Online-Journals.

org

JOE International Journal of

Online and Biomedical Engineering
iJOE | eISSN: 2626-8493 | Vol. 20 No. 12 (2024) |

https://doi.org/10.3991/ijoe.v20i12.49159

PAPER

Adaptive Control Using Radial Basis Function Neural

Networks for Pneumatic Artificial Muscle Systems

Minh-Duc Duong, ABSTRACT

Viet-Thanh Nguyen, This study introduces a novel adaptive controller employing neural networks, particularly
Quy-Thinh Dao(*) radial basis function (RBF) algorithms, to enhance the control performance of pneumatic
artificial muscle (PAM)-based systems. The proposed controller seeks to address the nonlin-
Hanoi University of
earities and hysteresis inherent in PAM-based systems by integrating neural approximation.
Science and Technology,
Experimental testing and comparisons with conventional controllers are conducted using an
Hanoi, Vietnam
antagonistic configuration of PAMs. The results illustrate the precision and reliability of the
[email protected] proposed controller, suggesting potential for future advancements in trajectory tracking con-
trol of PAM-based systems.

KEYWORDS
pneumatic artificial muscle (PAM), radial-basis-function neural network (RBFNN), neural
approximation, adaptive control

1 INTRODUCTION

Pneumatic artificial muscles (PAMs) are a type of soft actuator that has been
widely used in biorobot [1, 2], medical device applications [3–5], and industrial [6, 7]
due to their unique advantages. The most notable advantage of PAMs is their high
power-to-weight ratio, which allows them to generate large forces while remaining
lightweight. Additionally, PAMs are soft and flexible, allowing for safe human-robot
interaction and stable motion speed. PAMs can also adapt to different loads, provid-
ing adaptable compliance. Furthermore, PAMs possess properties similar to those
of human muscle, making them suitable for bio robotic applications that imitate
the morphology and physiology of humans. Another advantage is that it is rela-
tively low-cost to fabricate and easy to manufacture. Due to these advantages, PAMs
have been widely used in rehabilitation systems where precision, adaptability, and
human-like behavior are needed.
Despite the many advantages of PAMs, several disadvantages make model-
ing and controlling them difficult. One of the main disadvantages of PAMs is their

Duong, M.-D., Nguyen, V.-T., Dao, Q.-T. (2024). Adaptive Control Using Radial Basis Function Neural Networks for Pneumatic Artificial Muscle
Systems. International Journal of Online and Biomedical Engineering (iJOE), 20(12), pp. 109–123. https://doi.org/10.3991/ijoe.v20i12.49159
Article submitted 2024-03-17. Revision uploaded 2024-07-05. Final acceptance 2024-07-05.
© 2024 by the authors of this article. Published under CC-BY.

iJOE | Vol. 20 No. 12 (2024) International Journal of Online and Biomedical Engineering (iJOE) 109
Duong et al.

nonlinearity, making it difficult to model their behavior accurately. Additionally,

PAMs exhibit hysteresis, a non-ideal behavior resulting in a lag pertaining to the sys-
tem’s input and output. Another major disadvantage is the uncertainty of the model
parameters, which makes it difficult to design controllers that can adapt to different
conditions. Therefore, despite the many advantages of PAMs, the difficulties in mod-
eling and controlling them have been a significant obstacle to their widespread use
in practical applications.
Despite the difficulties in modeling and controlling PAMs, various control meth-
ods have been developed to overcome these challenges and achieve high perfor-
mance in PAM control. The classical proportional integral derivative (PID) controller
and its developed versions quickly became prevalent in various research domains.
Based on PID control theory, research [8] presented a position control method for
an artificial muscle that only used one pressure sensor. A pioneering PID-based
controller, as presented in [9], is introduced to improve tracking performance and
minimize hysteresis. Nevertheless, these PID controllers are only suitable for test-
ing the mechanical structure and basic tracking ability of low-quality artificial mus-
cle systems.
It is acknowledged that uncertainties and disturbances of pneumatic actuators
like PAMs can be effectively solved by sliding mode control (SMC) and its enhanced
version [10–14]. In addition, it essentially designs the controller in the discrete-time
domain to apply it to digital controllers. For those reasons, several studies on dis-
crete-time SMC have been conducted within the literature, focusing on various PAM-
based systems, including investigations of equivalent control methodologies [15, 16]
and reaching-law-based approaches [17, 18]. However, the chattering phenomenon
is a major issue with any version of SMC specified in PAM-based actuators when
tracking low-frequency trajectories.
Active disturbance rejection control (ADRC) methods have been considered to
deal with the inherent uncertainties and disturbances of PAM-based systems. The
fundamental idea of ADRC is to first consider the plant as a linear system and then
use an extended state observer to estimate and reject the ignored dynamics and
external disturbances as total disturbances. This approach does not require exten-
sive knowledge of the system model or its disturbances, making it a useful controller
for many applications. Both linear and nonlinear ADRCs have been used in litera-
ture [19–21]. However, research focusing on the practical aspects of the nonlinear
ADRC scheme has been limited.
Artificial intelligence (AI) is widely used in various fields [22–24], includ-
ing biomedical engineering [25–28]. Among AI techniques, fuzzy control, genetic
algorithms, and neural networks are commonly employed in the development of
intelligent control systems, showcasing their effectiveness over traditional control
methods [29–31]. The use of fuzzy logic with conventional controllers, such as fuzzy-
PID [32, 33] and fuzzy-SMC [34–36], has been widely implemented and achieved
good results in the past, but these solutions are not flexible enough for many dif-
ferent application cases. As a result, learning algorithms tailored for fuzzy systems
facilitate adaptive tuning of the system’s parameters according to the provided sig-
nal patterns. Adaptive control is another approach we examine, which is employed
by a controller that must adapt to a controlled system with variable or originally
unclear characteristics [37]. The radial-basis-function neural network (RBFNN) is a
popular choice among fuzzy logic neural networks and has been successfully used
in various applications. Due to its capability to approximate uncertainty in com-
plex model-uncertain systems with unidentified functions, it is viewed as a highly

110 International Journal of Online and Biomedical Engineering (iJOE) iJOE | Vol. 20 No. 12 (2024)
 Adaptive Control Using Radial Basis Function Neural Networks for Pneumatic Artificial Muscle Systems

effective solution for addressing diverse control challenges, notably dynamic uncer-
tainty [38]. Many neural network controllers are used with adaptive control tech-
niques for nonlinear systems to achieve high-precision performance [39–41]. With
the outstanding capabilities mentioned above of RBFNN, which have the potential to
mitigate the shortcomings of conventional controllers, we propose utilizing RBFNN
to develop an adaptive controller aimed at enhancing control quality for the PAM
system. This study employs indirect adaptive control methodologies, utilizing esti-
mated parameters to calculate the essential controller parameters [42]. Additionally,
we integrate the power of the RBF neural network to estimate uncertain elements
in the system.
In summary, this paper presents several noteworthy contributions. Firstly, it
introduces an adaptive controller, custom-designed to effectively handle antagonis-
tic configurations of PAMs through neural approximation techniques. Secondly, the
utilization of RBF algorithms is showcased to accurately approximate system uncer-
tainties and determine the necessary controller parameters. Lastly, the practicality
and potential impact of the proposed method are substantiated through experimen-
tal results conducted under diverse conditions. These results collectively underscore
the effectiveness and suitability of the presented approach in the context of rehabil-
itation applications.

2 SYSTEM MODELING

A full representation of the pneumatic experimental platform is depicted in

Figure 1. In the working object, a pair of PAMs is configured in an antagonistic
arrangement. The structure dimensions are 25 mm in diameter and 400 mm in
nominal length. The control valves (SMC, ITV2030-212S-X26) change and measure
the pressure inside both PAMs. When there exists a pressure differential between
the two PAMs, causing the pulley wheel to rotate, we employ a potentiometer (model
WDD35D8T) to measure and record the resulting angle. This information is then sent
to an embedded controller, specifically the MyRIO 1900 manufactured by National
Instruments. This allows us to perform system adjustments via the computer.

Fig. 1. The PAM-based actuator experimental platform

iJOE | Vol. 20 No. 12 (2024) International Journal of Online and Biomedical Engineering (iJOE) 111
Duong et al.

Due to the alignment in research focus, the mathematical model for the PAM-
based system is derived from prior research conducted in our laboratory [18].

x(t )  2 x (t )  1 x(t )  1u   0 (1)

where x is the chosen manipulated variable of the deviation angle, and u is the
voltage control signal, which expresses the pressure change in PAMs. The parame-
ters are described as follows:

 2F  2K L  R( B  B )  R
   1 1 0 1e 1f 

 1 J

 ( F1 PAP  K1 PAP L0 ) R
 0  J

  (2K 0  2K1 P  K1 PAP ) R2
 1 J

   B0e  B0 f  ( B1e  B1 f ) P0  B1e PAP )  R2
   
 2 J

where R and J are the pulley’s radius and inertia moment; L0 and P0 are the initial
value of upper PAM’s length and pressure; PAP is the initial different pressure of
both PAMs; Ki and Fi (i = 0,1) represent the spring and contractile elements of PAM’s
model; Bie and Bif are the components of the PAM’s damping coefficient. The nota-
tions e and f signify inflation and deflation of the muscle, respectively. For a more
detailed understanding of the mathematical model of a single PAM, please consult
the report [11].
The mathematical model’s identified parameters a0, a1, b1, b2 are synthesized
in Table 1.

Table 1. Identified model parameters

Parameters a0 a1 b1 b2
Value 205 18.01 -4.83 7.35 × 10-4

3 CONTROL DESIGN

The objective of this study is to introduce an adaptive controller utilizing online

RBF neural approximation. This approach aims to enhance control precision and
adjust to variations in parameters. The development of the adaptive law was
grounded in the principles of Lyapunov stability theory, ensuring the attainment
of stable closed-loop systems. Figure 2 provides a visual representation of the block
diagram for the envisioned closed-loop neural-based control system, demonstrating
the integration of the RBFNN in generating the control signal. The variable y rep-
resents the recorded angle, while yd corresponds to the intended or reference signal.
Due to the closed-loop control arrangement, the tracking error is fed back to both
the controller and the adaptive mechanism, which includes the RBFNN. Within the
adaptive cluster, the function f (x), with uncertainties, undergoes approximation
before being transmitted to the controller. The control output u is directed toward
two control valves, thereby altering the pressures P1 and P2 of the two PAMs arranged
antagonistically.

112 International Journal of Online and Biomedical Engineering (iJOE) iJOE | Vol. 20 No. 12 (2024)
 Adaptive Control Using Radial Basis Function Neural Networks for Pneumatic Artificial Muscle Systems

Fig. 2. System schematic of the proposed controller

A system with two PAMs arranged in an antagonistic configuration represents a

single input, single output system. To facilitate the construction of the control algo-
rithm, according to mathematical model (1), we described the PAM system in the
controllable canonical form:

 x  f ( x , x )  gu
 (2)
 y  x

Where, g = a1 is a known function as a result of the identification process,

while f ( x , x )  2 x (t )  1 x(t )   0 is an unknown nonlinear function.
Assume that yd is the desired position signal. Declare e, vector X, and vector
E indicate the tracking errors, the vector of the system’s states, and the vector of
errors’ states:

e  yd  y,X  [ x x ]T ,E  [e e]T

From (2), the inverse-based control algorithm can be developed as follows:

1
u  [  f ( X )  yd  K T E ] (3)
g

Where, K = [kp, kd]T is controller coefficient vector. Substituting equation (3) into
equation (2), we obtain the closed-loop system as follows:

e  kd e  k p e  0 (4)

To ensure the stability of the system and accurate output tracking, the polynomial
s2 + kds + kp = 0 must be Hurwitz. Therefore, the selection of K is made to ensure that
all the roots of the polynomial function lie to the left of the imaginary axis on the
complex plane. However, it is realized that the controller (3) depends on an unknown
factor f (X). Thus, in the next subsection, we will employ the RBFNN to estimate f (X).

3.1 Designing radial basis function neural networks

In this subsection, an RBFNN is selected to approximate the unknown components

of the PAM-based system. Figure 3 illustrates the RBF neural network’s architecture,

iJOE | Vol. 20 No. 12 (2024) International Journal of Online and Biomedical Engineering (iJOE) 113
Duong et al.

which comprises three layers, that has been proposed. For a more thorough expla-
nation, the neural network takes two input nodes to represent the tracking error and
its derivative. The hidden layer comprises m nodes, responsible for capturing the
signal range and computing the weight vector. The output layer has only one node
to present the approximated values. Specifically, Z = [z1 z2]T is the input vector,
h(Z) = [hj]T is the output of the hidden layer, and the specific values of each hj are the
values of the Gaussian functions for jth node in that layer:
 Z c 2

 j 
h j  exp    (5)
b2j
 
Where, cj = [c1j c2j]T denotes the central coordinate value of the jth node, and
bj represents the width value of the Gaussian function for the jth node with j = 1,
2, … , m. The approximation fˆ ( X ) of the uncertain element f (X) of the system serves
as the expected output. The output of the neural network is:
m
fˆ ( X )  W T h( Z )  W h
j 1
j j
(6)

Where, W = [w1, w2, … , wm]T is the weight vector.

Given the nature of RBF neural networks may be utilized to estimate uncertain-
ties in control systems, in [38], the RBF neural network’s input vector is chosen as
the tracking error and the rate of change of it, that is, Z= E= [e e]T, the equation for
utilizing the RBF algorithm to approximate f (X) is outlined below:

 T h( E )
fˆ ( X ) = W (7)

Where, W represents the calculated weight vector, subject to potential adjust-

ments through the adaptive technique, and h(E) is a Gaussian function.

Fig. 3. RBF neural network structure

3.2 Designing control law

The RBF neural network has represented the uncertain nonlinear factor f (X)
from the previous section. As a result, the control law is fully defined:

1 ˆ
u [  f ( X )  yd  K T E ] (8)
g

Where, fˆ ( X ) is the approximated parameter for f (X).

114 International Journal of Online and Biomedical Engineering (iJOE) iJOE | Vol. 20 No. 12 (2024)
 Adaptive Control Using Radial Basis Function Neural Networks for Pneumatic Artificial Muscle Systems

3.3 Designing adaptive law

In this section, the adaptive law is derived using the Lyapunov stability approach to
approximate f (X) with fˆ ( X ) while maintaining the stability of the system. As indicated
in equation (7), the approximation fˆ ( X ) is achieved through a radial basis function
(RBF) neural network. Consequently, the updating rule for W  will be formulated. By
substituting the control law (8) into equation (2), the closed-loop system is expressed as:

e   K T E  [ fˆ ( X )  f ( X )] (9)

To facilitate a more convenient transformation, we define two new matrices:

 0 1  0 
L= ,B =   (10)
−k −kd  1 
 p 

Equation (9) can be rewritten as follows:

.
E   E  B[ fˆ ( X )  f ( X )] (11)

Supposing that f (X) can be calculated by an ideal RBF neural network with ideal
T
weight values W ∗ = w1∗ , w2∗ , , wm*  :

f (X) = W *Th(E) (12)

Then equation (11) becomes:

.
  W * )T h( E )
E   E  B(W (13)

Considering the following Lyapunov function:

1 T 1   W *)
V E PE  (W  W * )T (W (14)
2 2
 − W * signifies the error in parame-
Where, g is a positive numerical value, W
ter estimation, and P is a symmetric and positive definite matrix that fulfills the
Lyapunov equation:

LT P + PL = -Q (15)

Where, Q is an arbitrary 2 × 2 positive definite matrix and L is given by

equation (10).
Differentiating equation (14) with respect to time, we have:
.
1 1 1  
V  E T PE  E T PE  (W  W * )T W (16)
2 2 

By substituting equation (15) into equation (16), one can obtain

.
1 1     E T PBh( E )]
V   E T QE  (W  W * )T [W (17)
2 

Select the adaptation (weight updating) law as:

.
   E T PBh( E )
W (18)

iJOE | Vol. 20 No. 12 (2024) International Journal of Online and Biomedical Engineering (iJOE) 115
Duong et al.

Then,

1
V   E T QE  0 (19)
2

The system is stable in the sense of Lyapunov.

4 EXPERIMENTAL RESULTS

A variety of situations are tested with different trajectories to evaluate the perfor-
mance of the recommended controller. Experiments are carried out in both cases,
without and with loads. The sampling cycle Ts is set at 5 ms throughout each exper-
imental case.
Acknowledging the intricate characteristics of the artificial neural system, we
have opted for a particular setup, specifically a 2-5-1 RBFNN. The neural network’s
input is composed of two nodes, which represent both the tracking error and its
rate of change in control. The hidden layer, identified as hj, comprises five nodes
meticulously selected to encompass the full signal range and facilitate efficient com-
putation of weight vectors. Careful consideration has been given to selecting values
for bj to guarantee optimal performance. Finally, the output layer consists of a sin-
gle node that signifies the estimated parameter of the pneumatic muscle actuator,
denoted as f (X).
RBFNN approximation requires careful selection of the parameters cj and bj to
ensure that the Gaussian function is effectively mapped to the input values. If these
parameters are not chosen appropriately, the RBF network may not be valid. To
avoid this, it is recommended to set the initial values of cj and bj within the range of
the input values for the RBF network. This way, only the weight value needs to be
updated while cj and bj remain fixed. The initial weight value is set to zero, and the
10 5 0 5 10 
parameters cj and bj are designed as [c j ]    , and bj = 2.
20 10 0 10 20 
During the system modeling process, the value of l0 was determined to be 18.01.
Two parameters, kp = 40, kd = 0.1, are chosen according to the rule described in
Section 3. Other parameters were fine-tuned through a meticulous adjustment pro-
20 0 
cess and finally chosen as follows: Q =   , and g = 120.
 0 20 
Because of the likeness in control concepts and research subject matter, all exper-
imental results will be compared with the meticulously studied fuzzy-PID controller
(FPIDC) presented in the work in [43], where the PID controller’s parameters are
adjusted by fuzzy logic.

4.1 Tracking signals without loads

In the subsequent phase of the study, desired trajectories for evaluating the con-
troller’s performance are selected as sinusoidal signals with an amplitude of 40°
and frequencies ranging from 0.2 Hz to 1.0 Hz. Triangular signals featuring identi-
cal amplitudes and frequencies of 0.2 Hz and 0.5 Hz are additionally employed as
desired trajectories. These signals serve the purpose of assessing the controller’s track-
ing performance and its ability to handle various input profiles. The experimental

116 International Journal of Online and Biomedical Engineering (iJOE) iJOE | Vol. 20 No. 12 (2024)
 Adaptive Control Using Radial Basis Function Neural Networks for Pneumatic Artificial Muscle Systems

results, corresponding to the aforementioned signal types, are depicted in Figures 4

and 5, respectively. Tracking performances are depicted in the upper sub-figures,
where the measured trajectories are reflected by the dashed red lines of the FPIDC
and the dashed blue lines of the adaptive RBFNN controller, compared to the green
line representing the desired signals. Meanwhile, the tracking errors are shown in
the lower sub-figures, and Table 2 provides the two controllers’ root-mean-square
tracking error (RMSE) values.

(a) 0.2 Hz (b) 1.0 Hz

Fig. 4. Experiment results when tracking sinusoidal trajectories

(a) 0.2 Hz (b) 0.5 Hz

Fig. 5. Experiment results for tracking triangular trajectories

It is evident from all the experimental outcomes that both the FPIDC and the
adaptive RBFNN controller can track the desired signals effectively. But inspecting
the steady-state errors and RMSE values deeply, we acknowledge that the RBFNN
controller brings better performance, especially when the desired signals have
a high frequency or are no longer smooth, like triangular ones. In the exclusive
assessment of the RBFNN’s performance, we observe a slight degradation in control
performance with increasing trajectory frequency, although still achieving precise
results. When following sine wave signals, the tracking errors are limited to approx-
imately ± 2.0° at low frequencies and never exceed 4.0° at higher frequencies, which
demonstrates commendable tracking performance.
For triangular trajectories, the results show slightly lower precision. However,
it’s worth noting that if the frequency exceeds 0.5 Hz, maintaining performance

iJOE | Vol. 20 No. 12 (2024) International Journal of Online and Biomedical Engineering (iJOE) 117
Duong et al.

becomes challenging, primarily due to the dynamic hysteresis and time-intensive

response characteristics of the pneumatic artificial muscles.

Table 2. Quantitative evaluation of two controllers when tracking periodic signals

Sinusoidal Triangular
Frequency
FPIDC RBFNN FPIDC RBFNN
0.2 Hz 1.20 0.91 1.62 1.31
0.5 Hz 1.84 1.47 2.64 2.10
0.8 Hz 2.04 1.88 – –
1.0 Hz 2.27 2.03 – –

A gait pattern is also used as the desired signal to evaluate the application of PAM
and the proposed controller in rehabilitation. The desired trajectories of hip and
knee joints are programmed based on human movement data from the well-known
textbook [44]. Two frequencies, 0.2 Hz and 0.5 Hz, of the desired signal are used for
this experimental condition. Figure 6 shows the tracking performance of the RBFNN
controller and its counterpart when tracking the 0.2 Hz gait-pattern trajectory.
We have a similar conclusion about the performance with two other types of
desired trajectories: sine and triangle signals. Both controllers can track the desired
signal, and the RBFNN-based one achieves better tracking with lower fluctuation than
the fuzzy-PID one. The quantity evaluation of both controllers in this experiment is
provided in Table 3. We can see that when tracking the hip trajectory, both controllers
present worse performance with RMSE are 2.47° and 1.84° with 0.2 Hz of frequency.
When the frequency increases to 0.5 Hz, these values are 3.27° and 2.28°. The reason
is that there is a big gap between the measured and desired trajectories at the startup.

Table 3. Quantitative evaluation of two controllers when tracking gait pattern signals

Hip Joint Knee Joint

Frequency
FPIDC RBFNN FPIDC RBFNN
0.2 Hz 2.47 1.84 1.45 1.10
0.5 Hz 3.27 2.28 2.84 1.66

(a) Hip Trajectory (b) Knee Trajectory

Fig. 6. Experiment results when tracking human movement’s trajectories

118 International Journal of Online and Biomedical Engineering (iJOE) iJOE | Vol. 20 No. 12 (2024)
 Adaptive Control Using Radial Basis Function Neural Networks for Pneumatic Artificial Muscle Systems

4.2 Tracking signals with loads

To assess the robustness of the proposed control approach, additional loads are
abruptly introduced to the system upon reaching the steady state. Loads of 3 kg and
5 kg are suspended by a rope and lifted by the operator. Subsequently, the system ini-
tially tracks a sinusoidal signal with an amplitude of 40° degrees and a frequency of
0.5 Hz. At the 4-second mark from the system reaching the steady state, the operator
releases the load, introducing it as an external disturbance. The experimental results
are visualized in Figure 7. It’s noticeable that the tracking performance experiences
fluctuations when the impact takes place. Eventually, both control strategies recover
from this fluctuation. Despite similar overall results, the RBF neural network exhib-
its a faster return to the intended trajectory compared to the comparative strategy.

(a) Load m = 3 kg (b) Load m = 5 kg

Fig. 7. Experiment results for tracking a 40° amplitude and 0.5 Hz frequency
sinusoidal trajectory with a load

The change in the applied weight notably impacts the performance, which directly
affects the system. As weight increases, achieving an optimal tracking trajectory
becomes more challenging and time-consuming. With a 3 kg load, it takes approx-
imately 6.0 seconds to restore the desired stability, while a heavier load requires
8.0 seconds (equivalent to 3 and 4 cycles within this 0.5 Hz frequency experimental
setting). The experimental results align with the rehabilitation criteria established
for the commercial LOKOMAT system [45]. The above evaluations are proved in
graphs, and the RMSE values are shown in Table 4.

Table 4. Quantitative evaluation of two controllers when tracking gait pattern signals

RMSE
Load Weight
FPIDC RBFNN
3 kg 1.94 1.51
5 kg 2.27 1.77

5 CONCLUSION

This study introduces an adaptive controller utilizing RBFNN approximation for

an antagonistic configuration of two pneumatic artificial muscles. The integration of

iJOE | Vol. 20 No. 12 (2024) International Journal of Online and Biomedical Engineering (iJOE) 119
Duong et al.

an RBF neural network in the design is aimed at effectively approximating unknown

elements in the control law. This allows the proposed system to adapt to different
rehabilitation subjects with the same performance without any manual adjustment.
Additionally, the stability of the system and the adaptive law have been rigorously
analyzed using Lyapunov’s stability theory. The entire system operates seamlessly
and efficiently as a cohesive unit.
In a comprehensive assessment, it is evident that the proposed controller gener-
ates adaptable control signals capable of accommodating a range of system condi-
tions while consistently achieving outstanding tracking performance. For example,
during the tracking of sinusoidal signals with a load-free scenario and an amplitude
of 40.0°, the tracking error consistently stays below 4.0° (equivalent to 10% of the
amplitude), irrespective of the signal frequency.
Adding loads to the system is a more realistic circumstance, which indicates that
the controller’s performance is still guaranteed to adhere to LOKOMAT’s accuracy,
as specified in the reference [45]. This promises to be able to build a complete reha-
bilitation robot for both the lower and upper limbs of patients. Taking a deeper
look inside, the foundation ideas of the control algorithm led to some specific upper
hands. The process by which RBF updates the parameters for control signals can
curtail the effects of unstable elements. Adaptability and stability are proven in tran-
sient-impact situations as well. Whereas dynamic hysteresis and slow response are
inherent weaknesses that need further study. In the future developments outlined in
this paper, our intentions include integrating more advanced RBF algorithms with
other sophisticated control strategies, leveraging the versatile capabilities of neural
networks. Furthermore, we intend to conduct a series of practical experiments to
serve rehabilitation purposes.

6 REFERENCES

[1] F. Escobar, S. Díaz, C. Gutiérrez, Y. Ledeneva, C. Hernández, D. Rodríguez, and R. Lemus,

“Simulation of control of a scara robot actuated by pneumatic artificial muscles using
rnapm,” Journal of Applied Research and Technology, vol. 12, no. 5, 2014. https://doi.
org/10.1016/S1665-6423(14)70600-5
[2] A. Rezoug, B. Tondu, and M. Hamerlain, “Experimental study of nonsingular termi-
nal sliding mode controller for robot arm actuated by pneumatic artificial muscles,”
IFAC Proceedings Volumes, vol. 47, no. 3, pp. 10113–10118, 2014. https://doi.org/10.3182/
20140824-6-ZA-1003.00730
[3] S. K. Banala, S. H. Kim, S. K. Agrawal, and J. P. Scholz, “Robot assisted gait training with
active leg exoskeleton (alex),” in 2008 2nd IEEE RAS & EMBS International Conference on
Biomedical Robotics and Biomechatronics, Scottsdale, AZ, USA, 2008, pp. 653–658. https://
doi.org/10.1109/BIOROB.2008.4762885
[4] K. Kadota, M. Akai, K. Kawashima, and T. Kagawa, “Development of a power-assist robot
arm using pneumatic rubber muscles with a balloon sensor,” in RO-MAN 2009- The
18th IEEE International Symposium on Robot and Human Interactive Communication,
2009, pp. 546–551. https://doi.org/10.1109/ROMAN.2009.5326335
[5] A. D. D. R. Carvalho, N. Karanth P, and V. Desai, “Characterization of pneumatic muscle
actuators and their implementation on an elbow exoskeleton with a novel hinge
design,” Sensors and Actuators Reports, vol. 4, p. 100109, 2022. https://doi.org/10.1016/
j.snr.2022.100109
[6] K. Deckers, P. Guillaume, D. Lefeber, G. De Roeck, and E. Reynders, “Modal testing of
bridges using low-weight pneumatic artificial muscle actuators,” in Proceedings of
IMAC 26, the International Modal Analysis Conference, 2008.

120 International Journal of Online and Biomedical Engineering (iJOE) iJOE | Vol. 20 No. 12 (2024)
 Adaptive Control Using Radial Basis Function Neural Networks for Pneumatic Artificial Muscle Systems

[7] J. Radojicic, D. Surdilovic, and G. Schreck, “Modular hybrid robots for safe human-robot
interaction,” Zenodo, 2009. https://doi.org/10.5281/zenodo.1077433
[8] K. Yokoyama and K. Kogiso, “Pid position control of McKibben pneumatic artificial mus-
cle using only pressure feedback,” in 2018 Annual American Control Conference (ACC),
2018, pp. 3362–3367. https://doi.org/10.23919/ACC.2018.8431631
[9] G. Andrikopoulos, G. Nikolakopoulos, and S. Manesis, “Advanced nonlinear pid-based
antagonistic control for pneumatic muscle actuators,” IEEE Transactions on Industrial
Electronics, vol. 61, no. 12, pp. 6926–6937, 2014. https://doi.org/10.1109/TIE.2014.2316255
[10] J. Lilly and P. Quesada, “A two-input sliding-mode controller for a planar arm actu-
ated by four pneumatic muscle groups,” IEEE Transactions on Neural Systems and
Rehabilitation Engineering, vol. 12, no. 3, pp. 349–359, 2004. https://doi.org/10.1109/TNSRE.
2004.831490
[11] K. Xing, J. Huang, Y. Wang, J. Wu, and Q. X. J. He, “Tracking control of pneumatic artifi-
cial muscle actuators based on sliding mode and nonlinear disturbance observer,” IET
Control Theory & Applications, vol. 4, no. 10, pp. 2058–2070, 2010. https://doi.org/10.1049/
iet-cta.2009.0555
[12] C.-J. Chiang and Y.-C. Chen, “Neural network fuzzy sliding mode control of pneumatic
muscle actuators,” Engineering Applications of Artificial Intelligence, vol. 65, pp. 68–86,
2017. https://doi.org/10.1016/j.engappai.2017.06.021
[13] C. P. Vo, X. D. To, and K. K. Ahn, “A novel adaptive gain integral terminal sliding mode con-
trol scheme of a pneumatic artificial muscle system with time-delay estimation,” IEEE
Access, vol. 7, pp. 141133–141143, 2019. https://doi.org/10.1109/ACCESS.2019.2944197
[14] J. Mohorcic and L. Dong, “Extended state observer-based pressure control for pneumatic
actuator servo systems,” Control Theory and Technology, vol. 19, pp. 64–79, 2021. https://
doi.org/10.1007/s11768-021-00038-y
[15] Q.-T. Dao, M.-L. Nguyen, and S.-I. Yamamoto, “Discrete-time fractional order integral slid-
ing mode control of an antagonistic actuator driven by pneumatic artificial muscles,”
Applied Sciences, vol. 9, no. 12, p. 2503, 2019. https://doi.org/10.3390/app9122503
[16] Q.-T. Dao, D.-H. Mai, D. K. Nguyen, and N. T. Ly, “Adaptive parameter integral sliding
mode control of pneumatic artificial muscles in antagonistic configuration,” Journal
of Control, Automation, and Electrical Systems, vol. 33, p. 1116–1124, 2022. https://doi.
org/10.1007/s40313-022-00902-5
[17] R. Fellag, M. Hamerlain, S. Laghrouche, and N. Achour, “Adaptive discrete sliding mode
control of a pneumatic artificial muscle robot,” in 2017 5th International Conference
on Electrical Engineering-Boumerdes (ICEE-B), 2017, pp. 1–6. https://doi.org/10.1109/
ICEE-B.2017.8192099
[18] Q.-T. Dao, T.-K. Le Tri, V.-A. Nguyen, and M.-L. Nguyen, “Discrete time sliding mode
control with power rate exponential reaching law of a pneumatic artificial muscle sys-
tem,” Control Theory and Technology, vol. 20, pp. 514–524, 2022. https://doi.org/10.1007/
s11768-022-00117-8
[19] L. Zhao, X. Liu, and T. Wang, “Trajectory tracking control for double joint manipulator
systems driven by pneumatic artificial muscles based on a nonlinear extended state
observer,” Mechanical Systems and Signal Processing, vol. 122, pp. 307–320, 2019. https://
doi.org/10.1016/j.ymssp.2018.12.016
[20] S. Wang, W. Tan, and D. Li, “Design of linear adrc for load frequency control of power sys-
tems with wind turbines,” in 2016 14th International Conference on Control, Automation,
Robotics, and Vision (ICARCV), 2016, pp. 1–5. https://doi.org/10.1109/ICARCV.2016.
7838795
[21] S. Aole, I. Elamvazuthi, L. Waghmare, B. Patre, and F. Meriaudeau, “Improved active
disturbance rejection control for trajectory tracking control of lower limb robotic
rehabilitation exoskeleton,” Applied Sciences, vol. 20, no. 13, p. 1287, 2020. https://doi.
org/10.3390/app12031287

iJOE | Vol. 20 No. 12 (2024) International Journal of Online and Biomedical Engineering (iJOE) 121
Duong et al.

[22] M. A. Alsharaiah, “Neural network prediction model to explore complex nonlinear

behavior in dynamic biological network,” Int. J. Interact. Mob. Technol. (iJIM), vol. 16,
no. 12, pp. 32–51, 2022. https://doi.org/10.3991/ijim.v16i12.30467
[23] T. Ganokratanaa and M. Ketcham, “An intelligent autonomous document mobile delivery
robot using deep learning,” Int. J. Interact. Mob. Technol. (iJIM), vol. 16, no. 21, pp. 4–22,
2022. https://doi.org/10.3991/ijim.v16i21.32071
[24] H. A. Abu-Alsaad, “CNN-based smart parking system,” Int. J. Interact. Mob. Technol. (iJIM),
vol. 17, no. 11, pp. 155–170, 2023. https://doi.org/10.3991/ijim.v17i11.37033
[25] O. M. Al-hazaimeh, A. Abu-Ein, N. Tahat, M. Al-Smadi, and M. Al-Nawashi, “Combining
artificial intelligence and image processing for diagnosing diabetic retinopathy in reti-
nal fundus images,” Int. J. Onl. Eng. (iJOE), vol. 18, no. 13, pp. 131–151, 2022. https://doi.
org/10.3991/ijoe.v18i13.33985
[26] N. Gharaibeh, A. Abu-Ein, O. M. Al-hazaimeh, K. M. Nahar, W. A. Abu-Ain, and
M. Al-Nawashi, “Swin transformer-based segmentation and multi-scale feature pyramid
fusion module for Alzheimer’s disease with machine learning,” Int. J. Onl. Eng. (iJOE),
vol. 19, no. 4, pp. 22–50, 2023. https://doi.org/10.3991/ijoe.v19i04.37677
[27] M. Kh. Khazaaleh et al., “Handling DNA malfunctions by an unsupervised machine
learning model,” Journal of Pathology Informatics, vol. 14, p. 100340, 2023. https://doi.
org/10.1016/j.jpi.2023.100340
[28] O. M. Al-Hazaimeh, M. Al-Nawashi, and M. Saraee, “Geometrical-based approach for
robust human image detection,” Multimed. Tools Appl., vol. 78, pp. 7029–7053, 2019.
https://doi.org/10.1007/s11042-018-6401-y
[29] P. Gupta and N. K. Sinha, “Intelligent control of robotic manipulators: Experimental
study using neural networks,” Mechatronics, vol. 10, nos. 1–2, pp. 289–305, 2000. https://
doi.org/10.1016/S0957-4158(99)00059-8
[30] G. Naganna and S. Kumar, “Conventional and intelligent controllers for robotic
manipulators,” in 2006 IEEE International Conference on Industrial Technology, 2006,
pp. 424–428. https://doi.org/10.1109/ICIT.2006.372240
[31] Zaitceva, Iuliia, and B. Andrievsky, “Methods of intelligent control in mechatronics and
robotic engineering: A survey,” Electronics, vol. 11, no. 15, p. 22443, 2022. https://doi.
org/10.3390/electronics11152443
[32] S. Chan, J. Lilly, D. Repperger, and J. Berlin, “Fuzzy PD+i learning control for a pneu-
matic muscle,” in the 12th IEEE International Conference on Fuzzy Systems, 2003, FUZZ ’03,
2003, vol. 1, pp. 278–283. https://doi.org/10.1109/FUZZ.2003.1209375
[33] H. P. H. Anh and K. K. Ahn, “Hybrid control of a pneumatic artificial muscle (pam) robot
arm using an inverse narx fuzzy model,” Engineering Applications of Artificial Intelligence,
vol. 24, no. 4, pp. 697–716, 2011. https://doi.org/10.1016/j.engappai.2010.11.007
[34] C. V. Kien, T. T. Huan, D. T. Thai, and H. P. H. Anh, “Implementation of adaptive fuzzy
sliding mode control for nonlinear uncertain serial pneumatic-artificial-muscle (PAM)
robot system,” in 2017 International Conference on System Science and Engineering (ICSSE),
2017, pp. 83–88. https://doi.org/10.1109/ICSSE.2017.8030842
[35] D. Liang, N. Sun, Y. Wu, G. Liu, and Y. Fang, “Fuzzy-sliding mode control for humanoid
arm robots actuated by pneumatic artificial muscles with unidirectional inputs, sat-
urations, and dead zones,” IEEE Transactions on Industrial Informatics, vol. 18, no. 5,
pp. 3011–3021, 2022. https://doi.org/10.1109/TII.2021.3111655
[36] A. Rezoug, S. Boudoua, and F. Hamerlain, “Application of fuzzy sliding mode to control of
manipulator robots actuated by pneumatic artificial muscles,” IFAC Proceedings Volumes,
vol. 42, no. 19, pp. 580–585, 2009. https://doi.org/10.3182/20090921-3-TR-3005.00099
[37] C. Cao, L. Ma, and Y. Xu, “Adaptive control theory and applications,” Journal of Control
Science and Engineering, vol. 2012, pp. 1–2, 2012. https://doi.org/10.1155/2012/827353

122 International Journal of Online and Biomedical Engineering (iJOE) iJOE | Vol. 20 No. 12 (2024)
 Adaptive Control Using Radial Basis Function Neural Networks for Pneumatic Artificial Muscle Systems

[38] J. Liu, Radial Basis Function (RBF) Neural Network Control for Mechanical Systems.
1st Edition. Berlin: Springer Berlin, Heidelberg, 2013. https://doi.org/10.1007/978-3-
642-34816-7
[39] W. He, Y. Chen, and Z. Yin, “Adaptive neural network control of an uncertain robot with
full-state constraints,” IEEE Transactions on Cybernetics, vol. 46, no. 3, pp. 620–629, 2016.
https://doi.org/10.1109/TCYB.2015.2411285
[40] W. He, Z. Yan, C. Sun, and Y. Chen, “Adaptive neural network control of a flapping wing
microaerial vehicle with disturbance observer,” IEEE Transactions on Cybernetics,
vol. 47, no. 10, pp. 3452–3465, 2017. https://doi.org/10.1109/TCYB.2017.2720801
[41] H. Yang and J. Liu, “An adaptive RBF neural network control method for a class of non-
linear systems,” IEEE/CAA Journal of Automatica Sinica, vol. 5, no. 2, pp. 457–462, 2018.
https://doi.org/10.1109/JAS.2017.7510820
[42] K. J. Astrom and B. Wittenmark, Adaptive Control, 2nd Edition, Addison Wesley,
1995.
[43] Q. T. Dao, D.-H. Mai, and D.-K. Nguyen, “Experimental validation for pneumatic artificial
muscles using fuzzy PID control,” Measurement, Control, and Automation, vol. 2, no. 1,
pp. 58–64, 2022.
[44] D. A. Winter, Biomechanics and Motor Control of Human Movement, 4th Edition. John
Wiley & Sons, Ltd, 2009. https://doi.org/10.1002/9780470549148
[45] R. Riener, L. Lunenburger, S. Jezernik, M. Anderschitz, G. Colombo, and V. Dietz, “Patient-
cooperative strategies for robot-aided treadmill training: First experimental results,”
IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 13, no. 3,
pp. 380–394, 2005. https://doi.org/10.1109/TNSRE.2005.848628

7 AUTHORS

Minh-Duc Duong received his Ph.D. degree in Electronic and Information

Engineering from Toyohashi University of Technology in 2008. Presently, he is
an associate professor at the Department of Automation Engineering, School of
Electrical and Electronic Engineering, Hanoi University of Science and Technology.
His research interests include rehabilitation robots, vibration suppression control,
and bilateral teleoperation robot systems.
Viet-Thanh Nguyen is a senior in the School of Electrical and Electronic
Engineering at the Hanoi University of Science and Technology. As a member of
the esteemed Talented Program in Automation and Control Engineering, Viet-Thanh
started his research journey during his sophomore year at the Motion Control and
Applied Robotics Laboratory (MoCAR). His main research orientation is advanced
control algorithms for the pneumatic artificial muscle system (PAM) using (RBFNNs).
Quy-Thinh Dao received both his B.S. and M.S. degrees in Electrical Engineering
from Hanoi University of Science and Technology, Vietnam, in 2007 and 2010,
respectively. Subsequently, he completed his Ph.D. in functional control systems at
Shibaura Institute of Technology, Tokyo, Japan, in 2019. Currently serving as a senior
lecturer in the Department of Automation Engineering within the School of Electrical
and Electronic Engineering at Hanoi University of Science and Technology, he spe-
cializes in captivating research areas such as smart actuators, control theory, and the
advancement of welfare and service robotics (E-mail: [email protected]).

iJOE | Vol. 20 No. 12 (2024) International Journal of Online and Biomedical Engineering (iJOE) 123