Article

Application to control high frequency rehabilitation

robot based on ADRC
Quy-Thinh Dao 1 *, Minh-Chien Trinh 2 , Van-Linh Nguyen 3 , and Viet-Cuong Tran 4,
1 Department of Industrial Automation, Hanoi University of Science and Technology, Vietnam
2 Department of Industrial Automation, Hanoi University of Science and Technology, Vietnam
3 Department of Industrial Automation, Hanoi University of Science and Technology, Vietnam
4 Department of Industrial Automation, Hanoi University of Science and Technology, Vietnam
∗ Correspondence: [email protected], Tel:+81-80-9307-9301

Version November 30, 2021 submitted to Journal Not Specified

1 Abstract: Pneumatic Artificial Muscles (PAMs), also known as actuators, are modeled after natural
2 human muscles, they have the ability to stretch, twist and change stiffness. As a result, they
3 are applied in various drive systems in areas such as rehabilitation, robots, industrial actuators,
4 bio-simulators... The technology has great breakthrough potential due to its flexibility, lightness,
5 compactness and smaller power dissipation ratio compared to other actuators. In this article, Active
6 Disturbance Rejection Control (ADRC) is used in the nonlinear model to improve the performance
7 and quality in tracking control performances of the rehabilitation robot.

8 Keywords: Pneumatic Artificial Muscle; Active Disturbance Rejection Control

9 1. Introduction
10 Nonlinear mechatronic systems are now used for many different purposes, bringing many benefits
11 to humans. Among them, PAM stands out with its outstanding features such as: simulation like a
12 human muscle, high performance, compactness and above all, extremely small power consumption.
13 That shows that the practical application of PAM in fields such as medical, industrial, office, etc. is
14 very effective, especially in the field of medical rehabilitation. With the use of the rehabilitation robot
15 introduced in this paper, it is possible to test the effectiveness of the proposed robot as well as the
16 algorithm. However, along with the above advantages, it is difficult to control due to high latency,
17 complex nonlinear components. Those are big problems that need to be addressed
18 To solve the problems on a wide range of controllers are proposed: The PID controller [1] was
19 used by G. Andrikopoulos and associates for the countervailing PAM mechanism with very good
20 results. The adaptive controller [2-5] used for the rehabilitation robot gives very satisfactory results
21 at low frequencies. Fuzzy controller [6-7] The set can predict the mapping between, force, length,
22 change in length and pressure inside the PMA , forward feedback control [8] can monitor the signal
23 fast frequency reference. This paper uses ADRC controller for model recovery robot. This robot has
24 2-DOF simulating human hip and knee joints to help patients practice recovery movements after
25 surgery and injury. With the use of PAM (4 PAM), the control for the robot to act like a human is very
26 complicated. In the previous paper, the robot control at high frequency still has many problems with
27 linear ADRC controller. Print this paper, a differential tracker is proposed and a nonlinear observer is
28 implemented. The comparison between linear and nonlinear observers is also given, along with the
29 Lyapolow control law, the experimental results are very good. The sample signals are all simulated in
30 human footstep, which is very suitable for the patient’s rehabilitation process. Robot control process,
31 measurement angle parameters, control signals are collected through NI Myrio kit with Labview
32 control software. This article is divided into five sections as follows: s rehabilitation process. Robot

Submitted to Journal Not Specified, pages 1 – 12 www.mdpi.com/journal/notspecified

 Version November 30, 2021 submitted to Journal Not Specified 2 of 12

33 control process, measurement angle parameters, control signals are collected through NI Myrio kit
34 with Labview control software. This article is divided into five sections as follows: s rehabilitation
35 process. Robot control process, measurement angle parameters, control signals are collected through
36 NI Myrio kit with Labview control software. This article is divided into five sections as follows:

37 • System model statement

38 • The tracking differentiator (TD).
39 • The nonliear Extended State Obsever .
40 • Controller ADRC
41 • Experimental results.

42 2. 2. System model statement

43 2.1. System structure

44 In this paper, the experimental system shown in Figure 1 is used, including the rehabilitation
45 robot and the suspension system, the control system (laptop, NI myrio toolkit). The robot is controlled
46 by a computer through the NI myrio toolkit. The robot is hung up by a mounting system that adjusts
47 the height to suit the patient. The robot’s skeleton is made of aluminum, designed with 2-DOF to help
48 clamp the patient’s thigh and leg, thereby helping the patient to move. With the help of the slider
49 between the hip joint and the knee joint, the length of the robot can be adjusted flexibly and fixed
50 through the screw. The range of motion of the hip joint is up to 75 degrees and the knee joint can be up
51 to 90 degrees.
52 The robot is designed from 2 pairs of opposing muscles fixed on the frame to help the robot move
53 flexibly. The PAM used in the system has a length of 50 cm and a diameter of 2.5 cm. The pressure in
54 the PAM is adjusted through the pressure regulator valve ITV-2030-212S-X26 of SMC. When there is
55 a change in pressure between the muscles, the muscle angle is measured through the potentiometer
56 WDD35D8 – 360°. The data is brought to NI Myrio-1900 of National Instrument (USA), then the
57 computer will perform the algorithm through Labview software from the collected data. The control
58 signal is then sent through the NI Myrio-1900 to the pressure regulators via analog signals.

59 2.2. System model

60 Mathematics of the model presented in [15] and system model shown in Figure 1
 Version November 30, 2021 submitted to Journal Not Specified 3 of 12

Figure 1. Rehabilitation Robot

61 3. The tracking differentiator (TD)

62 In order to avoid a sudden jump in error between the output signal and the reference of the
63 system. Or create a desirable reference orbit that is physically feasible.
The TD differential tracker is the most obvious part directly inherited from the PID. Because in
PID, D is often impractical due to the physical limitations of the model, which causes a PID controller
to generally use only PI. Normally, when D is used in the PID, the implementation of the differential
component of the signal r(t) is as follows:

1
y(t) ≈ (r (t) − r (t − τ )) ≈ ṙ (t) (3.1)
τ

Where τ is time constant (τ > 0 , τ very small). Set y1 (t) = r (t − τ ) Laplace on both sides:

y1 (s) = e−τs r (s) (3.2)

Use Taylor expansion to approximate:

1
y1 ( s ) ≈ r (s) (3.3)
1 + τs
 Version November 30, 2021 submitted to Journal Not Specified 4 of 12

If the signal r(t) contains a high frequency noise component of n(t) , the signal input becomes r(t) +
n(t), and the delay component r (t − τ ) can be filtered. high number:

1 1
y(t) ≈ (r (t) + n(t) − r (t − τ )) ≈ ṙ (t) + n(t) (3.4)
τ τ
Thus, the differential estimator y(t) is quite sensitive to noise in the signal r(t) because it is amplified by
the factor τ1 . To solve this difficulty, proposed a differential tracker that is resistant to noise :

r (t − τ1 ) − r (t − τ2 )
ṙ (t) ≈ (3.5)
τ2 − τ1

64 With condition: 0 < τ1 < τ2

Because the components in 3.5 all contain delay signals, we can completely use high-frequency
filters, using 3.2 and 3.3 have:

1 1 1
y(s) ≈ ( − )r (s), 0 < τ1 < τ2
τ2 − τ1 τ1 s + 1 τ2 s + 1
sr (s)
⇒ y(s) ≈ , 0 < τ1 < τ2 (3.6)
(τ1 s + 1)(τ2 s + 1)
sr (s)
⇒ y(s) ≈ ,0<τ
(τs + 1)2

Set y(t) = z1 (t), ẏ(t) = z2 (t) get the state space equation:



 ż1 (t) = z2 t
ż2 (t) = − τ12 [z1 (t) − ṙ (t)] − τ2 z2 (t), 0 < τ1 < τ2


 y(t) = z (t)
1
 (3.7)
 ż1 (t) = z2 t


z (t)
⇒ ż2 (t) = −w2 f (z1 (t) − ṙ (t), 2w ), 0 < τ1 < τ2


 y(t) = z (t)
1

65 Inside y(t) follow ṙ (t), ẏ(t) follow r̈ (t), w = τ12 > 0 is tuning parameter, f (.) is a suitable nonlinear
66 function. As mentioned in the previous section, the TD differential tracker 3.7 not only tracks the set
67 signal and filters out the noise, but it can also generate a new reference signal for the model that can
68 be traced to a different set of signals. reasonable way to avoid the setpoint jump in the PID. In other
69 words, in engineering applications, the reference trajectories are designed so that the pattern follows
70 the z1 (t) 3.7, rather than a jumpable r(t) as step function. This makes the reference trajectory smooth.
71 By making the reference trajectory smooth, the control signal can also be made smooth.
The 3.7 model is a special form of the nonlinear TD differential tracker. Although the first evidence
of convergence of the 3.7 model was reported in [8], it was later shown to hold true only for constant
signals. However, the effectiveness of the 3.7 differential tracker has been verified by numerous
numerical experiments and real-world practice [9] [10]. Another indirect but related proof is that the
linear differential (LTD) tracker shown in [11] is indeed convergent:
(
ż1 (t) = z2 (t)
(3.8)
ż2 (t) = −k1 R2 [z1 (t) − r (t)] − k2 Rz2 (t)

72 With k1 > 0, k2 > 0 are constant, R>0 is tunning parameter.

 Version November 30, 2021 submitted to Journal Not Specified 5 of 12

73 3.1. Linear Tracking Differential (LTD)

The differential tracking model was proposed:



 ż1 (t) = z2 t
ż2 (t) = z3 (t) (3.9)

ż3 (t) = −k1 R3 [z1 (t) − r (t)] − k2 R2 z2 (t) − k3 Rz3 (t)



With R>0 is tunning parameter. Set:




 e1 ( t ) = z 1 ( t ) − r ( t )
e2 (t) = z2 (t) − r (˙t) (3.10)

e3 (t) = z3 (t) − r (¨t)



(3.9) become:



 ė1 (t) = e2 (t)
ė2 (t) = e3 (t) (3.11)

e2 ( t ) ...
R3 [−k − k3 e3R(2t) ] − k2 R2 ṙ (t) − k3 Rr̈ (t) −

ė3 (t) = 1 e1 ( t ) − k 2 R r (t)


  
e˙1 (t) e1 ( t )
...
⇒ ė(t) = e˙2 (t) = AR3  e2R(t)  + B[k2 R2 ṙ (t) + k3 Rr̈ (t) + r (t)] (3.12)
   
e˙3 (t) e 3 ( t )
R2
   
0 1 0 0
Where A =  0 0 1  and B =  0  Assume that the reference signal r(t) satisfies
   
−k1 −k2 −k3 −1
supte[0,∞) |r (i) (t)| < M, M > 0 with i=1,2,3 represents the order differential i of the signal r(t), model
(3.12) will be asymptotically stable if A is a Hurwitz matrix. Apply the method of assigning poles to A
whose double poles are s TD = −1 have:

k1 = (−s TD )3 = 1, k2 = 3(s TD )2 , k3 = −3s TD = 3 (3.13)

74 4. Extended State Observer (ESO)

75 The ESO observer is an extension of the state observer in modern control theory. A state observer
76 is a supporting model that provides an estimate of the internal state variable of a given real system
77 based on inputs and outputs. The ESO observer is an improved development based on the ground
78 state observer, which not only estimates the state, but also estimates the "total disturbance". “Total
79 disturbance” can come from unmodeled system dynamics, unknown control coefficients, and external
80 noise, because in the ESO observer, “external disturbance” and state variables of the ESO When the
81 system is estimated simultaneously, we can design the output feedback control law without much
82 dependence on the mathematical model.
 Version November 30, 2021 submitted to Journal Not Specified 6 of 12

83 4.1. Nonliear Extended State Observer

According to [12], the proposed Nonlinear Extended State Observer for a 2∧ nd order nonlinear
object is as follows:

 y(t)− x̂1 (t)  a

  


 x̂˙ 1 (t) = x̂2 (t) + 3ε  2  sign(y(t) − x̂1 (t))
 ε 
 y(t)− x̂1 (t) 2a−1
 
x̂˙ 2 (t) = x̂3 (t) + u(t) + 3  2  sign(y(t) − x̂1 (t)) (4.1)

  3a−2 ε
 y(t)− x̂1 (t) 
x̂˙ 3 (t) = 1ε 

sign(y(t) − x̂1 (t))

 2 
ε

84 Where ε > 0 is the adjustment parameter and a e (0, 1).

85 4.2. Linear Extended State Observer

For use and design, [12] Gao simplified the ESO observer by introducing the LESO linear observer
for quadratic nonlinear systems as follows:



 x̂˙ 1 (t) = x̂2 (t) + a1 (y(t) − x̂1 (t))
x̂˙ 2 (t) = x̂3 (t) + a2 (y(t) − x̂1 (t)) + u(t) (4.2)

x̂˙ 3 (t) = a3 (y(t) − x̂1 (t))



86 Inside ai with i = 1, 2, 3 are the constants we need to choose to:

x̂1 → x1 (t), x̂2 (t) → x2 (t), x̂3 → x3 (t) when t → ∞ Set: xe1 (t) = x̂1 (t) − x1 (t), xe2 (t) = x̂2 (t) −
x2 (t), xe3 (t) = x̂3 (t) − x3 (t), have:



 xė1 (t) = xe2 (t) − a1 xe1 (t)
xė2 (t) = xe3 (t) − a2 xe2 (t) (4.3)


 xė3 (t) = − a3 xe1 (t) − ẋ3 (t)

⇒ xė(t) = Ae
x (t) + B ẋ3 (t) (4.4)
   
− a1 1 0 0
87 Inside: A = − a2 0 1 , B  0  Assumption that | ẋ3 (t)| ≤ M, M > 0 then the linear model
   
− a3 0 0 −1
88 will be globally asymptotically stable if it is a Hurwitz matrix. When have xe1 → 0, xe2 → 0, xe3 →
89 0, when t → ∞
90 ⇒ x̂1 (t) → x1 (t), x̂2 (t) → x2 (t), x̂3 (t) → x3 (t) when t → t The simple thing to chose constan ai
91 when A will have double poles s ESO = −K, K > 0 then:
92 det(sI − A) = s3 − s2 a1 − sa2 − a3 = (s − s ESO )3 = s3 − 3s2 s ESO + 3s(s ESO )2 − (s ESO )3

⇒ a1 = −3s ESO , a1 = 3(s ESO )2 , a3 = −(s ESO )3 (4.5)

So, the linear observer LESO can be rewritten as:


˙ 3
 x̂1 (t) = x̂2 (t) + ε (y(t) − x̂1 (t))


x̂˙ 2 (t) = x̂3 (t) + 3ε (y(t) − x̂1 (t)) + u(t) (4.6)

 x̂˙ (t) = 1 (y(t) − x̂ (t))

3 ε3 1

93 Inside: ε = −1 > 0 is the parameter

s ESO
 Version November 30, 2021 submitted to Journal Not Specified 7 of 12

94 5. Controller ADRC

95 5.1. Feedback Controller ADRC

The last important part of the ADRC controller is the feedback control based on the TD differential
monitor and the ESO extended observer. In the feedback loop, the unknown components of the model
called "total noise" will be removed using the ADRC controller’s estimation ability based on the ESO
observer. For simplicity, the ADRC controller for the 2nd order SISO system is as follows:



 ẋ1 (t) = x2 (t)
ẋ2 (t) = f (t, x1 (t), x2 (t), w(t)) + b(w(t))u(t) (5.1)


 y = x1 ( t )

Inside: x (t) = ( x1 (t), x2 (t))t is state variable, y(t) is output, u(t) is control signal, f (.) is model
function of the system, b(.) is is the gain of the control signal also has no unknown component, w(t) is
the disturbance acting on the model. Set b(w(t)) = b0 + ∆b(w(t)) have:


 ẋ1 (t) = x2 (t)

ẋ2 (t) = f (t, x1 (t), x2 (t), w(t)) + ∆b(w(t))u(t) + b0 u(t)


| {z } (5.2)
x3 ( t )




y = x1 ( t )


96 The control goal of the ADRC is based on a state feedback controller in which the state variables of the
97 model are estimated, so that the output y(t) follows the reference trajectory r (t) , and at the same time
98 x2 (t) follows ṙ (t).
To have an overview of the ADRC controller, we will recall its main components. The first
component is the differential tracker TD for the purpose of constructing a new reference orbital z(t)
and its differentials z(i) (t), i = 2, 3 through the original reference orbital r (t) :



 ż1 (t) = z2 (t)
ż2 (t) = z2 (t) (5.3)

z2 ( t ) z3 ( t )
ż3 (t) = R3 Ψ(z1 (t) − r (t),

R , R2 )


99 Where R is tunning parameter, the construction of this new reference orbit will make the system
100 physically easier to control, avoiding jump signals.
The second component is the ESO observer, which is used to estimate state variables x1 (t), x2 (t)
and "total disturbance" x3 (t) = f (t, x1 (t), x2 (t), w(t)) + ∆b(w(t))u(t) of the model:



 x̂˙ 1 (t) = x̂2 (t) + a1 g1 (y(t) − x̂1 (t))
x̂˙ 2 (t) = x̂3 (t) + b0 u(t) + a2 g2 (y(t) − x̂1 (t)) (5.4)

x̂˙ 3 (t) = a3 g3 (y(t) − x̂1 (t))



101 Where y(t) = x1 (t), g(.) is estimator function, ai (t), i = 1, 2, 3 is tunning parameter.
The final component of the ADRC controller design a feedback controller based on the ESO
observer, the control law is as follows:

1
u(t) = [ ϕ( x̂ (t) − z(t)) + z3 (t) − x̂3 (t)] (5.5)
b0
 Version November 30, 2021 submitted to Journal Not Specified 8 of 12

Inside: x̂ (t) = ( x̂1 , x̂2 ) and x̂3 are state variables of 5.1 function. z(t) = (z1 , z2 ) and z3 are state variables
of 5.4. x̂3 serves to cancel the x3 “total disturbance” effect of the object model. Function ϕ(.) is selected
accordingly to the following model asymptotic stabilization:
(
ė1 (t) = e2
, ϕ(0, 0) = 0 (5.6)
ė2 (t) = ϕ(e1 , e2 )

102 Inside: e1 = x1 − z1 , e2 = x2 − z2 . The goals of theADRC controller are state variables e = (e1 , e2 ) of
 xi (t) → x̂i (t) when t → ∞, i = 1, 2, 3


103 function 5.6 converges to the origin O = (0, 0) or xi (t) → zi (t) when t → ∞, i = 1, 2 .

 z (t) → r (t) in [τ, ∞], τ > 0

1
104 Figure 2 demonstrates the model of the ADRC controller.

Figure 2. ADRC controller model

105 5.2. Control law according to Lyapunov

106 There are many ways to choose a function ϕ(.) for (5.6) asymptotically stable to the origin:
(
ė1 (t) = e2
107 Considering the system: , ϕ(0, 0, 0) = 0 have Lyapunov function:
ė2 (t) = ϕ(e1 , e2 )
108 V (e1 , e2 ) = 21 [(e1 + e2 )2 + e12 ] > 0 , ∀ e1 > 0, e2 > 0 and V (0, 0) = 0
109 ⇒ V̇ = (e1 + e2 )(ė1 + ė2 ) + e1 ė2 = −e12 + (e1 + e2 )(e2 + e1 + ϕ(e1 , e2 ))
Function (5.6) will be asymptotically stable if V̇ = −k(e1 + e2 )2 , k > 0 or V̇ < 0, have: (e2 + e1 +
ϕ(e1 , e2 )) = −k (e1 + e2 )
⇒ ϕ(e1 , e2 ) = (−k − 1)e1 + (−k − 1)e2 (5.7)

ADRC feedback controller is selectable as follows:

1
u(t) = [(−k − 1)( x̂1 − z1 ) + (−k − 1)( x̂2 − z2 ) + z3 (t) − x̂3 (t)] (5.8)
b0

110 6. Experimental results

111 To prove the efficiency of the given algorithm. Applying control on a rehabilitation robot model
112 with a sample trajectory based on human steps. The system signals are sampled as T (s) = 5(ms). The
113 data is retrieved in about 4 cycles. All data exported from LabVIEW were processed using MATLAB
114 software (R2019a). Two sets of linear and nonlinear observers are compared. The results are shown in
115 Figure 3. The parameters of the two sets are shown in the table below:
 Version November 30, 2021 submitted to Journal Not Specified 9 of 12

Table 1. Parameters in Lyapunov controller with linear observer.

Parameters P0 Rh Rk kh kk bh bk εk εh Frequency
Value 2.5 25 25 35 30 15 90 0.015 0.01 0.5

Table 2. Parameters in Lyapunov controller with non linear observer.

Parameters P0 Rh Rk kh kk bh bk εk εh ah ak Frequency
Value 2.5 25 25 35 30 15 90 0.015 0.01 0.9 0.9 0.5

116 6.1. Results

Compared Knee
0
Desired
-5 Measured LESO
Knee Joint Angle (Degree)

Measured NLESO
-10

-15

-20

-25

-30

-35

-40
0 1 2 3 4 5 6 7 8
Time (s)
(a)

Compared Hip
20
Desired
15 Measured LESO
Measured NLESO
Hip Joint Angle (Degree)

10

-5

-10

-15
0 1 2 3 4 5 6 7 8
Time (s)
(b)

Figure 3. Experimental results : (a) Compared tracking trajectory between LESO and NLESO (Knee).
(b) Compared tracking trajectory between LESO and NLESO (Hip).
Version November 30, 2021 submitted to Journal Not Specified 10 of 12

Assess LESO Assess NLESO

20 20
x1 x1
x1 LESO x1 NLESO

15 15

10 10
Hip Joint Angle (Degree)

Hip Joint Angle (Degree)

5 5

0 0

-5 -5

-10 -10

-15 -15
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Time (s) Time (s)

(a) (b)

Assess LESO Assess NLESO

0 0
x1 x1
x1 LESO x1 NLESO
-5 -5

-10 -10
Knee Joint Angle (Degree)

Knee Joint Angle (Degree)

-15 -15

-20 -20

-25 -25

-30 -30

-35 -35

-40 -40
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Time (s) Time (s)

(c) (d)

Figure 4. Experimental results of observed variables : (a) Compared x1 and x̂1 (Hip LESO). (b)
Compared x1 and x̂1 (Hip NLESO). (c) Compared x1 and x̂1 (Knee LESO).(d) Compared x1 and x̂1
(Knee NLESO)

Assess LESO Assess NLESO

60 60
x2 x2
x2 LESO x2 NLESO

40 40

20 20
Hip Joint Angle (Degree)
Hip Joint Angle (Degree)

0 0

-20 -20

-40 -40

-60 -60

-80 -80
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Time (s) Time (s)

(a) (b)

Assess LESO Assess NLESO

200 150
x2 x2
x2 LESO x2 NLESO

150
100

100
Knee Joint Angle (Degree)
Knee Joint Angle (Degree)

50

50

-50
-50

-100
-100

-150 -150
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Time (s) Time (s)

(c) (d)

Figure 5. Experimental results of observed variables : (a) Compared x2 and x̂2 (Hip LESO). (b)
Compared x2 and x̂2 (Hip NLESO). (c) Compared x2 and x̂2 (Knee LESO).(d) Compared x2 and x̂2
(Knee NLESO)
 Version November 30, 2021 submitted to Journal Not Specified 11 of 12

Assess LESO Assess NLESO

1500 1500
x3 x3
x3 LESO x3 NLESO

1000 1000
Hip Joint Angle (Degree)

Hip Joint Angle (Degree)

500 500

0 0

-500 -500

-1000 -1000

-1500 -1500
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Time (s) Time (s)

(a) (b)

Assess LESO Assess NLESO

8000
15000
x3
x3 x3 NLESO
6000
x3 LESO
10000
4000
Knee Joint Angle (Degree)

2000

Knee Joint Angle (Degree)

5000
0

0 -2000

-4000
-5000
-6000

-8000
-10000

-10000

-15000
-12000
0 1 2 3 4 5 6 7 8 0 1 2 3 4 5 6 7 8
Time (s) Time (s)

(c) (d)

Figure 6. Experimental results of observed variables : (a) Compared x3 and x̂3 (Hip LESO). (b)
Compared x3 and x̂3 (Hip NLESO). (c) Compared x3 and x̂3 (Knee LESO).(d) Compared x3 and x̂3
(Knee NLESO)

Compared RMSE
1.6

1.4

1.2

1
RMSE

0.8

0.6

0.4

0.2

0
RMSE-LESO RMSE-NLESO
(a)

Figure 7. Compared RMSE between LESO and NLESO

117 Based on the experimental results, we can see: the NLESO observer is better than the LESO
118 observe, RMSE LESO = 1.598 and RMSE NLESO = 1.271.
 Version November 30, 2021 submitted to Journal Not Specified 12 of 12

119 7. Conclusion
120 In this paper, the ADRC control strategy has been used to improve the control quality of the
121 Rehabilitation Robot. First, the selection of the Tracking Differentiator helps to avoid the jump between
122 the sample signal and the output signal. Second, using two sets of linear and nonlinear observers
123 makes it possible to compare the efficiency of the two sets, the results have proven that the nonlinear
124 observer is better than the linear observer. Third, the control law Lyapunov is used in has shown
125 its effectiveness in controlling the rehabilitation robot in practice. However, the improvement of
126 long walks is still very limited and difficult to respond at high frequencies. This can be improved
127 by adjusting the mechanical system and developing new observers, which will be developed in the
128 future.
129 List of Acronyms:
PAMs Pneumatic Artificial Muscle
ADRC Active Disturbance Rejection Control
TD The Tracking Differentiator
130
LTD Linear tracking Differentiator
ESO Extended State Observer
NLESO Nonliear Extended State Observer

131 References
132 1. G. Andrikopoulos, G. Nikolakopoulos, and S. Manesis, “Advanced Nonlinear PID-Based Antagonistic
133 Control for Pneumatic Muscle Actuators”, IEEE Trans. Ind. Electron., vol. 61, no. 12 Dec. 2014
134 2. Q. Ai, C. Zhu, J. Zuo, W. Meng, Q. Liu, S. Q. Xie, and M. Yang, “Disturbance-estimated adaptive backstepping
135 sliding mode control of a pneumatic muscles-driven ankle rehabilitation robot”, Sensors, vol. 18, no. 1, p. 66
136 Dec. 2017
137 3. S.-I. Han and J. M. Lee, “Adaptive fuzzy backstepping dynamic surface control for output-constrained
138 non-smooth nonlinear dynamic system”, Int. J. Control Autom. Syst., vol. 10, no. 4, pp. 684–696 Aug. 2012
139 4. V. T. Yen, W. Y. Nan, P. Van Cuon, N. X. Quynh, and V. H. Thich, “Robust adaptive sliding mode control for
140 industrial robot manipulator using fuzzy wavelet neural networks”, Int. J. Control Automat. Syst., vol. 15, no.
141 6, pp. 2930–2941 Dec. 2017 Dec. 2017
142 5. L. Zhao, H. Cheng, and Y. Xia, “Angle Tracking Adaptive Backstepping Control for a Mechanism of
143 Pneumatic Muscle Actuators via an AESO”, IEEE Trans. Ind. Electron., vol. 66, no. 6 Jun. 2019
144 6. ] S.Q. Xie, and P.K Jamwal, “An iterative fuzzy controller for pneumatic muscle driven rehabilitation robot”,
145 Expert Systems with Applications, vol. 38, no. 7, p. 8128-8137 Jul. 2011
146 7. C. Chen, J. Huang, D. Wu, and Z. Song, “T-S Fuzzy Logic Control with Genetic Algorithm Optimization for
147 Pneumatic Muscle Actuator”, Int. Conference on Modeling, Identification and Control Jul. 2018
148 8. Quy-Thinh Dao, Shin-ichiroh Yamamoto, " Trajectory Tracking Control of a Robotic Orthosis for Gait
149 Rehabilitation: A Feedforward-Feedback Control Approach", The 10th Biomedical Engineering International
150 Conference (BMEiCON-2017), Hokkaido, Japan, 2017.
151 9. J. Q. Han, Control theory: the doctrince of model or the doctrine of control? J. Systems Sci. Math. Sci., 9:4
152 (1989), 328-335, (in Chinese).
153 10. J. Q. Han, A new type of controller: NLPID. Control and Decision, 6:6(1994), 401-407.
154 11. J. Q. Han, The improvement of PID control law by using nonlinearity. Information and Control, 24:6 (1995),
155 356-364.
156 12. Bao-Zhu Guo,Zhiliang Zhao,"Active disturbance rejection control: Theoretical perspectives"Communications
157 in Information and SystemsVolume 15, Number 3, 361–421, 2015
158 13. Guo, Bao-Zhu,Zhao, Zhi-Liang,"Active disturbance rejection control for nonlinear systems"
159 14. J.Q. Han, Active Disturbance Rejection Control–The Technique for Extimating and Compensating
160 Uncertainties, National Defence Industry Press, Beijing, 2008 (in Chinese).
161 15. Wen Yu and Jos Antonio Heredia,"PD Control of Robot with RBF Networks Compensation",2000
162 16. Quy-Thinh Dao, Duc-Canh Nguyen1, Minh-Chien Trinh1,Van-Vuong Dinh2, and Trong-Hieu Do,"Active
163 Disturbance Rejection Control of anAntagonistic Muscle", 2021
 Version November 30, 2021 submitted to Journal Not Specified 13 of 12

164 © 2021 by the author. Submitted to Journal Not Specified for possible open access publication
165 under the terms and conditions of the Creative Commons Attribution (CC BY) license
166 (http://creativecommons.org/licenses/by/4.0/).