Springer Nature 2021 LATEX template

Adaptive Corporation of Optimal Linear Quadratic Regulator

and Lumped Disturbance Rejection Estimator based Tracking
Control for Robotic Manipulators.
Abstract
This paper presents a collaborative control regime including an estimation and optimal tracking
control for the operation of robotic systems. A class of second order mechanical robotic systems is
considered and expressed with Euler-Lagrange equations. Based on this model, an adaptive observer,
with a structure of general proportional integral estimator, is derived to manage the influence of
disturbances, uncertain components in the working process of system. With the availability of the
designed observer, by using some computational techniques, we restructure dynamical model of the
studied system to be suitable for designs of feedback controller. To reach the optimal performance of
the closed-loop system, an optimal feedback control methodology is developed. The optimal controller
gain matrix is obtained by using an optimisation procedure. The effectiveness of the proposed method
is validated by undertaking comprehensive simulation with various case studies on the model of a
planar robot with two degrees of freedom.

Keywords: Optimal Tracking Controls, Adaptive Observer, Disturbances Rejection, Planar Robots

1 Introduction this reason, the fuzzy logic system were used with
the combinations of proportional derivative (PD)
Robotic systems have been increasingly used in and proportional integral derivative (PID) con-
industrial production to improve productivity and trollers to improve the adaptability of the system
product quality. The problem of smooth and effi- [7–9]. By adjusting the controllers’ parameters
cient control of robotic systems has always been according to the changes of system disturbances,
a problem that attracts much attention from undesirable influences of the disturbances on the
investors and research groups around the world outputs can be partially reduced.
[1]. Requirements for robot control systems are On the other hand, control methods for non-
not only to complete the prescribed tasks, but also linear systems have been researched and devel-
to achieve the optimal system’s performance. The oped based on Lyapunov’s theory to enhance
designed controller needs to be able to respond to the system’s responsiveness. Some typical non-
the effects of adverse impacts such as occurrence linear controllers such as sliding and back-stepping
of harmful disturbances and the unfavourable controllers were developed in literature [10–15],
system’s uncertainties [2]. For robotic systems but these methods cause an undesirable chatter-
where the dynamic structure and presentation are ing phenomenon. In addition, when the system’s
fully known and reliable, feedback controllers such parameters change during operation and/or are
as proportional derivative, proportional integral unknown, model-based controllers no longer retain
derivative controllers can be employed to attain their effectiveness. Therefore, adaptive control
the system performance [3–6]. These approaches regimes were investigated to improve the quality
is convenient and simple in designing controllers, of the system’s responses. The adaptive controller
nonetheless, they were ineffective in eliminating could be developed from backstepping and sliding
the adverse impacts of disturbances and noise on mode control strategies [16], but for manipula-
the performance of the closed-loop systems. For tor, the common adaptive control techniques were

1
 Springer Nature 2021 LATEX template

derived by using fuzzy logic theory or neural net- of linear quadratic regulation (LQR) and distur-
works [17–20]. Due to the property of being able bances observer was studied at [31–33]. However,
to approximate non-linear functions with arbi- the disturbances observers in [31] and [32] are only
trarily small errors, neural networks and fuzzy used to approximate the input disturbances, so
systems have great abilities in determining con- the influence from non-linear components in the
trol signals without requiring much the knowledge system still exists and the observer in [33] approx-
of the system’s information . However, computa- imate is used to approximate each of the different
tional complexity and large storage requirements quantities of the system so that the calculation
limit their applicability. volume will increase.
With the idea of using observer to eliminate Inspired from aforementioned works, in this
the effect of system disturbances and the system’s paper, an adaptive collaborative control regime
uncertainties, a controller design method was pro- comprising disturbance rejection observers and an
posed in [21] to achieve the control requirements optimal LQR controller is taken on a consideration
of non-linear systems. In this work, all effects to track prescribed paths and achieve the effi-
of noise, disturbances unknown components of cient performance of robotic manipulators. Firstly,
the dynamic model of the system were aggre- we consider a dynamical representation of a class
gated into a lumped disturbances and removed by of non-linear robotic systems with disturbances.
deploying observers. As a result, these dynamical In the following step, by using some advance
model of the non-linear systems were converted mathematical computation, we restructure origi-
into forms of linear systems leading to the ease nal system’s model to establish a unique Lumped
of controller design. Several other observer design disturbances which are time-varying functions
methods with similar purposes were proposed in containing external disturbances, uncertain fac-
[22–25]. Besides that, optimal control problem is tors and non-linear components. In the next part,
also considered [26, 27]. The above documents we propose an intelligent observer based on the
[22–27] have not considered the combination of available restructured system models to gener-
their observers with optimal control problems as ate the estimation of the varying functions by
well as other control algorithms. For non-linear applying technique of Taylor approximation. With
systems such as manipulator, the optimization the presence of estimated signals, we obtained
problems can be solved through neural networks an adaptive observer-based controller to achieve
with typical structure of actor-critic [28, 29]. The control requirements of closed-loop robotic sys-
neural network showed their ability to approx- tems. Furthermore, with the availability of the
imate the solution of the optimization problem designed observer, the impacts of the lumped
posed directly for the non-linear system, but disturbances can be removed, and the original
the complexity in structure as well as the com- non-linear system will be transformed a linear
putational volume is the biggest limitation of representation which is suitable for implementing
this method. A simpler way to design the opti- an additional feedback controller to enhance the
mal controller is to convert the system to linear performance of the systems. By using techniques
form by using linearization methods and Taylor of linear quadratic regulation, an extra adaptive
series approximation is an effective option. The optimal controller is developed to cooperatively
application of LQR controller in obtained lin- works with the designed adaptive observer-based
earization model by Taylor series approximation controller to achieve better performance of the
was shown in [30]. Although the computational robotic manipulator.
complexity is reduced, the control quality can The remaining of this paper is organized as fol-
still be affected by the deviation between the lows: Section II presents system description and
linearized model and the real model of the sys- restructured models of the studied robotic manip-
tem. Furthermore, it is necessary to consider the ulators, and the structure of adaptive control
occurrence of disturbances in the optimal con- strategy. The synthesis of adaptive control regime
trol problem. The solutions given in this case can including lumped disturbance rejection observer
include disturbances observers. The combination and optimal LQR are demonstrated in Section III
In Section IV, comprehensive simulations will be
 Springer Nature 2021 LATEX template

the vector of the control signals and the input dis-

turbances T e . Matrix M ∈ Rn×n is the inertial
matrix of the system (M is symmetric and posi-
𝑙2 tively definite), C ∈ Rn×n represents the mutual
and centrifugal component, G ∈ Rn×1 represents
the gravity component. System (1) describes a
𝑙𝑐2 general form for the dynamic model of a robotic
m2 g
𝑙1 system. In particular, for a two degree of free-
𝑞2 dom planar robot, q = [q1 q2 ] and n is selected
as n = 2. It is noted that for the simplicity
𝑙𝑐1
of presentation, we merely present the structure
m1 g of a system planar two-link robot arm, however,
𝑞1 research results in this paper can be extended for
general robotic systems.

2.2 Restructured models

Fig. 1 The structure of a planar two-link robot arm Regarding to the operation of a robotic system,
the robot will implement a task following pre-
undertaken on the model of a planar robotic sys- scribed purposes. To do that, a control system is
tem with two degrees of freedom to validate the developed to ensure the robot can realise given
effectiveness of our control regime. Finally, Section tracking tasks. In general, rotation angles, q, of
V concludes our paper. the planar robot will attain to achieve desir-
able angles (reference signals), r, after some sec-
2 System Description and onds of settling times. In this paper, we aim to
develop a control strategy to achieve the track-
Problem Statement ing requirements of predefined paths and stabilize
the closed-loop robotic systems containing highly
2.1 System Description non-linear, uncertain factors, and the fluctuation
Fig. 1 illustrates a model of a planar two- of disturbances (see (1)). To achieve this challeng-
link robotic arm. The operation and dynamic ing research goal, in the following part we will
equations of the robotic system can be found in restructure the studied system model (see (1)).
[22]. The notations in Fig. 1 are given as follows At first, we rewrite system (1) in the following
form
ˆ q1 , q2 are the rotation angle of the two corre-
sponding link of the planar robotic arm.  
−1
ˆ m1 , m2 are the mass of Link 1 and Link 2. q̈ = M (q) T + T e − C(q, q̇)q̇ − G(q) . (2)
ˆ l1 , l2 are the length of Link 1 and Link 2.
ˆ J1 , J2 are the moment of inertia of the first and Let we define ζ = r − q as a control error and by
second links. adding the second derivative, r̈, of r to both sides
ˆ g is the acceleration due to gravity of (2), we obtain
By applying by Euler-Lagrange method the fol-
lowing structure of the planar two-link robotic ζ̈ = r̈ − q̈
arm presented in Fig. 1 can be obtained as [22]
 
−1
= r̈ − M (q) T + T e − C(q, q̇)q̇ − G(q)



M (q)q̈ + C q, q̇ q̇ + G q = T + T e . (1)
= − M (q)−1 T + r̈ − M (q)−1 T e
where q ∈ Rn×1 is the state space fitting vari- + M (q)−1 C(q, q̇) + M (q)−1 G(q).
able vector, q̇ and q̈ are respectively the first and (3)
the second order derivatives of q , T ∈ Rn×1 is Now, we add a proportional term, Ωζ, and an inte-
gral term, Ψζ̇, into both sides of equation (3). Gain
 Springer Nature 2021 LATEX template

matrices Ω and Ψ with appropriate dimensions be derived as

will be determined later upon to specific control T = M (q)fˆ. (8)
requirements. As a result, system (7) becomes

ζ̈ + (Ωζ + Ψζ̇) = Ωζ + Ψζ̇ − M (q, q̇)−1 T

 
−1
ẋ = Ax − M (q) M (q)f + f = Ax. (9)
+ r̈ − M (q)−1 T e
(4)
+ M (q)−1 C(q, q̇)q̇
System (9) will be stable if matrix A is Hurwitz
+ M (q)−1 G(q). which has all negative eigenvalues. This stability
depends on the selections of pair matrices of Ψ
We denote a time-varying function, f , as follows and Ω. Therefore, if f is available, an adaptive
controller, based on f and appropriate matrices
f = r̈ − Ωζ − Ψζ̇ − M (q)−1 T e Ψ and Ω, can be implemented to guarantee the
(5) stability and tracking performance of the studied
+ M (q)−1 C(q, q̇) + M (q)−1 G(q).
robotic system (1).
However, it is known that f is unmeasurable
Accordingly, we obtain the following result from
and accessible. To deal with this challenging issue,
(4) and (5)
in this paper, we propose an observer design which
can reconstruct the unmeasured signals, fˆ of f .
ζ̈(t) = Ωζ + Ψζ̇ − M (q)−1 T + f . (6)
Furthermore, in order to improve the performance
T of the robotic system and relax the difficulty in
Let we define x = [ζ T ζ̇ ]T as a state variable selecting pair matrices Ψ and Ω, we propose a
vector, we can achieve the following system new adaptive cooperative a new control methodol-
     ogy which is a co-corporative control regime which
ζ̇ 0 I ζ combines an intelligent observer (estimator) and a
= n n − M (q)−1 T + f ,
ζ̈ Ω Ψ ζ̇ (7) feedback controller as depicted in Fig. 2.
ẋ = Ax − M (q)−1 T + f .
 
Intelligent 𝑓መ
0n In
where A = . It is noted that matrix A is Observer
Ω Ψ Reference
+
𝒯 Robotic
in a form of a non-singular matrix. 𝛴 𝑥 𝛴 𝑀
ሾ𝑟 𝑟ሶ ሿ𝑇+ System 𝑇
− ൣ𝑞 𝑞ሶ ൧
Optimal 𝜏𝑙𝑛 −
Remark 1 : As can be seen from (5), time-varying vec- Controller
tor, f , is the combination of system disturbance T e ,
some non-linearity components of the robotic systems, Adaptive cooperative control strategy
second-order derivation of reference signal, r̈. There-
fore, f is different to system disturbance vector, T e .
To distinguish from T e , time-varying functions, f will Fig. 2 The structure of proposed control strategy.
be called ”lumped disturbances” and be represented
as disturbances of system (7). Fig. 2 illustrates the structure of our pro-
posed control system. As can be seen that, in
Fig. 2, an intelligent observer will be developed
2.3 Structure of adaptive to reconstruct the estimation, f̂ , of f and another
controller will be design to generate control signal,
in-corporative control strategy
τ ln . By the collaboration of observer and con-
It can be seen that, system (7) contain f = troller, the adaptive control signal, T in (7) is
[f1 (t) f2 (t)]T , where f1 (t) and f2 (t) are time- expressed in the following form
varying functions. If f1 (t) and f2 (t) are both  
measured or accessible, an adaptive controller can T = M (q) − τ ln + f ˆ (10)
 Springer Nature 2021 LATEX template

By using the controller form in (10), a closed- first, an intelligent estimator (observer) will be
loop system of proposed robotic system can be designed to reconstruct the estimation f̂ of f , then
rewritten as follows an optimal controller will be developed to collabo-
rate with the observer to improve the performance
ẋ = Ax − M (q)−1 T + f and guarantee the performance of the closed-loop
 
system.
= Ax − M (q)−1 M (q) − τ ln + f̂ + f
(11)
= Ax + τ ln + (f − f̂ ) 3.1 Intelligent estimator (observer)
= Ax + Bu + ∆f .
design
  In this part, an intelligent observer is deployed to
0n estimate the time-varying term, f . In this paper, f
where B = , u = τ ln and ∆f = f − f̂ .
In is assumed to be approximated by a Taylor expan-
It is very clear that if ∆f converge to zeros, sion with a sufficiently large number of degrees
system (11) becomes a linear system as follows m, m is integer, which satisfies f m =0. In order to
design the observer, let we define some following
ẋ = Ax + Bu + ∆f notations
(12)
=Ax + Bu, 

 Θ1 = ζ,

and x(t) converge to zero after some seconds of Θ2 = ζ̇,




settling time, system (11) becomes stable. Accord-

Υ1 = f ,



ingly, the tracking tasks of the proposed robotic



system will be achieved and the proposed robot



 Υ2 = ḟ ,
.. ..

planar can implement any task based on pre-

.=. (13)
scribed requirement, r.  (k−1)
Υk = f ,





 k
Remark 2 : Since the lumped disturbances f different



 Υ(k−1) = f ,
from input disturbances T e , the observer designed to .. ..



.=.

approximate f will also have a different purpose from 


those which approximate T e . When the effect of f Υ(m−1) =f m ,


is estimated and eliminated, the nominal system will
be converted to (11) which is a linear system model.
where k ∈ [1, m] denotes the order derivative of f .
Thereby, the controller design methods for linear sys-
From (13), we obtain thier derivatives as fol-
tem can be used and in this article we also focus on the
optimal control problem so that the Linear quadratic lows
regulator (LQR) technique is a suitable and effective 
choice to solve optimal control problem for the linear 
 Θ̇1 = Θ2 ,

Θ̇2 = − M (q)−1 T + ΩΘ1 + ΨΘ2 + Υ1 ,

model. 




Υ̇1 = Υ2 ,




The main purpose of this paper now is to

Υ̇2 = Υ3 ,
develop an intelligent observer to estimate the 
 .. ..
value of function f and an optimal control which



 .=.
collaborate with the observer to guarantee the



 Υ̇(m−1) =Υ(m)
tracking requirements and stability of the pro-



Υ̇m = 0.

posed closed-loop robotic system.
(14)
3 Adaptive cooperative
control design
In this section, we will develop a cooperative con-
trol strategy for the proposed robotic system. At
 Springer Nature 2021 LATEX template

We employ the following structure of an intelligent By replacing values from (18) and (17) into (11),
observer (m+2)
we obtain Θ̂1 as follows
˙



 Θ̂1 = Θ̂2 + Γm+1 (Θ1 − Θ̂1 ), (m+2)
Θ̂1 = Θ̂2
(m+1)
+ Γm+1 ζΘ
(m+1)

˙

Θ̂2 = ΩΘ̂1 + ΨΘ̂2 − M (q)−1 T + Υ̂1

 (m+1) (m) (m)
= Γm+1 ζΘ + ΩΘ̂1 + ΨΘ̂2




+ Γm (Θ1 − Θ̂1 ),

(m) (m)
− (M (q)−1 T )(m) + Υ̂1

+ Γm ζΘ



˙


 Υ̂1 = Υ̂2 + Γm−1 (Θ1 − Θ̂1 ),
 (m+1) (m) (m)
= Γm+1 ζΘ + ΩΘ̂1 + ΨΘ̂2
˙
Υ̂2 = Υ̂3 + Γm−2 (Θ1 − Θ̂1 ),
  
 (m−1) (m−1)
 + Υ̂2 + Γm−1 ζΘ
.. ..



.=.


(m)

− (M (q)−1 T )(m) + Γm ζΘ


 ˙
Υ̂m−2 = Υ̂m−1 + Γ1 (Θ1 − Θ̂1 ),



 (m+1) (m) (m+1)

˙
 = Γm+1 ζΘ + ΩΘ̂1 + ΨΘ̂1
Υ̂m−1 = Γ0 (Θ1 − Θ̂1 ). (m)
(15) − ΨΓm+1 ζΘ − (M (q)−1 T )(m)
with Γk ∈ Rn×n , k = 1, 2, ..., m + 1 are diagonal

(m) (m−1)
matrices of which all the diagonals are positive. + Γm ζΘ + Γm−1 ζΘ + ... + Γ1 ζ̇Θ
Let we define ζΘ = Θ1 − Θ̂1 as observer error. 
In the following part, we will work out the exis- + Γ 0 ζΘ .
tence condition of designed observer. The (m+2)
(19)
order derivatives of Θ̂1 can be obtained from (14)
On the other hand, from (6), we have ζ̈ = Ωζ +
and (15) as follows
Ψζ̇ − M (q)−1 T + f . Therefore
(m+2) (m+1)
Θ̂1 = Θ̂2 + Γm+1 (Θ1 − Θ̂1 )(m+1)
(16) −(M (q)−1 T ) = ζ̈ − Ωζ − Ψζ̇ − f
(m+1) (m+1)
= Θ̂2 + Γm+1 ζΘ (2) (1)
= Θ1 − ΩΘ1 − ΨΘ1 − f
(m+2) (m)
The (m + 1)-order derivatives of Θ̂2 can be −(M (q)−1 T )(m) = Θ1 − ΩΘ1
obtained from (14) and (15) as follows (m+1)
− ΨΘ1 − f (m) .
(m+1) (m) (m) (20)
Θ̂2 = ΩΘ̂1 + ΨΘ̂2 − (M (q)−1 T )(m) Substitute (20) into (19), we obtain
(m)
+ Υ̂1 + Γm (Θ1 − Θ̂1 )(m)
(m+2) (m+1) (m) (m+1)
(m) (m) Θ̂1 = Γm+1 ζΘ + ΩΘ̂1 + ΨΘ̂1
= ΩΘ̂1 + ΨΘ̂2 − (M (q)−1 T )(m)
(m)
(m) (m) − ΨΓm+1 ζΘ − (M (q)−1 T )(m)
+ Υ̂1 + Γ m ζΘ 
(17) (m)
+ Γm ζΘ + Γm−1 ζΘ
(m−1)
+ ... + Γ1 ζ̇Θ
By deploying a similar computation method in
(16) and (17), from (14) and (15), we obtain

+ Γ 0 ζΘ ,
 (m) (m−1) (m−1)
 Υ̂1 = Υ̂2 + Γm−1 ζΘ , (m+1) (m) (m+1)

 = Γm+1 ζΘ + ΩΘ̂1 + ΨΘ̂1
(m−1) (m−2) (m−2)

Υ̂ = Υ̂3 + Γm−2 ζΘ ,

 (m) (m+2)
 2 − ΨΓm+1 ζΘ− f (m) + Θ1


.. .. (18) 
.=. (m) (m+1) (m)
− ΩΘ1 − ΨΘ1 + Γm ζΘ






 Υ̂m−1 = Υ̂m−1 + Γ1 ζ̇Θ ,
 
Υ̂m = Γ0 ζΘ . (m−1)

+ Γm−1 ζΘ + ... + Γ1 ζ̇Θ + Γ0 ζΘ .
(21)
 Springer Nature 2021 LATEX template

Therefore, With m = 3, the the structure of (15) will be

simplified to a system of equations consisting of
(m+2) (m+1)
f (m) = ζΘ + (Γm+1 − Ψ)ζΘ ˙



 Θ̂1 = Θ̂2 + Γ4 (Θ1 − Θ̂1 ),
(m) 
˙
+ (Γm − Ω − P siΓm+1 )ζΘ (22) 
Θ̂ = ΩΘ̂1 + ΨΘ̂2 − M −1 T + Υ̂1 + Γ3 (Θ1 − Θ̂1 ),
 2



(m−1) ˙
+ Γm−1 ζΘ + ... + Γ1 ζ̇Θ + Γ0 ζΘ Υ̂1 = Υ̂2 + Γ2 (Θ1 − Θ̂1 ),

˙


2 = Υ̂3 + Γ1 (Θ1 − Θ̂1 ),
Υ̂

The characteristic polynomials is 


˙

Υ̂3 = Γ0 (Θ1 − Θ̂1 )
H(s) = In sm+2 + (Γm+1 − Ψ)sm+1 (28)
The characteristic polynomials H(s) becomes
+ (Γm − Ω − ΨΓm+1 )sm (23)
+ Γm−1 s m−1
+ ... + Γ1 s + Γ0 H(s) = In s5 + (Γ4 − Ψ)s4 + (Γ3 − Ω − ΨΓ4 )s3
(29)
+ Γ2 s2 + Γ1 s + Γ0
Due to f (m) ≈ 0, according to (22), the observer and if the characteristic polynomials H is desired of
error, ζΘ , can be converged to zero if we can choose the form
the parameter matrices so that the polynomial H(s) = In s5 + WH4 s4 + WH3 s3 + WH2 s2 + WH1 s
H(s) is a Hurwitz form
+ WH0
(30)
Remark 3 : Even though the observer is designed with the parameter matrices will be determined respec-
a sufficiently large of parameter m, we can still sim- tively as 
plify the structure of the observer in cases where m Γ4 = WH4 + Ψ,


is small. The structure of observer when m = 1 and Γ3 = WH3 + Ω + ΨΓ4 ,



m = 3 will be presented as follow. Γ = WH2 , (31)
For m = 1, the structure of observer (15) becomes  2
Γ1 = WH1 ,


˙ 


Θ̂ = Θ̂2 + Γ2 (Θ1 − Θ̂1 ), 
Γ0 = WH0
 1


˙
 Θ̂2 = ΩΘ̂1 + ΨΘ̂2 − M −1 T + Υ̂1 + Γ1 (Θ1 − Θ̂1 ),
˙

Υ̂1 = Γ0 (Θ1 − Θ̂1 )
(24) 3.2 Optimal Controller Design
and the characteristic polynomials H(s) can be rewrit-
ten as
As we discussed in Section 2, the main purpose
of this paper now is to develop collaborative con-
H(s) = In s3 + (Γ2 − Ψ)s2 + (Γ1 − Ω − ΨΓ2 )s + Γ0 trol strategy including an intelligent observer to
(25)
estimate the value of function f and an optimal
We can see that choosing the parameter matrices
control which guarantee the tracking requirements
to make H Hurwitz is quite simple. If we want the
characteristic polynomials H to have the form of the proposed robotic system. From observer. we
can obtain the estimation f̂ of fˆ comprising the
H(s) = In s3 + WH2 s2 + WH1 s + WH0 (26) uncertainty components and system disturbance,
the parameter matrices will be determined respec- Te . Let we remind that the control signal T will
tively as  consist of 2 components satisfying
Γ2 = WH2 + Ψ,

 
Γ1 = WH1 + Ω + ΨΓ2 , (27)
 T = M (q) − τ ln + f̂ . (32)
Γ0 = WH0


Hence, the closed-loop system can be expressed as

in (11)
ẋ = Ax + Bτ ln (33)
In this section, we will derive an optimal controller
which collaborate with the operation of observer
to optimally ensure the performance of the studied
 Springer Nature 2021 LATEX template

robotic system. Generally, if matrix A is chosen as from the algorithm is to the exact solution of
a Hurwitz matrix, the system (33) will be stable, (35), meaning that lim Lk = L∗ then get the
k→∞
which means that the control errors will go to zero. coefficient matrix of the controller as close to the
We will continue to use the optimization algorithm optimal coefficient matrix lim Rk = R∗ . The
for the linear system to further improve the track- k→∞

ing quality for the studied robotic system. Hence, proof has been presented in [34] and we obtain the
our objective now is to design a Linear quadratic parameter matrix for the optimal controller LQR.
regulator (LQR) to minimize the following cost
function: Remark 4 : As we mentioned in the article, if ∆f con-
verges to zero the system (12) will become (33) and
Zt the controller is designed to stabilize system (33) will
1
xT Cx + τ Tln Dτ ln dτ


J= (34) also stabilize the original system. In this Remark, we
2 will provide the proof for the stability of closed loop
0
system even there exist an approximation error, ∆f
With C = C T ≥ 0 and D = DT > 0. From [34], between f and and the estimation, fˆ. Before proving
the task of finding the optimal control signal τ ln is the stability of closed loop system, we give some prop-
erties proven in [34, 35], for all positive integers k, we
equivalent to finding a positive definite symmetry
have
roots L∗ of the Riccati equation
ˆ A − BRk is Hurwitz
AT L + LA + C − LBD−1 B T L = 0 (35) ˆ lim Lk = L∗
k→∞
ˆ lim Rk = R∗
Then, the coefficient matrix of the optimal con- k→∞

troller R∗ is calculated by The stability of the closed loop system will be

proved as follows: assume the Kleiman’s algorithm ter-
R∗ = D−1 B T L∗ (36) minates at the K th iteration and we get coefficient
matrix of the controller RK , the system become
Here we will use the approximate solution method ẋ = (A − BRK ) x + ∆f (39)
proposed by Kleiman in [34] to approximate the
with ∆f = f − fˆ is estimation error. Matrix RK satisfy
solution of (35) instead of directly solving the
A − BRK is Hurwitz, which is equivalent to
often difficult Riccati equation. This method is
presented as follows: (A − BRK )T P + P (A − BRK ) = −Q (40)

ˆ Determine R0 as the matrix of the state feed- with Q is a positive definite matrix have always a
positive definite solution P .
back controller. If A has made the system
We use the following Lyapunov function
stable, then R0 can be chosen including all zero
elements. V = xT P x (41)
ˆ Solving the Lyapunov equation Taking the derivative of V , we have
h i
T V̇ = xT (A − BRK )T P + P (A − BRK ) x+2xT P ∆f
(A − BRk ) Lk + Lk (A − BRk )
(42)
= −C − RkT DRk (37) from (40) we deduce

To find the solution for Lk with k = 0, 1, ... V̇ = − xT Qx + 2xT P ∆f

(43)
ˆ Calculate Rk+1 from Lk using the formula ≤ − ∥x∥2 λmin + 2∥x∥∥P ∥∥∆f ∥
where λmin is the smallest
  eigenvalue of Q. 
Rk+1 = D−1 B T Lk (38)  2∥P ∥∥∆f ∥
When x ∈ / O = x∥x∥ < λmin ∥x∥ <
Repeat the second and third steps of the algorithm 2∥P ∥∥∆f ∥
we have V̇ < 0 then x will converge to O
λmin
until the error satisfies the condition Lk+1 −Lk < ϵ containing 0. And when ∥∆f ∥ also converge to zero
for a given arbitrarily small ϵ. With Kleiman’s according to the precision of the observer, the area O
algorithm, it can be proved that the larger the also is shrunk, thereby ensuring that x is always in a
number of iterations, the closer the solution found very small neighborhood of 0. The Proof is completed.
 Springer Nature 2021 LATEX template

Remark 5 : m2 l2
m22 = + J2
4  
Algorithm 1 LQR with adaptive observer c11 c12
The Centripetal and Coriolis matrix C =
c21 0
1. Initialization:
c11 = −q̇2 m2 l1 l2 sin q2
ˆ Choose 2 matrices Ω and Ψ such that A
is Hurwitz. m 2 l1 l2
ˆ Choose parameter m and desired form of c12 = −q̇2 sin q2
2
the characteristic polynomials, H. m2 l1 l2
c21 = q̇1 sin q2
2. Calculation: 2
 T
and the Gravitational vector G = g1 g2 with
ˆ Determine the parameter matrix R∗
for the optimal controller LQR using 
m1 gl1

m2 gl2
Kleiman approximation method. g1 = + m2 gl1 cos q1 + cos(q1 +q2 )
ˆ Calculate the parameter matrices 2 2
Γ0 , Γ1 , ..., Γm+1 of the observer. In the m2 gl2
case of m = 1 the parameter matrices g2 = cos(q1 + q2 )
2
can be calculated through (27) and in Parameters of the Robot are shown in Table 1.
the case of m = 3 can use (31)
3. Controller synthesis Parameter Value
m1 [kg] m2 [kg] l1 [kg] l2 [kg] J1 [kg] J2 [kg]
ˆ Output of LQR controller: τ ln = −R∗ x 0.452 0.269 0.2 0.18 0.122 0.0447
ˆ Output of observer: fˆ = Υ̂1 Table 1 The parameters of Planar Robot
ˆ Synthesize control signal vector T
according to (31).

4.2 Control validation

To visualise the effectiveness of the coopera-
4 Validation of Control tive method, the system disturbance, T e =
Scheme [τext1 τext2 ], are selected as incremental step
changes as illustrated in Fig. 3 and and a
4.1 Data for Robotic system time-varying disturbances which is the sum
of this step disturbances and a cyclic distur-
To verify the effectiveness of the proposed adap-
bances [0.5 (1 + sin (2πt)) ; 0.05 (1 + cos (2πt))] as
tive optimal controller, the algorithm will be vali-
depicted in Fig.4.
dated on the two degree of freedom Planar Robot
For this aim, one optimal controller and two
model described in Fig. 1. The dynamic model of
intelligent observers are developed. Firstly, by
the Planar Robot can be expressed in a general
using the proposed algorithm in Section 3.2, we
Euler-Lagrange form [22] as follow
obtain the optimal controller’s R as follows




M q q̈ + C q, q̇ q̇ + G q = T + T e . (44) 
1.9258 0 4.2777 0

R= (45)
  0 2.5913 0 5.8820
m11 m12
Where M = with
m21 m22 It is noted that, for the optimal controller
design, the error between two consecutive iter-
m1 l1 2 l2 2
 
2 ations for the Kleiman algorithm is chosen as
m11 = +J1 +m2 l1 + + l1 l2 cos q2 +J2
4 4 0.0001. Matrices C and D used in the algorithm
are
C = diag ([100; 100; 100; 100])
 
m2 l2 l2
m12 = m21 = + l1 cos q2 + J2 (46)
2 2 D = diag ([1; 1])
 Springer Nature 2021 LATEX template

1
ˆ Adaptive control strategy 1 (called Controller
1) is a combination of the optimal controller
1

0.5
ext

(45) and an intelligent observer (47) where the

0 parameters for Taylor approximation of f is
m=1.
0 2 4 6 8 10
ˆ Adaptive control strategy 2 (called Controller
2) is a collaboration of the optimal controller
0 (45) and an intelligent observer (48) (called
2
ext

-0.1 LQRAOB-3) where the parameters for Taylor

-0.2 approximation of f is m = 3 .
0 2 4 6 8 10 In this section, there are three following scenarios
Time (Seconds) of simulation are undertaken
Fig. 3 System disturbances T e (τext1 , τext2 )
, (Scenario I1 and I2 ) ˆ (Scenario I1 ): the system reference, r is con-
stant signals. In this case, r = [r1 (t) r2 (t)]T =
2 [1 1]T , which are equivalent to 57.29o . The sys-
tem disturbances, Te are step changes as shown
1
ext

1
in Fig. 3.
0 ˆ (Scenario I2 ): the system reference, r is
0 2 4 6 8 10 cyclic signals, r = [r1 (t) r2 (t)]T , where
r1 (t) = 0.5 sin(0.2πt) − sin(0.4πt) and r2 (t) =
0.2
0.5 sin(0.2πt) + sin(0.4πt). The system distur-
2

0 bances, Te are illustrated in Fig. 3.

ext

ˆ (Scenario I3 ): the system references, r is similar

-0.2
Scenario I2 , the system disturbances, Te , are the
0 2 4 6 8 10
complex signals as depicted in Fig. 4.
Time (Seconds)
Fig. 4 Time varying disturbances Te = [τext1 ; τext1 ], Figures 5 and 6 show the responses of Link
Scenario I3 . 1 and 2 angles, q1 (t), q2 (t) of the closed-loop
robotic system with constant references, r(t) and
Based on the observer’s design method in Section under impacts of system disturbances, T e . As can
3.1, we develop two observers with the parameters be seen from Figures 5 and 6 that our proposed
of Taylor approximation of f , m = 1 and m = 3. operative adaptive tracking optimal controller and
For m = 1, the observer’s parameters are intelligent observer can quickly bring the robotic
Links’ angles, q1 (t), q2 (t), to the desirable con-
Ω = diag (−10; −6) , Ψ = diag (−25; −9) stant references, r1 (t) and r2 (t). These simulation
Γ2 = 3W0 + Ψ, Γ1 = 3W02 + ΨΓ2 + Ω, (47) results validate the contributions of our paper.
On the other hand, the response of Link 1 and
Γ0 = W03 , W0 = diag (40; 40) 2 angles, q1 (t), q2 (t) of the closed-loop robotic
system with cyclic references are demonstrated in
On the other hand, with m = 3 the observer gains Figures. 7 and 8. As can be observed from Figures
are obtained as follows 7 and 8, our tracking Controllers I and I can effec-
tively regulate the robotic Links’ angles to follow
Ω = diag (−10; −6) , Ψ = diag (−25; −9)
the prescribe cyclic references, r(t). Furthermore,
Γ4 = 5W0 + Ψ, Γ3 = 10W0 2 + ΨΓ2 + Ω, under the influence of time-varying disturbances,
Γ2 = 10W0 3 , Γ1 = 5W0 4 , Γ0 = W0 5 , T e , the results in Figures 9 and 10 indicate that by
using the proposed control strategies, the Links’
W0 = diag ([40; 40])
angles,q1 (t), q2 (t) can follow the cyclic references,
(48)
r after some seconds of settling times. These
In this subsection, we have developed two follow-
outcome verifies the merits of our paper.
ing control strategies
 Springer Nature 2021 LATEX template

11

1.2 1.5
r2 (t)
1 1 q (t) - Controller I
2
0.8 0.5 q2 (t) - Controller II
r1 (t)
[rad]

[rad]
0.6 q1 (t) - Controller I 0
0.4 q1 (t) - Controller II
-0.5
0.2
-1
0
-1.5
0 2 4 6 8 10 0 2 4 6 8 10
Time (Seconds) Time (Seconds)
Fig. 5 The response of Link 1 angle, q1 (t) of the closed-
loop systems with constant references (Scenario I1 ) Fig. 8 The response of Link 2 angle, q2 (t) of the closed-
loop systems with cyclic references, r, (Scenario I2 )
1.2
1.5
1 r1 (t)
0.8 1 q1 (t) - Controller I
r2 (t)
0.5 q1 (t) - Controller II
[rad]

0.6 q2 (t) - Controller I [rad]

0.4 q (t) - Controller II 0
2

0.2 -0.5
0 -1
0 2 4 6 8 10
-1.5
Time (Seconds) 0 2 4 6 8 10
Fig. 6 The response of Link 2 angle, q2 (t) of the closed- Time (Seconds)
loop systems with constant references (Scenario I1 ) Fig. 9 The response of Link 1 angle, q1 (t) of the closed-
loop systems (Scenario I3 ).
1.5
r1 (t) 1.5
1 q1 (t) - Controller I r2 (t)
q1 (t) - Controller II 1 q2 (t) - Controller I
0.5
0.5 q2 (t) - Controller II
[rad]

0
[rad]

0
-0.5
-0.5
-1
-1
-1.5
0 2 4 6 8 10
-1.5
Time (Seconds) 0 2 4 6 8 10
Fig. 7 The response of Link 1 angle, q1 (t) of the closed- Time (Seconds)
loop systems with cyclic references (Scenario I2 ) Fig. 10 The response of Link 2 angle, q2 (t) of the closed-
loop systems (Scenario I3 ).
In conclusion, our cooperative controllers are
highly effective in the tracking operation of the To validate the observer’s performance in estimat-
robotic systems which demonstrate the advan- ing the lumped disturbances, fˆ we have under-
tages of our proposed control strategy. taken Scenario II when the observer works inde-
pendently. In this subsection, the LQR controller
4.3 Observer validation will be removed leading to τ ln in (21) becomes
 Springer Nature 2021 LATEX template

zero. Then, two control strategies are designed 30

include f2 (t)
20 f (t)
20 m2
f1 (t)

[rad/s2]
f (t)
10 m1
10
[rad/s2]

0
0

-10
-10
0 2 4 6 8 10
-20
Time (Seconds)
0 2 4 6 8 10 Fig. 12 Estimated results of f2 and fm2 with Controller
Time (Seconds) III under constant disturbances (Scenario II1
Fig. 11 The Responses of f1 and fm1 with controller III 40
under constant disturbances (Scenario II1 ) f1 (t)
20 fm1 (t)

ˆ Adaptive control strategy 3 (called Controller

[rad/s2]
0
III) is implemented based on the intelligent
observer where the parameters for Taylor -20
approximation of f is m = 1.
ˆ Adaptive control strategy 4 (called Controller -40
IV) is built according to the estimation of the
intelligent observer only where the parameters -60
0 2 4 6 8 10
for Taylor approximation of f is m = 3. Time (Seconds)
In Scenario II, for demonstrative purpose, sys- Fig. 13 Estimated results of f1 and fm1 with Controller
IV under constant disturbances (Scenario II1 )
tem references, r and system disturbance, Te are
taken from Scenario I. Two cases of simulations 80
will be undertaken as follows: in Scenario II1 , f2 (t)
60 fm2 (t)
the system disturbances are incremental change
signals as in Scenario I2 whereas the system dis- 40
[rad/s2]

turbance in Scenario II2 are cyclic signals (see

Scenario I3 ). The designed observers reconstruct 20
T
the time-varying signals, fˆ = [fm1 fm2 ] , of the 0
T
unmeasured time-varying functions, f = [f1 f2 ] . -20
In consequence, the estimations, fˆ are used for
the implementation of adaptive control strategies -40
0 2 4 6 8 10
3 and 4 (controllers III and IV). Time (Seconds)
As can be seen from Figures 11-18, the lumped
Fig. 14 Estimated results of f2 and fm2 with Controller
disturbances f is time-varying and the observers IV under constant disturbances (Scenario II1 )
give a close estimation to f . Figures 13-14 and 17-
18 show that the observer with m = 3 gives better
simulation results in this subsection have shown
approximation efficiency than m = 1. So it can
the effectiveness of the observer.
be say that the parameter m affects the accuracy
of the observer, especially when the disturbances
Te affecting the system is complex. However, just 5 Conclusion
with m = 3, the observer has performed excel-
In this paper, an adaptive cooperative control
lent approximation to lumped disturbances even
regime comprising an intelligent observer and an
though input disturbances is time-varying. The
 Springer Nature 2021 LATEX template

13

10 80
f2 (t)
5 60 f (t)
m2
0 40
[rad/s2]

[rad/s2]
-5 20
-10 0
f1 (t)
-15 -20
fm1 (t)
-20
0 2 4 6 8 10 0 2 4 6 8 10
Time (Seconds) Time (Seconds)
Fig. 15 Estimated results of f1 and fm1 with Controller Fig. 18 Estimated results of f2 and fm2 with Controller
III under under time-varying disturbances (Scenario II2 ) IV under time-varying disturbances (Scenario I3 )

30
f2 (t) put the unfavourable system disturbance and non-
fm2 (t) linear components into a unique lumped distur-
20
bance vector which are time-varying functions.
[rad/s2]

Accordingly, an intelligent observer with the tech-

10 nique of Taylor approximation has been derived to
generate the estimation of the time varying func-
0 tions. By using the available estimated signals, an
adaptive observer-based controller can be built to
-10
attain control requirements of closed-loop robotic
0 2 4 6 8 10 systems. On the other hand, with the availability
Time (Seconds) of the designed observer, the influence of lumped
Fig. 16 Estimated results of f2 and fm2 with Controller disturbance has been eliminated, and the original
III under under time-varying disturbances (Scenario II2 ) non-linear system is converted to a linear form
which is convenient in designing an extra feed-
20 back controller to improve the performance of the
systems. Therefore, in this paper, by using tech-
0 niques of linear quadratic regulation, an additional
optimal controller has been derived to cooper-
[rad/s2]

-20
atively works with the adaptive observer-based
controller to achieve better performance of the
f1 (t)
robotic systems. A comprehensive simulation has
-40
been conducted on the model of a planar robotic
fm1 (t)
system with two degrees of freedom to validate the
-60
0 2 4 6 8 10 effectiveness of our control regime.
Time (Seconds)
Fig. 17 Estimated results of f1 and fm1 with Controller References
IV under time-varying disturbances (Scenario I3 )
[1] J. Mattila, J. Koivumäki, D. G. Caldwell and
optimal feedback controller has been presented to C. Semini, A Survey on Control of Hydraulic
achieve the tracking control purpose and restore Robotic Manipulators With Projection to
the stability of non-linear robotic systems sub- Future Trends, IEEE/ASME Transactions on
jected to disturbances. Firstly, a model of robotic Mechatronics, 22 (2017) 669–680.
system has been provided and restructured to
 Springer Nature 2021 LATEX template

[2] J. Zhai and G. Xu, A Novel Non-Singular Ter- Control of Industrial Robotic Manipulators,
minal Sliding Mode Trajectory Tracking Con- IEEE Access, 9 (2021) 31921 - 31931.
trol for Robotic Manipulators, IEEE Trans-
actions on Circuits and Systems II: Express [11] J. Zhai, G. Xu, A Novel Non-Singular Termi-
Briefs, 68 (2021) 391-395 nal Sliding Mode Trajectory Tracking Con-
trol for Robotic Manipulators, IEEE Trans-
[3] I. David, G. Roble, PID control dynamics of actions on Circuits and Systems II: Express
a Robotic arm manipulator with two degrees Briefs, 68(1) (2020) 391 - 395.
of Freedom, Control de Procesos y Robotica
(2012), (2012) 1-7. [12] M. Rahmani, H. Komijani, M. H. Rahman,
New Sliding Mode Control of 2-DOF Robot
[4] K. Jayaswal, D. K. Palwalia, S. Kumar, Per- Manipulator Based on Extended Grey Wolf
formance investigation of PID controller in Optimizer, International Journal of Con-
trajectory control of two-link robotic manip- trol, Automation and Systems, 18 (2020)
ulator in medical robots, Journal of Interdis- 1572–1580.
ciplinary Mathematics, 24 (2021) 467-478.
[13] N. M. H. Norsahperi, K. A. Danapalasingam,
[5] F. Loucif, S. Kechida, A. Sebbagh, Whale An improved optimal integral sliding mode
optimizer algorithm to tune PID controller control for uncertain robotic manipulators
for the trajectory tracking control of robot with reduced tracking error, chattering, and
manipulator, Journal of the Brazilian Soci- energy consumption, Mechanical Systems
ety of Mechanical Sciences and Engineering and Signal Processing, 142 (2020) 106747.
volume, 42(1) (2020).
[14] U. Zakia, M. Moallem, C. Menon, PID-
[6] E. V. L. Nunes, L. Hsu, Global Tracking for SMC controller for a 2-DOF planar robot,
Robot Manipulators using a Simple PD Con- 2019 International Conference on Electrical,
troller Plus Feedforward, 28(1) (2010) 23-34. Computer and Communication Engineering
(ECCE),(2019).

[7] V. Mohan, H. Chhabra, A. Rani, V. Singh, [15] P. R. Ouyang, J. Tang, W. H. Yue, S.

An expert 2DOF fractional order fuzzy Jayasinghe, PD with sliding mode control
PID controller for nonlinear systems, Neu- for trajectory tracking of robotic system,
ral Computing and Applications, 31 (2019) IEEE International Conference on Advanced
4253–4270. Intelligent Mechatronics (AIM), 32(2) (2014)
189-200.
[8] P. J. Gaidhane, M. J. Nigam, A. Kumar,
P. M. Pradhan, Design of interval type-2 [16] P.R. Ouyang, J. Acob, V. Pano, Adaptive PD
fuzzy precompensated PID controller applied plus sliding mode control for robotic manip-
to two-DOF robotic manipulator with vari- ulator, 2016 IEEE International Conference
able payload, ISA Transactions, 89 (2019) on Advanced Intelligent Mechatronics (AIM),
169-185. 2016.

[9] A. T. Azar, F. E. Serrano, Fractional Order [17] A. O. Pizarro-Lerma, R. Garcia-Hernandez,

Two Degree of Freedom PID Controller for v. Santibanez, J. VillalobosChin, Experimen-
a Robotic Manipulator with a Fuzzy Type-2 tal Evaluation of a Sectorial Fuzzy Controller
Compensator, Proceedings of the Interna- Plus Adaptive Neural Network Compensa-
tional Conference on Advanced Intelligent tion Applied to a 2-DOF Robot Manipulator,
Systems and Informatics 2018, (2019) 77-88. IFAC-PapersOnLine, 52(29) (2019) 233-238.

[10] T. N. Truong, A. T. Vo, H. J. Kang, A [18] Q. Zhou, S. Zha,; H. Li, R. Lu, C. Wu,
Backstepping Global Fast Terminal Slid- Adaptive Neural Network Tracking Control
ing Mode Control for Trajectory Tracking for Robotic Manipulators With Dead Zone,
 Springer Nature 2021 LATEX template

15

IEEE Transactions on Neural Networks and [27] Y. Hu, W. Wang, H. Liu, L. Liu, Reinforce-
Learning Systems, 30(12) (2018) 3611 - 3620. ment Learning Tracking Control for Robotic
Manipulator With Kernel-Based Dynamic
[19] A. Jouila, K. Nouri, An adaptive robust non- Model, IEEE Transactions on Neural Net-
singular fast terminal sliding mode controller works and Learning Systems, 31(9) (2019)
based on wavelet neural network for a 2- 3570 - 3578.
DOF robotic arm, Journal of the Franklin
Institute, 357(18) (2020) 13259-13282. [28] Y. Hu and B. Si, A Reinforcement Learn-
ing Neural Network for Robotic Manipulator
[20] A. T. Azar, F. E. Serrano, Fractional Order Control, Neural Computation, 30(7) (2018)
Two Degree of Freedom PID Controller for 1983 - 2004.
a Robotic Manipulator with a Fuzzy Type-2
Compensator, Proceedings of the Interna- [29] Y. Hu, W. Wang, H. Liu and L. Liu,
tional Conference on Advanced Intelligent Reinforcement Learning Tracking Control
Systems and Informatics 2018, (2019) 77-88. for Robotic Manipulator With Kernel-Based
Dynamic Model, IEEE Transactions on Neu-
[21] W. Ha, J. Back, A Disturbance Observer- ral Networks and Learning Systems, 31(9)
based Robust Tracking Controller for Uncer- (2020) 3570 - 3578.
tain Robot Manipulators, International Jour-
nal of Control, Automation and Systems, 16 [30] C. Mahapatra and S. Chauhan, Tracking con-
(2018) 417–425. trol of inverted pendulum on a cart with
disturbance using pole placement and lqr,
[22] W. Ha, J. Back, A Robust Tracking Con- 2017 International Conference on Emerging
troller for Robot Manipulators: Embedding Trends in Computing and Communication
Internal Model of Disturbances, 2019 Interna- Technologies (ICETCCT), (2017).
tional Conference on Robotics and Automa-
tion (ICRA), (2019). [31] J. H. Choi, J. S. Kim, and K. H. Kim,
Robust tracking performance of linear induc-
[23] M. Ramı́rez-Neria, R. Madonski, A. Luviano- tion motor-based automatic picking system
Juárez, Z. Gao, H. Sira-Ramı́rez, Design of using a high-gain disturbance observer, IET
ADRC for Second-Order Mechanical Systems Electric Power Applications, (2016).
without Time-Derivatives in the Tracking
Controller, 2020 American Control Confer- [32] M. Kim, J. H. Kim, S. Kim, J. Sim, and
ence (ACC), (2020). J. Park, Disturbance observer based linear
feedback controller for compliant motion of
[24] D. Shi, J. Zhang, Z. Sun, G. Shen, Y. Xia, humanoid robot, 2018 IEEE International
Composite trajectory tracking control for Conference on Robotics and Automation
robot manipulator with active disturbance (ICRA), (2018).
rejection, Control Engineering Practice, 106
(2021) 104670. [33] V. P. Vu, V. T. Ngo, V. D. Do, D. N.
Truong, T. T. Huynh, and T. D. Do, Robust
[25] M. A. Y. Abdallah, R. Fareh, Tracking mppt observer-based control system for wind
Control of Serial Robot Manipulator using energy conversion system with uncertain-
Active Disturbance Rejection Control, 2019 ties and disturbance, IEEE Access,9 (2021)
Advances in Science and Engineering Tech- 96466–96477.
nology International Conferences (ASET),
(2019). [34] D. Kleinman, On an iterative technique for
riccati equation computations, IEEE Trans-
[26] E. Guech, S. Bouzoualegh, Y. Zennir, S. actions on Automatic Control, 13(1) (1968)
Blažic, MPC Control and LQ Optimal Con- 114–115.
trol of A Two-Link Robot Arm: A Compara-
tive Study, Machines, 6 (2018) 409-414.
 Springer Nature 2021 LATEX template

[35] M. Kim, J. H. Kim, S. Kim, J. Sim, and

J. Park, “Disturbance observer based lin-
ear feedback controller for compliant motion
of humanoid robot,” 2018 IEEE Interna-
tional Conference on Robotics and Automa-
tion (ICRA), 2018.