,g+1 = (18) angle error, then the position of robot end is calculated with forward
xi,g , else
kinematics.
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
func(x) = ‖ (P(K(XN + x ) − K(XN )) − ΔP) ‖2 (19) θi = ϕi − Δθi + δθi , i = 1, 2, ..., 6 (24)

K(XN + x) = 0 T7 + d0 T7 (20)

K(XN ) = 0 T7 (21) The purpose of error compensation is to reduce the deviation be­
tween theoretical position and actual position of robot end. This method
X = XN + x (22) converts the error caused byδdi , δai , δαi , δβi to deviation caused by offset
anglesΔθi . The key is to find out better offset angles, which make the
value of e as small as possible. The value of e is computed by Eq. (26).
The smaller the value, the closer the two positions.
Eq. (19) is the individual evaluation function. It’s a highly nonlinear
system. The smaller the function value, the higher the individual fitness, ∑
6
∂Ts
T = (Tn − Ts ) − Δθi (25)
that is, the better the obtained individual.Kis the forward kinematics i=1
∂θi
operator.P(K(XN + x) − K(XN ))are the position errors in x, y, and z
√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
calculated by forward kinematics with corrected parameters(XN + √ 4
√∑ 2
x)and nominal parametersXN of robot kinematics.ΔPis a vector of the e=√ Tij (26)
initial positioning errors, which is calculated by subtracting the theo­ i=1,j=1

retical position vector from the actual position vector of robot end in x,

5
G. Luo et al. Robotics and Computer-Integrated Manufacturing 71 (2021) 102165

Table 2. set empirical value to 100 for LM. For DE and LM-DEH, the empirical
Pre-set parameter errors. values vary with the population size and search range. In this simula­
Joints Δθ /(rad) Δd/mm Δa/mm Δα/(rad) Δβ/(rad) tion, empirical values are set to 20000 for them and SAA. After
compensating for parameter errors, calculating the robot positioning
1 − 0.00035 0.02 − 0.7 0.00 0
errors and comparing them with that before compensation. Another fifty
2 0.0002 − 0.05 − 0.4 0.000035 0.004
3 0.0001 0.03 0.5 0.0014 0 points are selected as the verification points and calculating their posi­
4 0.004 − 0.04 0.5 − 0.0014 0 tioning errors after compensation.
5 0.00035 0.01 − 0.1 0.00087 0 One group of pre-set parameter errors selected randomly from some
6 − 0.008 − 0.06 − 0.2 − 0.001 0 simulations is shown in Table 2. The identified deviations of kinematic
parameters are list in Table 3. The relationship between the number of
4. Simulation and experiment iteration and the mean positioning error is shown in Fig. 5. It can be seen
from this figure that LM-DEH and DE algorithms are both capable of
In order to validate the effectiveness and correctness of LM-DEH finding the optimal values of kinematic parameter deviations, but under
algorithm for kinematic parameters calibration, numerical simulation the condition of the same iteration number of algorithms, the values
studies were performed with MATLAB software based on the established found by LM-DEH algorithm are better than values found by DE algo­
kinematic model in Section 2. In this simulation, without any noise rithm. The SAA has a faster convergence speed at the initial stage of the
considered, the robot forward and inverse kinematic models were pre­ iteration, but it spends more time to find the accurate values. Moreover,
sented to simulate the robot controller. Comparing the parameter errors the iterative result of SAA is easily affected by the values of the algo­
(pre-set factitiously randomly and identified by the algorithms) and the rithm parameters. As a result, LM-DEH found out better values than
robot positioning errors (before and after compensation) to evaluate the SAA.
parameter identification algorithm and error compensation algorithm, The start mean positioning error of LM is about 3 mm, it is influenced
respectively. by the search direction and step length. Small step won’t cause big start
The procedure of calibration can be seen from Fig. 4. Because the mean error. The start mean positioning errors of DE and LM-DEH are
robot has six degrees of freedom, a total of 24 D-H parameters need to be influenced by the initial population. The initial population is randomly
identified. In addition, robot’s joint axesz2 andz3 are parallel, so one generated from the allowed domain. Individuals of population could be
more parameterβ2 needs to be identified. Each set of robot end position any random points within search range. DE has a bigger search range, as
can provide three equations. To identify the 25 parameters, at least nine the red circle showed in Fig. 6, so it has a bigger start mean error. For
group equations are needed. The more position provided, the higher the example, in the simulation of DE, the values of lower and upper de­
accuracy of identified parameters. Fifty sample points in the robot viations are set to 5 (mm) for length parameters, 0.14 (rad) for angle
workspace are selected randomly, and their simulated actual positions parameters. If all parameters are calculated with the maximum devia­
are calculated by the forward kinematic model with contaminated ki­ tion value, the positioning error will be calculated as about 460 mm. LM-
nematic parameters. L-M, DE, and LM-DEH algorithms are used to DEH has a smaller search range represented by the green circle by
identify the parameters respectively. SAA is a probability-based, widely minimizing the search range of DE through combination with LM.
used optimization algorithm. It is also reproduced to identify the kine­ Correspondingly, LM-DEH has a smaller start mean error.
matic parameters, to compare with the proposed method. Based on the The robot positioning improvements were analyzed after calibration,
parameter settings of the algorithms, empirical values of maximum the positioning errors of the robot after compensation have been
number of iterations are set for LM-DEH, DE, SAA and LM respectively. compared in Fig. 7. Table 4 provides a list of statistics about positioning
The empirical values ensure that the iterative results of algorithms error after compensation at around 16500 iterations.
approximately stable when the number of iterations is close to its The confirmation experiment was carried out on the FANUC
maximum value. In general, LM could converge quickly, it is enough to M710ic/50 robot after numerical simulation study. Actual positions of

Table 3.
The identified deviations of kinematic parameters.
algorithm joint Δθ/(rad) Δd/mm Δa/mm Δα/(rad) Δβ/(rad)

L-M 1 − 0.000388 0.025041 − 0.715231 − 0.00004 0


2 0.000207 − 0.031673 − 0.419713 0.000106 0.004011
3 0.000105 0.029991 0.500341 0.001405 0
4 0.00399 − 0.02786 0.510085 − 0.001397 0
5 0.000351 0.003814 − 0.115992 0.000877 0
6 − 0.008 − 0.036092 − 0.196405 − 0.001002 0
DE 1 − 0.000350 0.019683 − 0.700006 0.000003 0
2 0.000202 2.864259 − 0.399996 0.000035 0.004000
3 0.000100 2.944080 0.500009 0.001400 0
4 0.004000 − 0.040004 0.499977 − 0.001400 0
5 0.000350 0.009977 − 0.099983 0.000870 0
6 − 0.008000 − 0.060003 − 0.199999 − 0.001000 0
LM-DEH 1 − 0.000350 0.020155 − 0.700000 − 0.000002 0
2 0.000199 − 0.145955 − 0.400001 0.000035 0.004000
3 0.000100 − 0.065960 0.500000 0.001400 0
4 0.004000 − 0.040001 0.500000 − 0.001400 0
5 0.000350 0.010000 − 0.100001 0.000870 0
6 − 0.00800 − 0.059999 − 0.200001 − 0.01000 0
SAA 1 − 0.000349 0.024087 − 0.677334 − 0.000005 0
2 0.000173 − 0.083027 − 0.410617 0.000036 0.003999
3 0.000516 0.162475 0.006704 0.001552 0
4 0.003979 − 0.113885 0.496863 − 0.001148 0
5 0.015669 0.723961 − 2.756408 0.004850 0
6 − 0.138075 − 0.061173 − 0.198294 0.123776 0

6
G. Luo et al. Robotics and Computer-Integrated Manufacturing 71 (2021) 102165

Fig. 5. The relationship between the number of iteration and the positioning error.

The initial mean positioning error is calculated as 0.994 mm. Table 5


lists the actual deviations of parameters identified by the L-M, DE, SAA
and LM-DEH algorithms. Fig. 9 shows the comparison of actual posi­
tioning errors after calibration. Table 6 shows the maximum, minimum
and average values of positioning errors.
Through comprehensive comparison, what is obvious is that the
positioning accuracy has improved a lot after calibration with algo­
rithms. The robot positioning error decreased from 0.994 to 0.443,
0.439 mm and 0.350 mm after L-M, DE and SAA calibration respec­
tively. In contrast, the positioning accuracy has a greater improvement
after calibration with LM-DEH algorithm, robot positioning error
decreased from 0.994 mm to 0.262 mm. In simulation, considering no
noise, robot positioning accuracy has been greatly improved compared
Fig. 6. Minimizing the search range.
with the results of experiment. The robot positioning error dropped from
3.357 mm to less than 0.1 mm after L-M calibration, reaching 0.0702
robot end were measured by a FARO Vantage laser tracker, shown in
mm. The robot positioning accuracy improved to 0.055 mm from 3.357
Fig. 8. They will be influenced by the inevitable measurement noise,
mm after SAA calibration. The positioning accuracy after DE calibration
such as the measurement error of laser tracker, the coordinate system
has increased by 4.6% in performance over the positioning accuracy of
transformation error between the robot and laser tracker.
L-M algorithm, and the positioning accuracy after LM-DEH calibration
PEB is the actual position of robot end relative to the base, it is ob­ has also increased by 85%. Though the two DE and LM-DEH algorithms
tained by Eq. (27).PEL is the actual position of robot end relative to the both can find optimal values, the number of iterations of the former is at
laser tracker.TLB is a four-dimensional transformation matrix between least 2 times that of the latter, resulting in the poor efficiency. Those
laser tracker and the base, which is calculated by the ARTS-5025 mea­ results demonstrate that the proposed kinematic parameter calibration
surement software provided by the Premax company, complying with method is very effective and greatly improves the absolute positioning
the measurement standard GB/T 12642–2013. The initial mean posi­ accuracy of the robot.
tioning error is calculated by Eq. (28).(x,y,z)is the theoretical position of
robot end relative to the base computed by forward kinematics with 5. Application in robotic grinding of aero-engine blade
nominal parameters.nis the number of calibration configurations. The
experiment was completed following the steps performed in simulation. Since aero-engine blades are mostly composed of curved surfaces and
curves, the trajectory of robotic grinding of blades is difficult to accu­
PEB = [xr yr zr T
1] = TLB ⋅PEL (27)
rately obtain through teaching programming. For the path algorithms of
n √̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅
/ robotic machining, most of the existing results are based on offline

e= 2
(xri − xi ) + (yri − yi ) + (zri − zi ) 2 2
n (28) programming of a CAD model. Fig. 10 shows the trajectory generated
i=1 based on the tool location data of the existing model and the CAD

7
G. Luo et al. Robotics and Computer-Integrated Manufacturing 71 (2021) 102165

Fig. 7. The comparison of simulation results about positioning errors.

robot is low, there will be a deviation between the planned path and the
Table 4. actual path.
Statistics of positioning error after compensation. The improvement of the robot’s absolute positioning accuracy can
Mean(mm) Min(mm) Max(mm) make the contour of the blade after machining within the tolerance
Simulation Before calibration 3.357 0.494 7.086 range, as shown by the red dashed line in Fig. 13, but the contour shape
After L-M calibration 0.0702 0.0386 0.1251 may have a large deviation from the theoretical contour. Dynamic
After DEcalibration 0.067 0.024 0.099 calibration and compensation methods can further improve the path
After SAA calibration 0.055 0.006 0.125 accuracy so that the contour shape of the blade after grinding and the
After LM-DEH calibration 0.0105 0.0018 0.0194
theoretical profile approximately coincide.
Poor absolute positioning accuracy will cause the phenomenon of
software through offline programming method. Robotic machining of grinding misplacement. Fig. 14 shows an example of misplacement of
blades requires that robot can follow the planned complex path and grinding when the blade edge is machined by a robot. The area enclosed
guarantee the fine path accuracy. by black boundary lines is the area to be machined. Blue line is the
The main factors affecting the path accuracy are errors related to the theoretical trajectory generated through offline programming, and the
robot kinematic model and dynamic model. Robot kinematic calibration red line is the actual grinding trajectory of the robot, calculated with
can ensure to make the planned path and the actual path as close as pre-set parameter errors. It can be seen from the figure that there is an
possible, thereby improving the absolute accuracy of path to a certain average deviation about 1.4547 mm between the actual trajectory and
extent, as shown in Fig. 11. The planned path is generated by connecting the theoretical trajectory, causing that a certain of the non-machining
discrete points in sequence, showed in the Fig. 12. This method is called area is processed, while part of the area should be machined remains.
parametric method. The actual path is generated by connecting actual Robot kinematic calibration can reduce the deviation by compen­
discrete points reached by robot. If the absolute positioning accuracy of sating errors. Compensated trajectories are generated after calibration.

Fig. 8. Experimental system.

8
G. Luo et al. Robotics and Computer-Integrated Manufacturing 71 (2021) 102165

Table 5.
Actual deviations of kinematic parameters identified by algorithms.
algorithm joints Δθ/(rad) Δd/mm Δa/mm Δα/(rad) Δβ/(rad)

L-M 1 0.000111 1.222202 − 0.435222 0.000114 0


2 0.000359 − 0.049608 0.583631 0.000056 − 0.000185
3 0.001574 0.049417 − 1.827721 − 0.000029 0
4 0.000045 − 0.791819 − 0.027113 − 0.000237 0
5 − 0.052606 8.042368 9.198078 0.047134 0
6 0.000276 − 0.008693 0.510776 − 0.000 0
DE 1 0.001672 1.305055 − 0.426379 − 0.000660 0
2 0.000529 − 2.842804 0.343183 0.000 − 0.001478
3 − 0.000571 − 1.275925 0.309864 − 0.000807 0
4 − 0.000187 − 0.386244 − 0.111792 − 0.000907 0
5 0.026450 − 0.094273 − 4.599326 0.000458 0
6 0.000253 0.059337 0.108804 0.100000 0
LM-DEH 1 0.000061 1.426262 − 0.407457 0.000153 0
2 0.000409 0.440833 0.567402 0.000055 − 0.000168
3 0.001078 0.532052 − 1.371941 0.000054 0
4 0.000061 − 0.694273 − 0.025637 − 0.000374 0
5 − 0.055473 8.183331 9.690090 0.048125 0
6 0.000271 0.038356 0.544841 0.014414 0
SAA 1 − 0.000010 − 0.112934 − 0.183599 0.000101 0
2 − 0.000007 − 3.016715 0.514422 0.000036 − 0.000070
3 0.004325 − 3.181672 − 4.498543 0.000057 0
4 0.000457 − 1.101572 − 0.040260 − 0.000572 0
5 0.003763 2.365600 − 0.668226 0.014786 0
6 0.089401 − 0.009181 0.021152 0.098925 0

Fig. 9. The comparison of actual positioning errors after calibration.

Table 6
. The positioning errors.
Mean(mm) Min(mm) Max(mm)

Experiment Before calibration 0.994 0.402 1.394


After L-M calibration 0.443 0.241 0.822
After DEcalibration 0.439 0.181 0.746
After SAA calibration 0.350 0.053 0.633
After LM-DEH calibration 0.262 0.036 0.702

The trajectory marked in green is generated after LM-DEH calibration.


The medium olive and pale maroon ones are generated after L-M and DE
calibration respectively. The comparison of the compensated trajec­
tories, generated with no noise considered, with the actual trajectory
and theoretical trajectory is shown in Fig. 15. After calibration with L-M
and DE algorithms, the deviations have dropped to about 0.06 mm and
0.001 mm, the declining contents of deviations after LM-DEH calibra­
Fig. 10. Offline programming for trajectory planning based on CAD model.
tion is almost the same as the decrease after DE calibration. The
compensated trajectories shown in Fig. 16 are generated with mea­
surement noise considered. In this case, the deviation dropped to 0.38 M calibration. It can be seen in the partial enlarged images that the
mm after L-M calibration, and both the deviations after DE and LM-DEH deviation between the compensated trajectory and theoretical trajectory
calibration have decreased by about 47.4% compared with that after L- has significantly decreased compared with the actual trajectory. In

9
G. Luo et al. Robotics and Computer-Integrated Manufacturing 71 (2021) 102165

Fig. 11. Path accuracy.

Fig. 15. The comparison of the trajectories with no noise considered.

Fig. 12. Path generation.

Fig. 16. The comparison of the trajectories with measurement


noise considered.

The method proposed in this paper only calibrates the geometrical


error of the robot, without taking non-geometrical error into account.
Non-geometrical error caused by like flexibility of links or stiffness of
gears or thermal deformation should be further considered to compen­
Fig. 13. Impact of calibration on trajectory.
sate for non-geometrical errors, thereby further improving robot posi­
tioning accuracy. The selection of sampling points is also worthy of
being studied. If the number of sampling points is too small, or the
location of the sampling points is too concentrated in space, the
improvement of positioning accuracy will be affected. In addition, the
intensive study about path accuracy will be conducted. In the next stage,
more calibration methods will be reproduced and the comparison of the
proposed method with other methods will be done.

6. Conclusion

This paper proposed a novel hybrid algorithm for robot kinematic


calibration, and a series of simulations and experiments have been
conducted to validate the algorithm. The conclusions are summarized as
follows.

Fig. 14. The deviation between the actual trajectory and the theoret­ (1) Under the condition of taking no noise into account, in simula­
ical trajectory.
tion, the positioning accuracy is 3.357 mm before calibration.
And it has improved by 4.5%, 21.65% and 85% after DE, SAA and
contrast with trajectory generated after L-M calibration, trajectories LM-DEH calibration respectively over the positioning accuracy
generated after LM-DEH and DE calibration are closer to the theoretical after L-M calibration, which has improved to 0.0702 mm. These
trajectory. directly show that the proposed hybrid algorithm is correct and
The comparisons show that kinematic calibration can significantly effective.
improve the positioning accuracy of robot. The improvement of posi­ (2) After calibration experiment involving measurement noise, the
tioning accuracy is beneficial to enhance the absolute accuracy of path, robot positioning accuracy has been improved from 0.994 mm to
thereby making better improvement on the phenomenon of misplace­ 0.443 mm, 0.439 mm and 0.350 mm after L-M, DE and SAA
ment of grinding, and further improving the processing quality of the calibration respectively. In contrast, the positioning accuracy has
blade.

10
G. Luo et al. Robotics and Computer-Integrated Manufacturing 71 (2021) 102165

a greater improvement after calibration with LM-DEH algorithm, [9] L. Wu, X. Yang, K. Chen, H. Ren, A Minimal POE-Based Model for Robotic
Kinematic Calibration with Only Position Measurements, IEEE Trans. Autom. Sci.
robot positioning accuracy has improved from 0.994 mm to
Eng. 12 (2) (2015) 758–763.
0.262 mm. The improved positioning accuracy calibrated with [10] X. Chen, Q. Zhang, Y. Sun, Non-kinematic calibration of industrial robots using a
effective algorithms can improve the trajectory accuracy to a rigid–flexible coupling error model and a full pose measurement method, Robot.
certain extent. Comput. Integr. Manuf. 57 (2019) 46–58.
[11] W. Wang, H. Song, Z. Yan, L. Sun, Z. Du, A universal index and an improved PSO
(3) In terms of trajectory deviation, after calibration with L-M algo­ algorithm for optimal pose selection in kinematic calibration of a novel surgical
rithm, the average deviation between the theoretical and actual robot, Robot. Comput. Integr. Manuf. 50 (2018) 90–101.
trajectory is reduced from 1.4547 mm to 0.38 mm. The DE and [12] T. Messay, R. Ordóñez, E. Marcil, Computationally efficient and robust kinematic
calibration methodologies and their application to industrial robots, Robot.
LM-DEH algorithms have better performance with both the Comput. Integr. Manuf. 37 (2016) 33–48.
average deviations dropped from 1.4547 mm to around 0.2 mm. [13] H.-.N. Nguyen, P.-.N. Le, H.-.J. Kang, A new calibration method for enhancing
If the noise is not considered, the average deviations after cali­ robot position accuracy by combining a robot model–based identification approach
and an artificial neural network–based error compensation technique, Adv. Mech.
bration are no longer 0.38 mm and 0.2 mm, but 0.06 mm and Eng. 11 (1) (2019).
0.001 mm. The reduction of the deviation is beneficial to improve [14] Y. Cho, H.M. Do, J. Cheong, Screw based kinematic calibration method for robot
the processing quality of the blade. manipulators with joint compliance using circular point analysis, Robot. Comput.
Integr. Manuf. 60 (2019) 63–76.
[15] Y. Gan, J. Duan, X. Dai, A calibration method of robot kinematic parameters by
drawstring displacement sensor, Int. J. Adv. Rob. Syst. 16 (5) (2019).
Declaration of Competing Interest [16] G.B. Gao, H.W. Zhang, H.J. San, G.Q. Sun, X. Wu, W. Wang, Kinematic calibration
for industrial robots using articulated arm coordinate machines, Int. J. Model.
Identif. 31 (1) (2019) 16–26.
The authors declare no conflict of interest. [17] Q. Zhu, X. Xie, C. Li, G. Xia, Q. Liu, Kinematic Self-Calibration Method for Dual-
Manipulators Based on Optical Axis Constraint, IEEE Access 7 (2019) 7768–7782.
[18] C. Kanzow, N. Yamashita, M. Fukushima, Levenberg–Marquardt methods with
Acknowledgements strong local convergence properties for solving nonlinear equations with convex
constraints, J. Comput. Appl. Math. 172 (2) (2004) 375–397.
This study was supported by the National Natural Science Founda­ [19] J. Fan, J. Huang, J. Pan, An Adaptive Multi-step Levenberg–Marquardt Method,
J. Sci. Comput. 78 (1) (2018) 531–548.
tion of China (Grant No. 52075059) and the Natural Science Foundation [20] J. Shawash, D.R. Selviah, Real-Time Nonlinear Parameter Estimation Using the
of Chongqing (cstc2020jcyj-msxmX0266). Levenberg–Marquardt Algorithm on Field Programmable Gate Arrays, IEEE Trans.
Indust. Electron. 60 (1) (2013) 170–176.
[21] W. Tian, M. Mou, J. Yang, F. Yin, Kinematic calibration of a 5-DOF hybrid
References kinematic machine tool by considering the ill-posed identification problem using
regularization method, Robot. Comput. Integr. Manuf. 60 (2019) 49–62.
[1] D. Zhu, X. Feng, X. Xu, Z. Yang, W. Li, S. Yan, H. Ding, Robotic grinding of complex [22] G. Zak, B. Benhabib, R.G. Fenton, I. Saban, Application of the Weighted Least
components: a step towards efficient and intelligent machining – challenges, Squares Parameter Estimation Method to the Robot Calibration, J. Mech. Des. 116
solutions, and applications, Robot. Comp.- Integrat. Manuf. 65 (2020). (3) (1994) 890–893.
[2] R.P. Judd, A.B. Knasinski, A Technique to Calibrate Industrial Robots with [23] A.R. Khaparde, Analysis of New Distributed Differential Evolution Algorithm with
Experimental-Verification, IEEE Trans. Robot. Automat. 6 (1) (1990) 20–30. Best Determination Method and Species Evolution, Procedia Comput. Sci. 167
[3] Z. Roth, B. Mooring, B. Ravani, An overview of robot calibration, IEEE J. Robot. (2020) 263–272.
Automat. 3 (5) (1987) 377–385. [24] L. Cui, G. Li, Q. Lin, J. Chen, N. Lu, Adaptive differential evolution algorithm with
[4] S.H. Kim, E. Nam, T.I. Ha, S.-.H. Hwang, J.H. Lee, S.-.H. Park, B.-.K. Min, Robotic novel mutation strategies in multiple sub-populations, Comput. Oper. Res. 67
Machining: a Review of Recent Progress, Int. J. Precis. Eng. Manuf. 20 (9) (2019) (2016) 155–173.
1629–1642. [25] Y. Li, S. Wang, B. Yang, An improved differential evolution algorithm with dual
[5] Z. Jiang, W. Zhou, H. Li, Y. Mo, W. Ni, Q. Huang, A New Kind of Accurate mutation strategies collaboration, Expert Syst. Appl. (2020) 153.
Calibration Method for Robotic Kinematic Parameters Based on the Extended [26] T. Yaghoobi, Parameter optimization of software reliability models using improved
Kalman and Particle Filter Algorithm, IEEE Trans. Indust. Electron. 65 (4) (2018) differential evolution algorithm, Math. Comput. Simul. 177 (2020) 46–62.
3337–3345. [27] S.A. M’hiri, N.M. Ben Romdhane, T. Damak, New Forward Kinematic Model of
[6] G. Gao, F. Liu, H. San, X. Wu, W. Wang, Hybrid Optimal Kinematic Parameter Parallel Robot Par4, J. Intell. Robot. Syst. 96 (2) (2018) 283–295.
Identification for an Industrial Robot Based on BPNN-PSO, Complexity (2018). [28] H.L. Yuan, X. Li, H. Hu, W. Ding, R.K. Agarwal, P. Tandon, E.X. Wang, Two Error
[7] R. Wang, A. Wu, X. Chen, J. Wang, A point and distance constraint based 6R robot Models for Calibrating SCARA Robots based on the MDH Model, MATEC Web of
calibration method through machine vision, Robot. Comput. Integr. Manuf. 65 Conferences (2017) 95.
(2020). [29] J. Qi, B. Chen, D. Zhang, Compensation for absolute positioning error of industrial
[8] I.M. Chen, G.L. Yang, C.T. Tan, S.H. Yeo, Local POE model for robot kinematic robot considering the optimized measurement space, Int. J. Adv. Rob. Syst. 17 (2)
calibration, Mech. Mach. Theory 36 (11–12) (2001) 1215–1239. (2020).

11

You might also like