IEEE TRANSACTIONS ON ROBOTICS, VOL. 36, NO.

5, OCTOBER 2020 1577

An Iterative Approach for Accurate Dynamic Model

Identification of Industrial Robots
Yong Han , Jianhua Wu , Chao Liu , and Zhenhua Xiong

Abstract—Dynamic model has broad applications in motion high-speed motions or heavy payloads. Dynamic models are
planning, feedforward controller design, and disturbance observer also utilized to constitute disturbance observers in addition to
design. Particularly, with the increasing application of model- many advanced controllers. For instance, a virtual force sensor
based control in industrial robots, there has been a resurgence
of research interest in accurate identification of dynamic models. was built by using the robot dynamic model [2] so that contact
However, on the one hand, most existing identification methods force can be estimated without the use of expensive force/torque
directly rely on least squares or weighted least squares (WLS), sensors. However, accurate parameters of the dynamic model of
which suffer from outliers and could lead to physical infeasible a robot-like mass, center of mass, and inertia tensor are typically
solutions. On the other hand, nonlinearity of the friction model is unknown even to the manufacturer of the robot. Even though
seldom treated in a unified way with linear regression. Moreover,
recent researches have shown that proper exciting trajectories are these parameters may be estimated from a computer-aided-
crucial to the identification accuracy, but few of previous works take design (CAD) software [3], the values are not accurate due to
measurement noise into consideration when optimizing the exciting the existence of motors, reducers, wires, and other components.
trajectories. In this article, we propose an iterative approach which Moreover, joint frictions can only be identified by experiments.
integrates WLS, iteratively reweighted least squares with linear Hence, accurate identification of dynamic models is of great
matrix inequality constraints, and nonlinear friction models so
that the above-mentioned issues can be properly solved altogether. importance and necessity for industrial robots.
Our research also reveals that performance can be improved by Dynamic model identification of an industrial robot has a
including priori knowledge of measurement noise in the optimiza- long history since the 1980s. Major achievements have been
tion of exciting trajectories. The proposed approach is supported made on derivation of a linear model with base parameters,
by experimental analysis of four different combinations within the exciting trajectories design, robust estimators, physical feasi-
framework on a 6-DoF industrial robot.
bility constraints of model parameters, and joint frictions. A
Index Terms—Dynamic model identification, friction model, comprehensive review on these topics was summarized in [4].
industrial robots, linear regression. As an experimental endeavor, Atkeson et al. [5] proposed
a method to estimate inertia parameters of manipulator’s links
I. INTRODUCTION and loads and provided experimental validations. Two numerical
approaches were used to deal with the rank-deficient problem
ANY applications of industrial robots require the avail-
M ability of a dynamic model. For instance, offline pro-
gramming with task optimization and advanced robot controller
in the regressor matrix. Later on, Gautier et al. [6], [7] provided
an analytical solution to this problem by recursively grouping
inertial parameters of two adjoint links to directly calculate
design were introduced as two typical applications based on
the minimum set of serial robots’ inertial parameters, which is
robots’ dynamic models [1]. In the former, realistic simulation
corresponding to the maximum linear independent columns in
and optimization of the robot manipulator’s motion, considering
the regressor matrix. Furthermore, a software called SYMORO
its physical constraints, requires an accurate dynamic model.
was developed to make the derivation of the regressor matrix
In the latter, advanced model-based controllers are proposed to
easier [8].
compensate the dynamic forces like inertial force, centrifugal
Then, exciting trajectories [9]–[11] have to be optimized to
and Coriolis forces so that tracking errors can be reduced for
persistently excite the regressor matrix since the identification
accuracy intrinsically depends on the excitation properties of the
Manuscript received October 21, 2019; revised February 17, 2020; accepted
April 15, 2020. Date of publication May 20, 2020; date of current version recorded motion [12]. Most works optimize a criterion related to
October 1, 2020. This work was supported in part by the National Natural Science the condition number of the whole regressor matrix [13]–[16].
Foundation of China under Grant U1813224 and Grant 91848106 and in part by However, Venture et al. [17] have shown that regressors which
the Program of Shanghai Academic/Technology Research Leader under Grant
18XD1401700. This article was recommended for publication by Associate share similar condition numbers can have different excitation
Editor A. M. Zanchettin and Editor P. Robuffo Giordano upon evaluation of the properties and subregressors are defined to characterize these
reviewers’ comments. (Corresponding authors: Jianhua Wu; Zhenhua Xiong.) properties. Bonnet et al. [12] used the condition numbers of
The authors are with the State Key Laboratory of Mechanical Sys-
tem and Vibration, School of Engineering, Shanghai Jiao Tong University, the subregressors to characterize the optimization criterion and
Shanghai 200240, China (e-mail: [email protected]; [email protected]; divided the optimization process into two static and dynamic
[email protected]; [email protected]). phases, leading to excellent results. Indeed, the inclusion of the
Color versions of one or more of the figures in this article are available online
at https://ieeexplore.ieee.org. subregressors’ condition number in optimization criterion not
Digital Object Identifier 10.1109/TRO.2020.2990368 only optimizes the condition number of the whole regressor

1552-3098 © 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
1578 IEEE TRANSACTIONS ON ROBOTICS, VOL. 36, NO. 5, OCTOBER 2020

matrix but also optimizes the regressor’s internal structure. joint. An interesting work that does not need to separate iden-
When weighted least squares (WLS) is applied, however, the tification of nonlinear friction models for each joint was com-
optimization should be performed on the weighted regressor. mitted in [44]. Recently, friction models with thermal and load
Therefore, it could be interesting to include the priori knowledge dependency have been studied in [45]–[47], where nonlinear
of measurement noise in the optimization process, which is due viscous friction (with respect to joint velocity) is also observed.
in the literature. However, how to combine these nonlinear friction models with
Although WLS or least squares (LS) estimators are widely linear regression in a unified way remains covered.
applied in dynamic model identification [1], [18], [19], it is In this article, we propose a new iterative method for accurate
well-known that they are susceptible to outliers [20], [21] dynamic model identification of industrial robots. The major
and, when the regressor matrix is corrupted with noise, could contributions are as follows. 1) The proposed method offers a
lead to a biased estimation [22]. About the first matter, set unified identification framework, which seamlessly integrates
membership estimation techniques [23], [24] are introduced. WLS, iteratively reweighted least squares (IRLS) with LMI
However, special care should be taken when choosing prior constraints and friction models, for improved robustness and
error bounds. Janot et al. [25] introduced robust regressions and accuracy in identification results. Particularly, both linear and
residual analysis to verify the reliability of LS estimation for nonlinear friction models can be addressed with linear regression
dynamic model identification of industrial robots. The results, in a unified way. 2) Subregressors based on the type of physical
however, are only verified on a haptic interface of one DoF and quantities and priori knowledge of measurement noise of joint
there is a lack of criterion on how to tune H parameter of the torques can be taken into account by the optimized exciting
Huber’s weight function. About the second matter, mathematical trajectories. 3) Experimental analysis of four different combina-
methods [26]–[28] are proposed. Nonetheless, it is hard to apply tions is conducted on their computational cost and performance,
them for dynamic model identification since errors appear in the whose results confirm that the proposed nonlinear friction model
regressor matrix in a nonlinear way. In practice, we either resort and weight function show the best overall performance within
to filtering [1], [18] or use the simulation data of a direct dynamic the framework.
model [29], [30]. The rest of the article is arranged as follows. Section II
Another drawback of WLS or LS estimators is that they introduces the background of the inverse dynamic model, the
could lead to physically unfeasible solutions [31], [32]. An WLS method, and LMI constraints on inertial parameters.
important work that enforces physical feasibility constraints Section III introduces the proposed iterative approach for ac-
on base parameters is based on the linear matrix inequality curate model identification of industrial robots. Section IV pro-
(LMI)-semidefinite programming (SDP) framework [33]. The vides experimental results, where residual analysis is heavily
results of the fully developed version of the framework show that used and comparisons of four combinations under the framework
enforcement of the physical feasibility constraints can mitigate are made. Finally, Section V concludes this article.
overfitting [34]. When priori knowledge of the standard inertial
parameters is available (such as the ones derived from CAD II. BACKGROUND
models), it is possible to combine physical feasibility constraints The inverse dynamic model of a robot with n moving links
with these parameters either as a regularizer [35], [36] or as a calculates the joint forces/torques vector τ as a function of
priori to retrieve standard dynamic parameters [37], [38]. In [39], generalized coordinates and their derivatives as follows:
an additional condition that ensures the triangle inequality of the
inertia matrices is implemented. This new constraint together τ = M (q) · q̈ + C(q, q̇) · q̇ + g(q) + f (1)
with the positive definiteness constraint is reformulated as an where M (q) · q̈ is the inertia force/torque, C(q, q̇) · q̇ is the
LMI constraint and bounding-volume constraints on the mass Coriolis and centrifugal force/torque, g(q) is the gravity, and f
distribution are enforced in [35]. Since solutions of previous is the friction. Traditionally, for jth joint, the friction is modeled
methods probably lie on the boundaries of the constraints, Lee as follows:
and Park [36] proposed a geometric algorithm based on the
Riemannian metric so that more robust and physically consistent fj = F cj · sign(q̇j ) + F vj · q̇j + Bj
estimation can be achieved even from insufficient excitation.
= [sign(q̇j ) q̇j 1] · [F cj F vj Bj ]T (2)
Unluckily, this algorithm takes a large amount of time and fails
to include frictions. Finally, as concluded in [35], it would be with F cj being the Coulomb’s friction, F vj being the viscous
interesting to incorporate the physical feasibility constraints into friction coefficient, and Bj being the bias or offset accounting
a framework where stochasticity gets a quantified treatment. for nonsymmetrical friction and nonzero force/torque offsets.
Since most industrial robots rely on geared drives where fric- Equation (1) can be rewritten as a linear relation to a set of
tions are present, it is of high importance to model and identify standard dynamic parameters [48] as follows:
frictions. Conventional methods of dynamic model identifica-
τ = W (q, q̇, q̈) · θ (3)
tion for industrial robots generally adopt Coulomb’s friction
plus linear viscous friction model [1], [18], [19]. However, such where θ is a (14n × 1) vector of standard dynamic
friction models are not adequate [40]. Although more complex parameters with θ j = [XXj XYj XZj Y Yj Y Zj ZZj
friction models [41]–[43] are adopted in the identification, they M Xj M Yj M Zj Mj Iaj F vj F cj Bj ]T , which is composed
require separate identification of friction parameters of each of, for each link j, six components of the inertia tensor,

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
HAN et al.: ITERATIVE APPROACH FOR ACCURATE DYNAMIC MODEL IDENTIFICATION OF INDUSTRIAL ROBOTS 1579

(XXj XYj XZj Y Yj Y Zj ZZj ); three components of the where Λjj is the jth element of the diagonal matrix Λ, Ej is
first moment (M Xj M Yj M Zj ); the mass Mj ; the total the jth row of the matrix E, and || · ||2 is the 2-norm operator.
inertia moment for rotor and gears Iaj ; the friction parameters It should be noted that Λ can also be applied to Kalman filter
F vj , F cj , and Bj . design of an external force observer [49].
According to [6] and [7], (3) can be further simplified as Finally, the regression coefficients can be estimated by the
follows: WLS estimator
τ = Y (q, q̇, q̈) · π (4) π = (YT · D−1 · Y)−1 · YT · D−1 · T (13)
where Y (q, q̇, q̈) is the regressor matrix composed of the where D is a (mn × mn) block diagonal matrix with its diagonal
maximum linear independent columns of W and π is base blocks being m Λ estimated by (12). Estimation of the regression
parameters. coefficients’ variance changes accordingly
From (4), the dynamic model identification can be formulated
as a multiple linear regression problem, i.e., varπ = diag((YT · D−1 · Y)−1 ). (14)

τ = Y (q, q̇, q̈) · π + ε (5) According to [34] and [35], the following LMI constraint (for
jth link) should be satisfied in order to make the identified results
where τ is the (n × 1) response vector, Y (q, q̇, q̈) is the (n × p) physically feasible:
regressor matrix, π is the (p × 1) regression coefficients, and ε ⎡ ⎤
tr(Lj )
is the (n × 1) error vector with the following property:
⎣ 2 I − Lj l j ⎦ > 0 (15)
Cεε = E(ε · εT ) = Λ (6) lTj Mj

where E(·) is the expectation operator and Λ is a (n × n) where I is the (3 × 3) identity matrix, tr(·) is the trace function,
diagonal matrix with each diagonal element being the expected and Lj , lj are defined as follows:
variance of the corresponding measured joint force/torque noise. ⎡ ⎤ ⎡ ⎤
Assuming that m measurement points are obtained during an XXj XYj XZj M Xj
⎢ ⎥ ⎢ ⎥
experiment, the observation matrix and observation response Lj = ⎣ XYj Y Yj Y Zj ⎦ , lj = ⎣ M Yj ⎦ .
can be computed as follows: XZj Y Zj ZZj M Zj
⎡ ⎤
Y1
⎢ ⎥ III. AN ITERATIVE APPROACH FOR DYNAMIC
⎢Y 2 ⎥
Y= ⎢ . ⎥
⎢
⎥, (7) MODEL IDENTIFICATION
⎣ .. ⎦ A. The Normalized Linear Regression
Ym
In the above section, it is assumed that noises in the measured
T= [τ T1 τ T2 ... τ Tm ]T (8) forces/torques are independent among joints, which leads to a
diagonal covariance matrix. Here we generalize the assumption
where Y is of size (mn × p) and T is of size (mn × 1). by allowing noises in the measured forces/torques of different
The regression coefficients can then be computed as follows joints to be correlated so that the covariance of the error vector
by LS: does not have to be diagonal, i.e.,
π = (YT · Y)−1 · YT · T (9) Cεε = E(ε · εT ) = Ω. (16)
and an estimation of the regression coefficients’ variance can be Accordingly, it can be estimated as follows:
computed as follows:
E · ET
varπ = diag((YT · Y)−1 ) (10) Ω= . (17)
m−p
where varπ is a (p × 1) vector and the diag(·) is an operator that For convenience, we formulate the original linear regression
obtains the diagonal elements of a square matrix. problem as normalized one by left-multiplying (5) with the
For better results, WLS is introduced [19]. First, residuals are inverse of Ω’s matrix square root, i.e.,
computed, i.e.,
Ω− 2 · τ = Ω− 2 · Y (q, q̇, q̈) · π + Ω− 2 · ε
1 1 1
(18)
−1
R = T − Y · π = {I − Y · (Y · Y) T
·Y }·T
T
(11)
with a little abuse of notation, the above equation can be con-
where R is the (mn × 1) residual, which can be reshaped as a cisely expressed as follows:
(n × m) matrix called reshaped residual E here.
Second, the variance of the error vector is estimated as τ ∗ = Y ∗ (q, q̇, q̈) · π + ε∗
follows: where τ ∗ is the normalized response vector, Y ∗ (q, q̇, q̈) is
||Ej ||22 the normalized regressor matrix, and ε∗ is the normalized error
Λjj = (12)
m−p vector. Similarly, the normalized observation matrix and the

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
1580 IEEE TRANSACTIONS ON ROBOTICS, VOL. 36, NO. 5, OCTOBER 2020

normalized observation response can be computed as follows:

⎡ ∗ ⎤
Y 1
⎢ ∗ ⎥
⎢ Y 2⎥
Y∗ = ⎢ ⎥
⎢ .. ⎥ , (19)
⎣ . ⎦
Y ∗m

T∗ = [τ ∗T
1 τ ∗T
2 . . . τ ∗T
m] .
T
(20)

B. The IRLS Method

It is well-known that LS or WLS estimators are vulnera-
ble to outliers and influential points, which are due to model
Fig. 1. Chosen weight function Ψ(·).
uncertainties, vibrations/measurement noise in dynamic model
identification. In this article, to deal with this problem, we resort
to a robust IRLS method [21], which was previously introduced exactly coincides with the LS solution, i.e.,
in robot dynamic identification by Janot et al. [25]. π = (Y#T · Y# )−1 · Y#T · T# , (24)
The IRLS generally puts additional weights on data points so
that the outliers contribute less to the final results. Assuming that varπ = diag((Y#T · Y# )−1 ) (25)
we have the weight matrix W of size (mn × 1) and its extended
matrix We whose size is (mn × p) with each column being W , where π = [π Tb π Tf ]T . Due to the existence of model uncertain-
applying W and We to T∗ and Y∗ respectively, we get ties and vibrations/measurement noise, however, the solution of
problem (23) is often different from the LS solution but is more
Y# = We · ∗Y∗ , (21) robust to model uncertainties and vibrations/measurement noise
∗ (or overfitting mitigation as mentioned in [34]).
T = W · ∗T
#
(22)
When the solution is ready, the normalized residual can be
where ·∗ is the elementwise multiplication operator. computed as follows:
Then, the following optimization problem can be formulated:
R# = T# − Y# · π (26)
arg(π b , π f , π d ) min(||Y · #
[π Tb π Tf ]T − T ||2 )
#
and according to (16), the error vector’s covariance matrix can
−1 be updated as follows:
subject to : πs = G · [π Tb π Td ]T
1 1
π f (F vj ) > 0 Ω 2 · E# · E#T · Ωold
2

Ωnew = old (27)

π f (F cj ) > 0 m−p

π s (Iaj ) > 0 where E# is the (n × m) reshaped normalized residual. The

⎡ ⎤ process how Ω is updated is described in the inner loop (L1 ) of
tr(π s (Lj )) Fig. 5.
⎣ I − π s (Lj ) π s (lj ) ⎦
2 >0 With the normalized residual R# as inputs, the weight matrix
π s (lj )T π s (Mj ) W can be updated according to a weight function Ψ(·), i.e.,
j = 1, 2, 3, . . . , n, (23) Wnew = Min(Wold , Ψ(R# )) (28)
where π b denotes the inertial base parameters, π f denotes where Min(·) returns the smaller number of its two operands
the friction parameters, π d denotes the nonbase parameters, elementwise. The Min(·) is necessary because it guarantees
π s denotes the standard inertial parameters, and the matrix G convergence of W . Despite the lack of rigorous proofs, an
defines a linear bijective mapping between π s and [π Tb π Td ]T intuitive explanation for this choice is that the new weight on the
[33]. In our experiments, G can be derived from Appendix. In same data point should not be larger than the previous weight
the constraints, we also define abstract operators like π f (·) or (or a previous outlier should not be considered as an inlier now).
π s (·), which return results in a similar pattern to their operands The process how W is updated is described in the middle loop
but with the corresponding numerical values contained in π f or (denoted as L2 ) of Fig. 5.
π s . For instance, π s (lj ) returns lj but with numerical values There are many kinds of weight functions as reported in [21].
whose components (like M Xj , M Yj and M Zj ) are contained Instead of the Huber weight function that is used in the work [25],
in π s . a T class hard redescender is chosen to be the weight func-
The above convex optimization problem is readily solved tion in this article, as shown in Fig. 1. The main reason for
by CVX [50], [51] in the semidefinite programming mode. In such a choice is that the measurement noise is supposed to be
the ideal case, i.e., there is no model uncertainties and vibra- normally distributed (refer to Fig. 2) and we want to filter out
tions/measurement noise, all the constraints in problem (23) outliers that do not fit in the model (refer to Fig. 3) while the
should be inactive, which means the solution of problem (23) Huber weight function is meant for cases where the normally

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
HAN et al.: ITERATIVE APPROACH FOR ACCURATE DYNAMIC MODEL IDENTIFICATION OF INDUSTRIAL ROBOTS 1581

Fig. 2. Different choices of the threshold. (a) Good choice of the threshold 3. (b) Bad choice of the threshold 3.5.

Fig. 3. Experimental example to show outliers. (a) Joint torque and residual. (b) ∼ W and joint velocity.

distributed noise is contaminated with another unknown distri- 5.3 s, respectively. Accordingly, outliers are observed around
bution [52]. Though the detected outliers corresponding to that these positions in Fig. 3(b), where ∼ W is the opposite of W
vibrations/measurement noise (refer to Fig. 3) can be thought elementwise. After careful analysis, we conclude that outliers
of an unknown distribution, most outliers are due to model around 2.5 and 8.8 s are due to vibrations/measurement noise,
uncertainties (refer to Fig. 12; outliers concentrate at the low while outliers around 1.2 and 5.3 s are due to the model uncer-
joint speed). With the chosen weight function, we try to accent tainties of frictions. The above example also justifies the choice
data with large joint velocity while ignore data around zero joint of the T class hard redescender as the weight function. With
velocity where friction is poorly modeled. Correspondingly, IRLS, therefore, only the data points, which fit in the model
the identified dynamic model behaves poorly around zero joint and have low vibrations/measurement noise, contribute to the
velocity while behaves excellently elsewhere. identification results.
The reason why a threshold of 3 is chosen here is that the Remark 1: It is necessary to point out that IRLS and the
calculated normalized residual should follow the normal distri- aforementioned WLS are quite different. IRLS puts weights on
bution after weighting the data points. As shown in Fig. 2, the the data points, while the aforementioned WLS puts weights on
normal probability plot presents straight lines with unit slope measurement channels or joints. The weight function chosen
within the threshold when a proper threshold is set while there here is the simplest and the most straightforward, but other
could be “heavy tails” when an improper threshold is set. What weight functions (e.g., Huber weight function) may also apply
is more, the threshold should be as large as possible so that the with little changes within the framework.
number of outliers is kept to the smallest. According to [21], a
threshold of 2.795 could keep 95% of the data points.
In order to show which data points are assigned with zero C. Friction Estimation and Modeling
weight (or treated as outliers), an experimental example is given, Normally, Coulomb’s friction plus linear viscous friction
as shown in Fig. 3. The measured torque and the calculated model is adopted for the dynamic model identification. In this
residual is shown in Fig. 3(a), from which large measurement research, however, we find that the traditional friction model
noise/vibrations are observed around 2.5 and 8.8 s, respectively, accounts for the major nonlinearity between the residual and the
and abrupt changes in the torque are observed around 1.2 and predicted response (refer to Fig. 14). Hence, we should estimate

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
1582 IEEE TRANSACTIONS ON ROBOTICS, VOL. 36, NO. 5, OCTOBER 2020

and model the friction with high accuracy to improve dynamic

model identification results.
One way to estimate the friction is to set the friction param-
eters to zero after identifying π and then subtract the torques
related to π i from the measured ones, i.e.,
Fest = T − Y · [π Tb 0T ]T . (29)
Then, more accurate friction model can be utilized to fit (29).
In this article, we adopt the following nonlinear friction model
(for jth joint), which can fit the estimated friction well in our
experiments except for the data whose joint velocity is low
fj = (F cj + F vj · |q̇j |αj ) · sign(q̇j ) + Bj
Fig. 4. Friction estimation and modeling.
= [sign(q̇j ) |q̇j |αj · sign(q̇j ) 1] · [F cj F vj Bj ]T .
(30)
Note that such a friction model is also used in Farhat’s work [44],
where the dynamic model identification is treated as a nonlin-
ear optimization problem. This friction model differs from the
traditional one (2) in that the viscous friction is not linear to the
joint velocity but linear to some unknown αj power of the joint
velocity. Therefore, it is a generalization of the traditional one,
i.e., the nonlinear friction model degenerates into the traditional
one when the parameter αj is exactly one. Whenever αj is
known, moreover, by grouping [sign(q̇j ) |q̇j |αj · sign(q̇j ) 1]
into Y and [F cj F vj Bj ]T into π, the chosen friction model
is again linear with respect to F cj , F v j , and Bj , and the linear
regression can be applied with small changes. However, it should
be noted that the regressor matrix Y is also a function of α now.
A practical approach to estimate and update αj is to fit the
chosen friction model to the estimated friction, using the interior
point method [53] with F cj , F vj , Bj in the already identified
π f and the previous αj as a good start point, i.e.,
arg(αj , F cj , F vj , Bj ) min(||fj − Fj ||2 )
subject to : F cj > 0
F vj > 0
αj > 0
j = 1, 2, 3, . . . , n (31)
where fj is the friction model described in (30) and Fj is the
estimated friction derived from Fest calculated by (29). In spite of
the nonconvexity, the globally optimal solution can be achieved
due to the good choice of a start point.
One such example is shown in Fig. 4. After the α is estimated,
new observation matrix can be computed using the updated α.
This process, which is described in the outer loop (denoted as
L3 ) in Fig. 5, will repeat until the nonlinear friction parameter
α converges.
Remark 2: This section actually presents a general method
for estimating and modeling frictions and the nonlinearity of
friction models is treated in a unified way with linear regressions.
This method applies in the presence of a more complex friction
model, which can be divided into linear parts and nonlinear
parts. The linear parts are included in base parameters, while Fig. 5. Overall approach for dynamic model identification.

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
HAN et al.: ITERATIVE APPROACH FOR ACCURATE DYNAMIC MODEL IDENTIFICATION OF INDUSTRIAL ROBOTS 1583

the nonlinear parts are contained in the regressor matrix. The is chosen as T class redescender due to its explicit explanation
parameters of the nonlinear parts can be estimated by fitting the (refer to Fig. 3) and sound statistical meaning (refer to Fig. 2). As
friction model to the estimated friction [refer to (29)] and used iterations go on within this loop, the weight matrix W converges
to compute the new regressor matrix until the nonlinear friction to a constant matrix and only data points that fit in the friction
parameters converge. model and have low vibrations/measurement noise will reside.
Thus, the solution π is more robust to model uncertainties and
D. Residual Analysis vibrations/measurement noise. At each iteration of L2 , the inner
loop has to be re-evaluated due to the changes in the weight
The major assumptions that we have made thus far in the
matrix W .
linear regression model, i.e. (5), are as follows.
In the outer loop (denoted as L3 ), nonlinear friction param-
1) The relationship between the response vector τ and the
eters are updated by solving (31). As iterations go on within
regressor matrix Y is linear, at least approximately.
this loop, nonlinear friction parameters approximate to their
2) The error vector ε has zero mean in each channel.
true values. Therefore, frictions get properly modeled (refer to
3) The error vector ε has a constant covariance matrix Ω.
Fig. 12) and inertial parameters can be estimated with higher
4) The residual is serially uncorrelated, i.e., each row of the
precision (refer to Fig. 14). Since nonlinear friction parameters
reshaped residual E is serially uncorrelated.
are contained in the regressor matrix Y , the observation matrix
5) The residual is normally distributed, i.e., each row of the
Y has to be recomputed at each iteration within this loop. As
reshaped residual E is normally distributed.
a result, the middle loop L2 and the inner loop L1 have to be
Assumptions 3 and 4 imply that the error vector ε is an
re-evaluated as well.
independent random vector. Assumption 5 is required for the
Instead of merging L1 and L3 into a nonlinear optimization
hypothesis testing and interval estimation.
problem like the work [44], the separation of L1 and L3 is more
To ensure a faithful and accurate identification, we should
art than technique. First, it guarantees the problem (23) is convex
always consider the validity for these assumptions. Residual
since the nonlinear parameter α is used as a priori instead of a
analysis, especially the graphical analysis of residuals, gives us
parameter to be optimized. This is important because convexity
a very effective way to investigate the adequacy of the fit of a
of a problem is a prerequisite for the CVX solver. Second, the
regression model and check the underlying assumptions [54]. In
problem can be solved faster by scaling the original big nonlinear
this article, the following residual plots are adopted.
optimization into smaller ones like the problem (31). Since the
1) Normal probability plot of E# is used to dynamically show
nonlinear parameter α is used as a priori in problem (23), the
how the IRLS removes outliers and regulates the residuals
weight matrix W is initialized to ones every time α gets updated.
(refer to Fig. 9). It also reveals the normal distribution of
Thus, the uncertainty of α is ignored (i.e., we assume there is
the residuals.
no uncertainty in α).
2) Auto-correlation plot of E# is used to check the serial
Remark 3: The proposed iterative approach in fact provides
independence assumption (refer to Fig. 13).
a general framework for dynamic model identification. It still
3) Plot of E against the fitted response is used to check
applies even if a different weight function or friction model is
linearity of the model and the constant covariance of
adopted.
measurement noise (refer to Fig. 14).
4) Plot of E against joint velocities is used to estimate friction
(with π f being zero; refer to Fig. 12) and check the outliers IV. EXPERIMENTS
(refer to Fig. 10). A. Experimental Setup
The robot we used is a Foxconn robot, Foxbot A-05-2, which
E. The Overall Approach
is an industrial robot with six revolute joints. As shown in Fig. 6,
Fig. 5 demonstrates the overall approach for dynamic model the modified Denavit–Hartenberg (DH) convention is utilized to
identification, which consists of three loops. build the kinematic model of the Foxbot and the modified DH
In the inner loop (denoted as L1 ), (23) is solved to get the parameters are listed in Table I.
identified base parameters, with which normalized residual R# Experiments are conducted on the dSPACE MicroLabBox
can be computed by (26) and the covariance matrix Ω can platform, which is an all-in-one development system for labo-
be updated by (27). Initially, Ω is an identity matrix and the ratory with high performance and versatility. The original con-
solution π is close to the solution of ordinary least squares. troller of Foxbot is removed. Moreover, all six motor drivers are
As iterations go on within this loop, Ω converges to a constant configured as torque control mode and rewired to the MicroLab-
matrix whose diagonal elements denote the estimated variance Box. Fig. 7 shows the overall setup. The host computer is running
of the corresponding joint/channel and the solution π is close MATLAB/Simulink to generate code for the MicroLabBox.
the solution of WLS. The LMI constraints on base parameters After compiling, the MicroLabBox implements the code and
are also implemented with this loop so that the solution π is outputs control signal to the motor drivers via digital to analog
always physically feasible. converters. Meanwhile, it receives the encoder signal from the
In the middle loop (denoted as L2 ), the weight matrix W is motor drivers to form a closed loop. Some digital inputs/outputs
updated by (28), according to the normalized residual R# with and switches are also included for dedicated operations like
chosen weight function Ψ(·). In our case, the weight function servo ON/OFF. The sample time for the overall control system

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
1584 IEEE TRANSACTIONS ON ROBOTICS, VOL. 36, NO. 5, OCTOBER 2020

matrix can be well conditioned. In this article, each joint of the

robot tracks a series of sine curves, whose frequencies are 0.1,
0.2, and 0.4 Hz. Thus, joint trajectories can be characterized as
follows:

3
qj = [μji · sin(ωi · t) + νij · cos(ωi · t)] + δ j (32)
i=1

where the superscript j denotes the joint number, while the

subscript i denotes the frequency number. Since each joint
trajectory has seven parameters, there are 42 parameters for
all the joints. Here, we use β, a 42 × 1 vector, to stand for
these parameters, with β j = [μj1 μj2 μj3 ν1j ν2j ν3j δ j ]T being the
parameters for joint j. Trajectories of a full period of 10 s are
discretized into 100 points.
Normally, condition number of Y is chosen as the criterion of
exciting trajectories optimization. As shown in [17], however,
this is not enough. As a result, condition numbers of subregres-
sors are used to regulate the regressor matrix [12].
Fig. 6. Modified DH model of the Foxbot A-05-2. When WLS is applied, the following inequality stands:
cond(Y ∗ ) = cond(Ω− 2 · Y ) ≤ cond(Ω− 2 ) · cond(Y ). (33)
TABLE I 1 1

MODIFIED DH PARAMETERS OF FOXBOT A-05-2

Although an optimal condition number of Y sets an upper
bound on that of Y∗ , it would be interesting to perform opti-
mization on Y∗ directly by including priori knowledge of the
measurement noise Ω0 , which can be roughly estimated from
the torque noises (at joint side) of all joints. Invoked by [12],
moreover, optimization should be performed iteratively so that
optimal motions at the present step could complement previous
motions. Consequently, the following optimization problem is
formulated:
arg(β) min(10 · cond(Y∗ (β)) + 5 · cond(Y∗i (β))
+ cond(Y∗g (β)) + cond(Y∗f (β)))

3
subject to : ωi · μji = 0
i=1


3
μji 2 + νij 2 + δ j ≤ qjmax
i=1


3
− μji 2 + νij 2 + δ j ≥ qjmin
i=1


3
ωi · μji 2 + νij 2 ≤ q̇jmax
Fig. 7. Experimental setup. i=1


3

is 1 ms. The position controller implemented in the MicroLab- ωi2 · μji 2 + νij 2 ≤ q̈jmax
i=1
Box is the traditional proportional-integral differential controller
with fine-tuned gains. During the experiment, an additional load j = 1, 2, 3, 4, 5, 6, (34)
is calibrated and used to compare identified loads and their ∗
where Y is the normalized observation matrix computed with
theoretical counterparts.
Ω0 , Y∗i is the subregressor of Y∗ corresponding to the inertial
components (links’ inertia and rotors’ inertia) of π, Y∗g is the sub-
B. Data Acquisition
regressor of Y∗ corresponding to the gravity components (links’
It is important to choose appropriate trajectories so that the first moment) of π, and Y∗f is the subregressor of Y∗ correspond-
regressor matrix can be persistently excited and the observation ing to the friction components of π. Different from works [12]

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
HAN et al.: ITERATIVE APPROACH FOR ACCURATE DYNAMIC MODEL IDENTIFICATION OF INDUSTRIAL ROBOTS 1585

TABLE II
COMPARISONS OF DIFFERENT CRITERIA FOR TRAJECTORIES OPTIMIZATION1∗

1∗
The friction parameters αj of all joints are set to 1 during the optimization.
2∗
cond(Y).
3∗
10 · cond(Y) + 5 · cond(Yi ) + cond(Yg ) + cond(Yf ).
4∗
10 · cond(Y∗ ) + 5 · cond(Y∗i ) + cond(Y∗g ) + cond(Y∗f ).

and [17], where the subregressors are designed according to TABLE III
MOTOR GAINS AND TRANSMISSION RATIOS
link segments, the subregressors here are designed according to
the type of physical quantities so that parameter scaling can be
decoupled in a similar way. The choice of coefficients among
the subregressors are based on heuristics by observing their
condition numbers. In addition, the first constraint means each
joint starts at zero velocity; the second and third constraints put
upper and lower bounds on the angle of each joint; the fourth
a fourth-order Butterworth filter with a cutoff frequency of
constraint puts upper bounds on the velocity of each joint; and
6 Hz, 15 times of the highest frequency (0.4 Hz) of exciting
the fifth constraint puts upper bounds on the acceleration of each
trajectories, is chosen to filter the position data back and forth
joint.
(to keep zero-lag); then, the filtered position is processed by
Due to the iterative nature, Y∗ has the following structure:
a central difference filter to obtain joint velocities and accel-
At the first iteration, Y∗ is of size (600 × 58); at the second
erations; afterwards, all joint motion data (position, velocity,
iteration, Y∗ is of size (1200 × 58), i.e., the last half is to be
and acceleration) are downsampled by a scale of 10 and only
determined at the present iteration, while the first half has already
one full period of motion data is reserved for identification or
been determined at the previous iteration; at k iteration, Y∗ is of
validation. The commanded currents/torques are converted to
size (600 k × 58), i.e., the last 600 rows are to be determined at
joint side torques according to Table III. Then, one full period
the present iteration, while others have already been determined
of torque data is reserved in a similar way after downsampling.
at previous iterations. During each iteration, the optimization
runs at a random selected start point. If the cost is decreased, we
accept this iteration. Otherwise, we deny this iteration. If the cost C. Data Processing
cannot be decreased after ten trials, we stop the optimization and The collected data are processed according to the procedure
consider that a globally optimal solution is achieved. Therefore, shown in Fig. 5. In this section, we will show how the covariance
a successful optimization could take half an hour1 by the interior matrix Ω converges in the loop L1 , how the weight matrix W
point method [53]. converges in the loop L2 , how the friction model parameter α
Comparisons of different criteria for trajectories optimization converges in loop L3 , and how the residual analysis is applied
are shown in Table II. From the table, it can be observed that to checking the assumptions and model adequacy.
all criteria achieve similar condition numbers of Y at the final Fig. 8 shows the convergence process of these three loops.
iteration. However, due to the fact that criterion 2 and criterion 3 In Fig. 8(a), the condition number of Ω is logged during the
give extra regulation on the subregressors, the regressor matrix execution L1 for one loop. It shows that the covariance matrix Ω
is more equilibrated. Furthermore, since criterion 3 includes the converges after ten iterations. In Fig. 8(b), the sum of the weight
priori knowledge of the measurement noise, i.e. Ω0 , it gives the matrix W is logged during the execution of L2 for one loop.
best performance on Y∗ . Since the chosen weight function is a T class hard redescender,
Nine motions are experimented to collect data, including the the sum of W simply indicates the number of inliers. Therefore,
five motions of criterion 3 for identification and the four motions Fig. 8(b) tells us that outliers are iteratively zero weighted against
of criterion 2 for validation. The experiments are conducted iterations. In Fig. 8(c), the sum of the weight matrix at the end
within an hour with fixed intervals to escape the influence of of L2 is logged during the execution of L3 for one loop. It
temperature. Meanwhile, commanded currents/torques of mo- shows that more data points are taken as inliers as the friction
tors and joint positions of motors’ encoders are recorded. Since parameter α is updated. Depending on the stop criteria of the
computation of Y requires joint velocities and accelerations, three loops, (23) may be solved for hundreds of times to get the
final results. In the shown case, (23) is solved for hundreds of
1 It was tested on Window 10 with processor Intel(R) Core(M) i7-6700HQ times, which takes a long time. If we resort to the LS solution,
CPU @ 2.6 GHz and 16-GB RAM. i.e. (24), however, the execution time can be greatly reduced.

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
1586 IEEE TRANSACTIONS ON ROBOTICS, VOL. 36, NO. 5, OCTOBER 2020

Fig. 8. Explanations of the three loops of the identification approach. (a) Loop L1 (first iteration in L2 and L3 ). (b) Loop L2 (first iteration in L3 ). (c) Loop L3 .

Fig. 9. Normal probability plot of E# (first iteration in L3 ). (a) First iteration in L2 . (b) 15th iteration in L2 . (c) 14th iteration in L2 .

Fig. 10. Residual versus joint velocity of joint 4 (j4) (first iteration in L3 ). (a) First iteration in L2 . (b) 5th iteration in L2 . (c) 14th iteration in L2 .

Figs. 9 and 10 further explain how IRLS works. Fig. 9 is that it can produce more compatible estimation by grouping
the normal probability plot of E# , from which we can see that them into π than that by taking them as a priori.
the distribution of the residual is not normal at the beginning, Fig. 13 checks the serial uncorrelation assumption on residu-
i.e., there are too many “heavy tails.” After several iterations, als. In Fig. 13, auto-correlation of residuals of all joints are less
however, the “heavy tails” are removed and the normalized than a threshold of 0.15, which reveals that the residuals can be
residual is clamped between negative three and positive three taken as serially uncorrelated. It also displays some periodicity,
due to the chosen weight function (refer to Fig. 1). At a glimpse which could be due to the periodicity of exciting motions.
of Fig. 10(a), we can see that large residuals are concentrating Plot of E against the fitted response is used to check the model
around the zero velocity. This is because these data points do not adequacy, as shown in Fig. 14. The residual E is computed by
fit in the nonlinear friction model [refer to (30)]. Therefore, these (11) and the fitted response or predicted response is computed by
data points are considered as outliers and iteratively removed, Y · π. When π f is set to zero, only the dynamic forces/torques
as shown in Fig. 10(b) and (c). without friction are computed, while all the forces/torques are
Fig. 11 shows the convergence of friction model parameter computed otherwise. From Fig. 14(a) and (c), we can conclude
α and Fig. 12 shows the final estimation of the nonlinear that the dynamic model is linear with respect to the inertial
friction parameter α and the fitting results. Although the friction parameters, i.e. π i , since the residual against the fitted response
parameters F cj , F vj , and Bj can also be estimated when fitting without friction zero is contained in a horizontal band, which
the nonlinear friction model (30) to the estimated friction, only is an ideal situation. Comparing Fig. 14(a) and (b), we can
αj is taken as a priori to compute the observation matrix Y when conclude that the traditional friction model [refer to (2)] is not
performing the linear regression. Note that the nonlinear friction adequate since the residual against the fitted response with the
model is linear with respect to F cj , F vj , and Bj which means traditional friction model indicates nonlinearity, which means

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
HAN et al.: ITERATIVE APPROACH FOR ACCURATE DYNAMIC MODEL IDENTIFICATION OF INDUSTRIAL ROBOTS 1587

Fig. 11. Convergence of α in L3 . (a) Joints 1–3. (b) Joints 4–6.

Fig. 12. Final estimation of α in L3 . (a) Joint 1. (b) Joint 2. (c) Joint 3. (d) Joint 4. (e) Joint 5. (f) Joint 6.

Fig. 13. Serial uncorrelation assumption check of E∗ . (a) Joint 1. (b) Joint 2. (c) Joint 3. (d) Joint 4. (e) Joint 5. and (f) Joint 6.

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
1588 IEEE TRANSACTIONS ON ROBOTICS, VOL. 36, NO. 5, OCTOBER 2020

Fig. 14. Model accuracy check. (a) and (b) Traditional friction model (2). (c) and (d) Nonlinear friction model (30).

that other regressor variables are needed in the model. It also TABLE IV
RMS OF RESIDUALS IN THE IDENTIFICATION1∗
indicates that improper friction models make major contribu-
tions to the nonlinearity between the residual and the predicted
response. Comparing Fig. 14(c) and (d), we can conclude that
the nonlinear friction model [refer to (30)] is appropriate for
this robot system since the residual against the fitted response
with the nonlinear friction model is contained in a horizontal
band. Comparing Fig. 14(a) and (b) with Fig. 14(c) and (d), we 1∗
The size of Y is 30 000 × 58 for 5000 selected samples from trajectories optimized
can conclude that the variance of the residual is significantly re- by criterion 3.
duced by using the nonlinear friction model. The above analysis
reveals that improper friction models make major contributions TABLE V
RMS OF RESIDUALS IN THE IDENTIFICATION AFTER EXCLUDING
to the nonlinearity of residuals (against fitted response) and an LOW VELOCITY SAMPLES1∗
inadequate friction model will inevitably increase the variance
of residuals.

D. Experiment Results Analysis

In this section, experiment results are summarized and com-
pared. Four different combinations under the proposed frame- 1∗
Only residuals of which the absolute joint velocity is greater than 0.05 rad/s are
work are applied to identify the dynamic model parameters, i.e., collected.
Method 1 uses linear friction model (2), the proposed weight
function (Fig. 1) and includes no constraints (24); Method 2
reveals the influence of the constraints. The performance and
uses nonlinear friction model (30), the proposed weight function
computational cost of these four combinations are provided to
(Fig. 1) and includes no constraints (24); Method 3 uses non-
offer a best choice for users.
linear friction model (30), the Huber weight function2 [25] and
Identification results of the four combinations are listed in
includes no constraints (24); Method 4 uses nonlinear friction
Table VI. Root mean squares (rms) of residuals in the identi-
model (30), the proposed weight function (Fig. 1) and includes
fication in Tables IV and V where samples at low velocity are
the constraints (23). These four combinations are representative
excluded. Comparing Method 1 with Method 2 in Tables IV
because comparison between Method 1 and Method 2 reveals
and V, we can find that the Method 2 is better than Method 1
the influence of different friction models, comparison between
since the rms of Method 2 is smaller than that of Method 1,
Method 2 and Method 3 reveals the influence of different weight
thanks to the advantage of the nonlinear friction model over
functions, and comparison between Method 2 and Method 4
the linear friction model. Comparing Method 2 with Method
3 in Table IV, one may conclude that Method 3 is better than
2 The H parameters is fixed at the most common choice 1.345. Method 2 since the rms of Method 3 is smaller than that of

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
HAN et al.: ITERATIVE APPROACH FOR ACCURATE DYNAMIC MODEL IDENTIFICATION OF INDUSTRIAL ROBOTS 1589

TABLE VI
IDENTIFICATION RESULTS1∗

1∗
All the units are in SI unit.
The size of Y is 30 000 × 58 for 5000 selected samples from trajectories optimized by criterion 3.

Method 2 for all joints except joint 2. However, after excluding similar rms are achieved, which is not surprising since the
samples at low velocity (where friction is poorly modeled) as identified values of the two methods are similar, as listed in
listed in Table V, we can find that Method 2 is actually better Table VI. However, the identified values with Method 4 are
than Method 3 for all joints. Comparing Method 2 with Method guaranteed to be physically feasible due to the constraints. It
4 in Tables IV and V, the performances are quite close since is also interesting to note that the parameters with different

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
1590 IEEE TRANSACTIONS ON ROBOTICS, VOL. 36, NO. 5, OCTOBER 2020

Fig. 15. Cross validation (from the top to the bottom are joints 1–6).

signs of values identified by Method 2 and Method 4 (like π21 one since it does not have to implement loop L3 due to the linear
and π25 ) are very small in significance, which make them more friction model. Method 3 is faster than Method 2 and it turns out
vulnerable to model uncertainties and vibrations/measurement that different weight functions can affect convergence rate of
noise. the whole framework. Method 4 is the slowest one because it
The computation cost for the four combinations is listed in takes more time to solve problem (23) instead of the analytical
Table VII. From the table, we can tell Method 1 is the fastest solution, i.e., (24).
Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
HAN et al.: ITERATIVE APPROACH FOR ACCURATE DYNAMIC MODEL IDENTIFICATION OF INDUSTRIAL ROBOTS 1591

TABLE VII TABLE IX

COMPUTATION COST OF THE FOUR COMBINATIONS IN THE IDENTIFICATION1∗ RMS OF RESIDUALS IN THE CROSS VALIDATION AFTER EXCLUDING
LOW VELOCITY SAMPLES1∗

1∗
The size of Y is 30 000 × 58 for 5000 selected samples from trajectories optimized
1∗
by criterion 3. Only residuals of which the absolute joint velocity is greater than 0.05 rad/s are
2∗
They were tested on Window 10 with processor Intel(R) Core(M) i7-6700HQ CPU collected.
@ 2.6 GHz and 16-GB RAM. Results are round off to seconds.
3∗
The stop criterion is set to 10−7 , applying to the largest element in abs(Ωnew −
Ωold ). excluding low velocity samples, as listed in Table IX. Now, even
4∗
The stop criterion is set to 10−5 , applying to the largest element in abs(Wnew −
Wold ). for joints 4 and 6, Method 2 and Method 4 outperform Method
5∗
The stop criterion is set to 10−3 , applying to the largest element in abs(αnew − 3, which justifies our choice of the weight function.
αold ). The above results indicate that, within the proposed frame-
work, 1) a proper friction model is crucial for improving the
TABLE VIII
RMS OF RESIDUALS IN THE CROSS VALIDATION1∗
accuracy of identification results, 2) different weight functions
can affect the convergence rate of the whole framework, while
Huber weight function compromises in accuracy for better con-
vergence rate, and 3) taking computation cost and performance
into account, Method 2 is considered to be the best among
the four investigated combinations, and, in practice, it can be
combined with BPFC or implement (23) in the final iteration to
1∗
The size of Y is 24 000 × 58 for 4000 selected samples from trajectories optimized guarantee physical feasibility property of inertial parameters.
by criterion 2.

V. CONCLUSION
In order to compare the performance of the identified dynamic In this article, an iterative approach of dynamic model identi-
model parameters, cross validation is conducted, as shown in fication was proposed, which is applicable for different weight
Fig. 15. In the figure, the gray solid lines represent the measured functions or friction models. Within the framework, robustness
joint torques; the black solid lines represent joint velocities; to model uncertainties and physical feasibility of model pa-
the red solid lines denote the residual using identified values rameters were simultaneously addressed by IRLS with LMI
by Method 1; the blue dot-dashed lines denote the residual constraints. Residual analysis showed that traditional friction
using identified values by Method 2; the magenta dotted lines model is not adequate (refer to Fig. 14), hence the need of a more
denote the residual using identified values by Method 3; and accurate nonlinear friction model, which could be a problem for
the green dashed lines denote residual using identified values WLS. In this article, however, nonlinearity of the friction model
by Method 4. From this figure, we can visually find that the was treated in a unified way with linear regression by splitting off
residuals of the Methods 2, 3, and 4 are generally smaller than the nonlinear parameter into the regressor matrix and updating
that of Method 1, which reveals that a proper choice of friction it in an iterative manner. Four combinations under the proposed
models is crucial for accurate dynamic model identification. framework were investigated to show, respectively, the influence
Large errors present at low speed (like joint 4 at 30 s) due to the of friction models, weight functions, and physical feasibility
fact that both friction models (2) and (30) lose accuracy under constraints on the identification results. Our research also re-
low speed. It is interesting to note that residual 2 nearly coincides vealed that performance can be improved by including priori
with residual 4, which means the performance of Method 2 is knowledge of measurement noise in the optimization of excit-
almost the same as Method 4. Since Method 4 takes several ing trajectories. Experimental results showed that the proposed
hours to get the solution while Method 2 only takes less than approach is robust and accurate, given a proper friction model
3 min, a compromise can be made by first using Method 2 and weight function. Since friction is dependent on temperature
to get the solution and then correcting the solution with the and load [47], a particular application of the proposed approach
physical feasibility constraints [also known for base parameters was to investigate such dependencies of friction models.
feasibility correction (BPFC) [33]] or just solve problem (23) Experimental results showed that large prediction errors
in the final iteration, before which the analytic solution (24) is present at low speed (refer to Fig. 15). One may try to model
adopted. For more detailed comparisons, the rms of the four the Stribeck effect in a friction model to improve accuracy.
residuals are shown in Table VIII, from which we can check However, Stribeck curves are commonly obtained by sampling
that rms of Methods 2, 3, and 4 are smaller than that of Method some constant speed where frictions are operating in the sliding
1. RMS of Method 2 and Method 4 are quite similar and they mode, while in dynamic model identification, frictions may
are smaller than that of Method 3 for all joints except for joints interchange frequently between the presliding mode and the
4 and 6. Since both the linear and nonlinear friction models lose sliding mode, which would mess up the Stribeck curve, as shown
accuracy under low speed, we also collect rms of residuals after by the outliers at low speed in Fig. 12. Therefore, it would be
Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
1592 IEEE TRANSACTIONS ON ROBOTICS, VOL. 36, NO. 5, OCTOBER 2020

interesting to include a dynamic friction model such as a LuGre π21 = ZZ5 + Y Y6 , (42)
model [55] or a GMS model [56] into our framework, which
could be our future work. π22 = XX6 − Y Y6 , (43)
When priori knowledge of inertial parameters is available, it π23 = XY6 , (44)
can appear in the objective function of the optimization problem
(23) as a regularizer, which penalizes the distance between iden- π24 = XZ6 , (45)
tified inertial parameters and priori inertial parameters. When a π25 = Y Z6 , (46)
proper metric is defined, the regularizer would draw the distri-
bution of identified inertial parameters toward the distribution π26 = ZZ6 , (47)
of the priori inertial parameters, which will be beneficial for 
6
complex systems like humanoids and multilegged robots where π27 = M X2 + Mj · a 3 , (48)
measurement data could be limited and parameter excitability j=3
could be poor [57].
π28 = M Y2 ,
APPENDIX 
6

THE SYMBOLIC EXPRESSIONS OF BASE PARAMETERS π29 = M X3 + Mj · a 4 , (49)

OF FOXBOT A-05-2 j=4


6
π30 = M Y3 + M Z4 + Mj · d 4 , (50)

6 
6 
6
j=4
π1 = Mj · a22 + Mj · a23 + Mj · a24
j=2 j=3 j=4 π31 = M X4 , π32 = M Y4 − M Z5 , π33 = M X5 ,

3 π34 = M Y5 + M Z6 , (51)
+ Y Yj + Ia1 + ZZ1 , (35)
j=2
π35 = M X6 , (52)
π36 = M Y6 , (53)

6
π2 = − Mj · a23 + XX2 − Y Y2 , (36) π37 = Ia3 , π38 = Ia4 , π39 = Ia5 , π40 = Ia6 ,
j=3
π41 = F v1 , π42 =; F v2 , π43 = F v3 , π44 = F v4 ,
π3 = XY2 , π45 = F v5 , π46 = F v6 , π47 = F c1 , π48 = F c2 ,
π4 = XZ2 − M Z3 · a3 , π49 = F c3 , π50 = F c4 , π51 = F c5 , π52 = F c6 ,
π5 = Y Z 2 , π53 = B1 , π54 = B2 , π55 = B3 , π56 = B4 ,


6 π57 = B5 , π58 = B6 .
π6 = Mj · a23 + Ia2 + ZZ2 , (37)
j=3
In the above equations, a2 , a3 , a4 , and d4 are the modified DH
kinematic parameters of the Foxbot and their values are 0.075,

6
0.27, 0.11, and 0.306 m, respectively, as listed in Table I. Other
π7 = 2M Z4 · d4 + Mj · (d24 − a24 ); symbols are the standard dynamic parameters, as mentioned in
j=4
Section II.
+ XX3 − Y Y3 + Y Y4 , (38) Besides, the numbered equations are the base parameters that
are explicitly affected by payloads, while the unnumbered ones

6
are not.
π8 = XY3 − M Z4 · a4 − Mj · a 4 · d 4 , (39)
j=4
ACKNOWLEDGMENT
π9 = XZ3 , π10 = Y Z3 ,
The authors would like to thank the reviewers for their valu-
π11 = Y Y4 + ZZ3 + 2M Z4 · d4 able advice on the improvement of this article.

6
+ Mj · (a24 + d24 ), (40) REFERENCES
j=4
[1] J. Swevers, W. Verdonck, and J. D. Schutter, “Dynamic model identi-
π12 = XX4 − Y Y4 + Y Y5 , fication for industrial robots,” IEEE Control Syst. Mag., vol. 27, no. 5,
pp. 58–71, Oct. 2007.
π13 = XY4 , π14 = XZ4 , π15 = Y Z4 , [2] E. Magrini, F. Flacco, and A. De Luca, “Estimation of contact forces using
a virtual force sensor,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst.,
π16 = Y Y5 + ZZ4 , Sep. 2014, pp. 2126–2133.
[3] A. Albu-Schaffer, W. Bertleff, B. Rebele, B. Schafer, K. Landzettel, and
π17 = XX5 − Y Y5 + Y Y6 , G. Hirzinger, “Rokviss—Robotics component verification on ISS current
experimental results on parameter identification,” in Proc. IEEE Int. Conf.
π18 = XY5 , π19 = XZ5 , π20 = Y Z5 , (41) Robot. Autom., May 2006, pp. 3879–3885.

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
HAN et al.: ITERATIVE APPROACH FOR ACCURATE DYNAMIC MODEL IDENTIFICATION OF INDUSTRIAL ROBOTS 1593

[4] J. Wu, J. Wang, and Z. You, “An overview of dynamic parameter identifica- [29] M. Gautier, P. Vandanjon, and A. Janot, “Dynamic identification of a 6
tion of robots,” Robot. Comput.-Integr. Manuf., vol. 26, no. 5, pp. 414–419, DoF robot without joint position data,” in Proc. IEEE Int. Conf. Robot.
2010. Autom., May 2011, pp. 234–239.
[5] C. G. Atkeson, C. H. An, and J. M. Hollerbach, “Estimation of inertial [30] A. Janot, P. Vandanjon, and M. Gautier, “A generic instrumental variable
parameters of manipulator loads and links,” Int. J. Robot. Res., vol. 5, approach for industrial robot identification,” IEEE Trans. Control Syst.
no. 3, pp. 101–119, 1986. Technol., vol. 22, no. 1, pp. 132–145, Jan. 2014.
[6] M. Gautier and W. Khalil, “Direct calculation of minimum set of inertial [31] K. Yoshida and W. Khalil, “Verification of the positive definiteness of
parameters of serial robots,” IEEE Trans. Robot. Autom., vol. 6, no. 3, the inertial matrix of manipulators using base inertial parameters,” Int. J.
pp. 368–373, Jun. 1990. Robot. Res., vol. 19, no. 5, pp. 498–510, 2000.
[7] W. Khalil and F. Bennis, “Comments on ‘direct calculation of minimum [32] V. Mata, F. Benimeli, N. Farhat, and A. Valera, “Dynamic parameter
set of inertial parameters of serial robots’,” IEEE Trans. Robot. Autom., identification in industrial robots considering physical feasibility,” Adv.
vol. 10, no. 1, pp. 78–79, Feb. 1994. Robot., vol. 19, no. 1, pp. 101–119, 2005.
[8] W. Khalil and D. Creusot, “Symoro: A system for the symbolic modelling [33] C. D. Sousa and R. Cortesão, “Physical feasibility of robot base inertial
of robots,” Robotica, vol. 15, no. 2, pp. 153–161, 1997. parameter identification: A linear matrix inequality approach,” Int. J.
[9] B. Armstrong, “On finding exciting trajectories for identification exper- Robot. Res., vol. 33, no. 6, pp. 931–944, 2014.
iments involving systems with nonlinear dynamics,” Int. J. Robot. Res., [34] C. D. Sousa and R. Cortesão, “Inertia tensor properties in robot dynamics
vol. 8, no. 6, pp. 28–48, 1989. identification: A linear matrix inequality approach,” IEEE/ASME Trans.
[10] M. Gautier and W. Khalil, “Exciting trajectories for the identification of Mechatronics, vol. 24, no. 1, pp. 406–411, Feb. 2019.
base inertial parameters of robots,” Int. J. Robot. Res., vol. 11, no. 4, [35] P. M. Wensing, S. Kim, and J. E. Slotine, “Linear matrix inequalities
pp. 362–375, 1992. for physically consistent inertial parameter identification: A statistical
[11] J. Swevers, C. Ganseman, D. B. Tukel, J. de Schutter, and H. V. Brussel, perspective on the mass distribution,” IEEE Robot. Autom. Lett., vol. 3,
“Optimal robot excitation and identification,” IEEE Trans. Robot. Autom., no. 1, pp. 60–67, Jan. 2018.
vol. 13, no. 5, pp. 730–740, Oct. 1997. [36] T. Lee and F. C. Park, “A geometric algorithm for robust multibody inertial
[12] V. Bonnet, P. Fraisse, A. Crosnier, M. Gautier, A. González, and G. Venture, parameter identification,” IEEE Robot. Autom. Lett., vol. 3, no. 3, pp. 2455–
“Optimal exciting dance for identifying inertial parameters of an anthro- 2462, Jul. 2018.
pomorphic structure,” IEEE Trans. Robot., vol. 32, no. 4, pp. 823–836, [37] M. Gautier, S. Briot, and G. Venture, “Identification of consistent standard
Aug. 2016. dynamic parameters of industrial robots,” in Proc. IEEE/ASME Int. Conf.
[13] C. Presse and M. Gautier, “New criteria of exciting trajectories for robot Adv. Intell. Mechatronics, Jul. 2013, pp. 1429–1435.
identification,” in Proc. IEEE Int. Conf. Robot. Autom., vol. 3, May 1993, [38] C. Gaz, M. Cognetti, A. Oliva, P. R. Giordano, and A. De Luca, “Dynamic
pp. 907–912. identification of the Franka Emika Panda robot with retrieval of feasible
[14] W. Rackl, R. Lampariello, and G. Hirzinger, “Robot excitation trajectories parameters using penalty-based optimization,” IEEE Robot. Autom. Lett.,
for dynamic parameter estimation using optimized b-splines,” in Proc. vol. 4, no. 4, pp. 4147–4154, Oct. 2019.
IEEE Int. Conf. Robot. Autom., May 2012, pp. 2042–2047. [39] S. Traversaro, S. Brossette, A. Escande, and F. Nori, “Identification of fully
[15] J. Jin and N. Gans, “Parameter identification for industrial robots with a fast physical consistent inertial parameters using optimization on manifolds,”
and robust trajectory design approach,” Robot. Comput.-Integr. Manuf., in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Oct. 2016, pp. 5446–5451.
vol. 31, pp. 21–29, 2015. [40] H. Olsson, K. Å ström, C. C. de Wit, M. Gäfvert, and P. Lischinsky,
[16] K. Ayusawa, A. Rioux, E. Yoshida, G. Venture, and M. Gautier, “Friction models and friction compensation,” Eur. J. Control, vol. 4, no. 3,
“Generating persistently exciting trajectory based on condition num- pp. 176–195, 1998.
ber optimization,” in Proc. IEEE Int. Conf. Robot. Autom., May 2017, [41] P. Hamon, M. Gautier, P. Garrec, and A. Janot, “Dynamic modeling and
pp. 6518–6524. identification of joint drive with load-dependent friction model,” in Proc.
[17] G. Venture, K. Ayusawa, and Y. Nakamura, “A numerical method for IEEE/ASME Int. Conf. Adv. Intell. Mechatronics, Jul. 2010, pp. 902–907.
choosing motions with optimal excitation properties for identification of [42] P. Hamon, M. Gautier, and P. Garrec, “New dry friction model with load-
biped dynamics—An application to human,” in Proc. IEEE Int. Conf. and velocity-dependence and dynamic identification of multi-DoF robots,”
Robot. Autom., May 2009, pp. 1226–1231. in Proc. IEEE Int. Conf. Robot. Autom., May 2011, pp. 1077–1084.
[18] M. Gautier, “Dynamic identification of robots with power model,” in Proc. [43] N. Kammerer and P. Garrec, “Dry friction modeling in dynamic identi-
Int. Conf. Robot. Autom., vol. 3, Apr. 1997, pp. 1922–1927. fication for robot manipulators: Theory and experiments,” in Proc. IEEE
[19] M. Gautier and P. Poignet, “Extended Kalman filtering and weighted least Int. Conf. Mechatronics, Feb. 2013, pp. 422–429.
squares dynamic identification of robot,” Control Eng. Practice, vol. 9, [44] N. Farhat, V. Mata, A. Page, and F. Valero, “Identification of dynamic
no. 12, pp. 1361–1372, 2001. parameters of a 3-DoF RPS parallel manipulator,” Mechanism Mach.
[20] P. J. Huber, “The 1972 Wald lecture robust statistics: A review,” Ann. Theory, vol. 43, no. 1, pp. 1–17, 2008.
Math. Statist., vol. 43, no. 4, pp. 1041–1067, 1972. [45] A. Wahrburg et al., “Modeling speed-, load-, and position-dependent
[21] P. W. Holland and R. E. Welsch, “Robust regression using iteratively friction effects in strain wave gears,” in Proc. IEEE Int. Conf. Robot.
reweighted least-squares,” Commun. Statist. Theory Methods, vol. 6, no. 9, Autom., May 2018, pp. 2095–2102.
pp. 813–827, 1977. [46] S. Wolf and M. Iskandar, “Extending a dynamic friction model with
[22] J. P. Norton, An Introduction to Identification. New York, NY, USA: Dover nonlinear viscous and thermal dependency for a motor and harmonic drive
Publications, Inc., 2009. gear,” in Proc. IEEE Int. Conf. Robot. Autom., May 2018, pp. 783–790.
[23] N. Ramdani and P. Poignet, “Robust dynamic experimental identification [47] M. Iskandar and S. Wolf, “Dynamic friction model with thermal and load
of robots with set membership uncertainty,” IEEE/ASME Trans. Mecha- dependency: Modeling, compensation, and external force estimation,” in
tronics, vol. 10, no. 2, pp. 253–256, Apr. 2005. Proc. Int. Conf. Robot. Autom., May 2019, pp. 7367–7373.
[24] N. Ramdani and P. Poignet, “Experimental parallel robot dynamic [48] H. Mayeda, K. Yoshida, and K. Osuka, “Base parameters of manipula-
model evaluation with set membership estimation,” in Proc. 14th IFAC tor dynamic models,” in Proc. IEEE Int. Conf. Robot. Autom., vol. 3,
Symp. Identification Syst. Parameter Estimation Volumes, vol. 39, 2006, Apr. 1988, pp. 1367–1372.
pp. 967–972. [49] M. Capurso, M. M. G. Ardakani, R. Johansson, A. Robertsson, and
[25] A. Janot, P. O. Vandanjon, and M. Gautier, “Using robust regressions and P. Rocco, “Sensorless kinesthetic teaching of robotic manipulators assisted
residual analysis to verify the reliability of LS estimation: Application in by observer-based force control,” in Proc. IEEE Int. Conf. Robot. Autom.,
robotics,” in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Oct. 2009, May 2017, pp. 945–950.
pp. 1962–1967. [50] I. CVX Research, “CVX: Matlab software for disciplined convex program-
[26] P. Boggs, R. Byrd, and R. Schnabel, “A stable and efficient algorithm for ming, version 2.0,” Aug. 2012, [Online]. Available: http://cvxr.com/cvx
nonlinear orthogonal distance regression,” SIAM J. Sci. Statist. Comput., [51] M. Grant and S. Boyd, “Graph implementations for nonsmooth convex
vol. 8, no. 6, pp. 1052–1078, 1987. programs,” in Recent Advances in Learning and Control (Lecture Notes
[27] R. Fierro, “The total least squares problem: Computational aspects and in Control and Information Sciences series), V. Blondel, S. Boyd, and
analysis (S. Van Huffel and J. Vandewalle),” SIAM Rev., vol. 35, no. 4, H. Kimura, Eds. Springer-Verlag Limited, 2008, pp. 95–110. [Online].
pp. 660–662, 1993. Available: http://stanford.edu/ boyd/graph_dcp.html.
[28] G. Calafiore and M. Indri, “Robust calibration and control of robotic ma- [52] P. J. Huber, “Robust estimation of a location parameter,” in Breakthroughs
nipulators,” in Proc. Amer. Control Conf., vol. 3, Jun. 2000, pp. 2003–2007. in Statistics. Amsterdam: Netherlands: Springer, 1992, pp. 492–518.

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.
1594 IEEE TRANSACTIONS ON ROBOTICS, VOL. 36, NO. 5, OCTOBER 2020

[53] R. H. Byrd, J. C. Gilbert, and J. Nocedal, “A trust region method based Chao Liu received the B.E. degree in agricultural
on interior point techniques for nonlinear programming,” Math. Program., mechanization and automation from Shandong Agri-
vol. 89, no. 1, pp. 149–185, Nov. 2000. cultural University, Taian, China, in 2007, the M.E.
[54] D. C. Montgomery, E. A. Peck, and G. G. Vining, Eds., Introduction to degree in mechanical manufacturing and automation
Linear Regression Analysis, 5th Ed. Hoboken, NJ, USA: John Wiley & from Jilin University, Changchun, China, in 2009, and
Sons, Inc., 2012. the Ph.D. degree in mechatronics from Shanghai Jiao
[55] K. Johanastrom and C. Canudas-de-Wit, “Revisiting the Lugre friction Tong University, Shanghai, China, in 2015.
model,” IEEE Control Syst. Mag., vol. 28, no. 6, pp. 101–114, Dec. 2008. He is currently an Assistant Professor with the
[56] F. Al-Bender, V. Lampaert, and J. Swevers, “The generalized Maxwell-slip School of Mechanical Engineering, Shanghai Jiao
model: A novel model for friction simulation and compensation,” IEEE Tong University. His research interests focus on the
Trans. Autom. Control, vol. 50, no. 11, pp. 1883–1887, Nov. 2005. high-performance motion control algorithm, embed-
[57] T. Lee, P. M. Wensing, and F. C. Park, “Geometric robot dynamic identi- ded motion controller design, and drive-control integrated robot controller
fication: A convex programming approach,” IEEE Trans. Robot., vol. 36, design.
no. 2, pp. 348–365, Apr. 2020.

Yong Han received the B.E. degree in mechanical en- Zhenhua Xiong received the B.E. degree in aircraft
gineering from the Nanjing University of Science and design and M.E. degree in vibration engineering from
Technology, Nanjing, China, in 2014. He is currently the Department of Aircraft Design, Nanjing Univer-
working toward the Ph.D. degree in mechatronics at sity of Aeronautics and Astronautics, Nanjing, China,
the State Key Laboratory of Mechanical Systems, and in 1995 and 1998, respectively, and the Ph.D. degree
Vibration, Shanghai Jiao Tong University, Shanghai, in mechatronics from the Electrical and Electronics
China. Engineering Department, Hong Kong University of
His research interests include motion control, Science and Technology, Clear Water Bay, Hong
physical human robot interaction, and model Kong, in 2002.
identification. He is currently a Professor with the School of
Mechanical Engineering, Shanghai Jiao Tong Uni-
versity, Shanghai, China. His research interests include servo motion control
and intelligent manufacturing.

Jianhua Wu received the B.E. degree in engineering

mechanics and the Ph.D. degree in mechatronics from
Shanghai Jiao Tong University, Shanghai, China, in
2002 and 2007, respectively.
He is currently an Associate Professor with the
School of Mechanical Engineering, Shanghai Jiao
Tong University. His research interests include high-
speed and high-precision motion control, high-
performance motor driver design, and servo self-
tuning technology, which are applied to numerical
control systems.

Authorized licensed use limited to: UB Siegen. Downloaded on January 26,2024 at 09:17:48 UTC from IEEE Xplore. Restrictions apply.