robotics

Article
Finger Joint Stiffness Estimation with Joint Modular Soft
Actuators for Hand Telerehabilitation
Fuko Matsunaga 1 , Shota Kokubu 1 , Pablo Enrique Tortos Vinocour 1 , Ming-Ta Ke 2 , Ya-Hsin Hsueh 3 ,
Shao Ying Huang 4 , Jose Gomez-Tames 1,5 and Wenwei Yu 1,5, *

1 Department of Medical Engineering, Graduate School of Engineering, Chiba University, Chiba 263-8522, Japan
2 Graduate School of Intelligent Data Science, National Yunlin University of Science and Technology,
Yunlin 64002, Taiwan
3 Department of Electronic Engineering, National Yunlin University of Science and Technology,
Yunlin 64002, Taiwan
4 Engineering Product Development Department, Singapore University of Technology and Design,
Singapore 487372, Singapore
5 Center for Frontier Medical Engineering, Chiba University, Chiba 263-8522, Japan
* Correspondence: yuwill@faculty.chiba-u.jp; Tel.: +81-43-290-3231

Abstract: In a telerehabilitation environment, it is difficult for a therapist to understand the condition

of a patient’s finger joints because of the lack of direct assessment. In particular, not enabling the
provision of spasticity evaluation significantly reduces the optimal performance of telerehabilitation.
In a previous study, it has been proposed that finger stiffness could be estimated using an analytical
model of a whole-finger soft actuator. However, because the whole-finger soft actuators require
high air pressure for high bending performance and are costly to customize for each patient, using
joint modular soft actuators for telerehabilitation turns to be a necessity, though stiffness estimation
with joint modular soft actuators has not been studied yet. Another problem is caused by using
a marker-based joint angle measurement, which requires the markers to be attached to the exact
Citation: Matsunaga, F.; Kokubu, S.; positions, and limits its application in telerehabilitation. In this study, we proposed a procedure of
Tortos Vinocour, P.E.; Ke, M.-T.; finger joint stiffness estimation that combines information acquired from a joint modular soft actuator
Hsueh, Y.-H.; Huang, S.Y.; and a marker-less hand joint position acquisition device. Correction parameters were added to the
Gomez-Tames, J.; Yu, W. Finger Joint previous analytical model for -the bending analysis of a joint assisted using a joint modular soft
Stiffness Estimation with Joint actuator. Moreover, a multi-variate regression model was implemented for correcting joint angles
Modular Soft Actuators for Hand
obtained from the hand joint position acquisition device. As a result, a reasonable accuracy of stiffness
Telerehabilitation. Robotics 2023, 12,
estimation was achieved for rehabilitation with the joint modular soft actuators, which suggests
83. https://doi.org/10.3390/
the possibility of using the proposed method to evaluate the finger spasticity in a telerehabilitation
robotics12030083
environment. This is a big step forward towards optimal hand telerehabilitation.
Academic Editors: Kensuke Harada,
Shuxiang Guo, Yunchao Tang, Keywords: soft actuator; leap motion controller; telerehabilitation; finger joint stiffness
Mingjie Dong and Wei Feng

Received: 16 May 2023

Revised: 4 June 2023
Accepted: 6 June 2023 1. Introduction
Published: 7 June 2023 Stroke is one of the leading causes of death and disability in adults worldwide [1].
At least 65% of stroke patients have hand impairment [2]. It has been reported that
hand dysfunction can be a major obstacle in daily life and labor [3]. Therefore, it is
important for patients to undergo hand rehabilitation in order to recover from stroke
Copyright: © 2023 by the authors.
side effects [4,5]. Constant high-quality rehabilitation is necessary for effective hand
Licensee MDPI, Basel, Switzerland.
rehabilitation during motor paralysis [6]. Generally, rehabilitation is performed under the
This article is an open access article
guidance and assistance of a therapist; thus, a facility where therapists work with the other
distributed under the terms and
conditions of the Creative Commons
specialists is usually the place for rehabilitation. However, the recent spread of infectious
Attribution (CC BY) license (https://
diseases and the shortage of visiting therapists have reduced the frequency of rehabilitation
creativecommons.org/licenses/by/
that can be provided to patients [7]. Furthermore, most therapists are often located in
4.0/).

Robotics 2023, 12, 83. https://doi.org/10.3390/robotics12030083 https://www.mdpi.com/journal/robotics

Robotics 2023, 12, 83 2 of 25

urban centers, so patients living in underpopulated areas may be unable to participate in

rehabilitation [3].
Telerehabilitation is gaining attention as a new means to address such problems, im-
proving access for patients living in remote areas and reducing the financial and time costs
associated with outpatient rehabilitation [8]. However, in a telerehabilitation environment,
the therapist and patient are normally far away from each other, making it difficult for the
therapist to palpate and understand the condition of the patient’s hand [9].
Finger flexor spasticity, a motor disorder that causes involuntary muscle contractions
due to impaired reflex function, is a common aftereffect of stroke [10]. Spasticity not only
increases the stiffness of the fingers but also decreases their range of motion, affecting
the recovery of the patient’s motor function [11]. Therefore, understanding a patient’s
spasticity condition can help therapists determine optimal rehabilitation [12]. However,
spasticity is difficult to evaluate without direct contact with the therapist, as is the degree
of muscle tone that fluctuates from day to day [13]. Therefore, a method to objectively
quantify spasticity by non-palpation is necessary in telerehabilitation. Up to now, dedicated
stiffness measurement devices that have been developed are large and heavy because they
consist of many rigid body parts [14–16]. When rehabilitation is performed at a patient’s
home, the available space is assumed to be limited, and it is desirable to minimize the
number of devices by accomplishing simultaneous functions, such as the evaluation of
spasticity condition and rehabilitation using the same device.
Heung et al. developed a system that combines the two functions of finger motor
support and spasticity condition evaluation by using a soft elastic composite actuator
(SECA), a hand rehabilitation device, to estimate finger joint stiffness [17]. Soft actuators
are suitable devices for hand rehabilitation in a telerehabilitation environment because
of their flexibility, lightweight, high affinity with the body, and portability. In [18], the
stiffness estimation method using a SECA was found to agree with standard joint stiffness
quantification methods. Furthermore, it was suggested that a SECA could be used to
quantitatively assess the stiffness of the passive metacarpophalangeal (MCP) joint during
the performance of various tasks in hand rehabilitation. The stiffness estimation method
using a SECA estimates stiffness from an analytical model using the input air pressure of
the soft actuator and a finger joint angle, which requires a structure-dependent soft actuator
analytical model and a joint angle measurement. In addition, there are several technical
problems with its application in telerehabilitation.
Firstly, a whole-finger soft actuator with a single-pocket structure with multiple air
chambers connected to each other, such as the SECA, is not suitable for home environ-
ments because it requires high air pressure to achieve high bending performance [17,19,20].
Secondly, this type of actuator has only one pneumatic input and cannot support individ-
ual joints. Thirdly, the dimensions of the actuator need to be changed to accommodate
individual differences such as finger size, but this change may cause unintended changes
in actuator motion [21]. Finally, it is costly to customize the actuator for each patient.
Yun et al. and Kokubu et al. developed joint modular soft actuators that are divided into
parts corresponding to each finger joint [20,21]. By using spacers to connect each actuator,
the actuators can accommodate differences in the length of the user’s fingers without
changing the dimensions of the actuator. This enables personalized customization at a low
cost. In addition, the actuator does not require high air pressure because it transmits force
better than a whole-finger soft actuator does [20]. Based on these considerations, the joint
modular soft actuator can be considered a suitable soft actuator for use in a telerehabili-
tation environment because it provides high bending performance with minimal energy
and can accommodate individual differences with minimal effort. However, it is necessary
to incorporate an estimation of the stiffness of the finger joints (i.e., spasticity condition
evaluation function), though stiffness estimation with joint modular soft actuators has not
been studied yet. In this study, we designed a modular version of the SECA for actuating
individual joints (Modular-SECA). However, as explained in the following paragraph, the
Robotics 2023, 12, 83 3 of 25

analytical model of SECA for estimating finger joint stiffness is not completely applicable
to this new type of actuator.
The analytical model of the SECA is derived from the air chamber unit energy con-
servation law, so the length of the air chamber is included in the analytical model, but
not the length of the actuator itself. When the lengths of the air chambers are equal, the
analytical model can be interpreted as indicating that the SECA and Modular-SECA have
the same bending performance. Therefore, it is possible to estimate the stiffness of the
Modular-SECA using the original analytical model of the SECA. However, unlike the SECA,
the joint modular soft actuator has rigid connectors at both ends, and the length ratio of the
actuator body to the chamber is different [17,20]. Therefore, the two types of soft actuators
may exhibit different bending performances, and the analytical model of the SECA cannot
be applied directly to that of the Modular-SECA.
Another problem of the stiffness estimation method described in [17] is caused by
using a marker-based joint angle measurement. However, it is difficult for the patient
or his/her family members to attach the markers at precise positions, which limits its
application to telerehabilitation.
Furthermore, using 2D markers limits the acquisition of joint angles on a flat surface,
making it impossible to measure the joint angles of the middle and ring fingers, and thus
making it difficult to estimate the joint stiffness of them. Zhou et al. have developed a
system that simultaneously obtains the joint angles of multiple fingers and estimates their
stiffness by integrating a flex sensor into an actuator [22]. Flex sensors do not interfere with
the bending response of soft actuators too much, but they must be calibrated individually
for each actuator and cannot measure finger or hand orientation [23,24]. Moreover, using
this system, it is not possible to measure joint angles when a soft actuator is not attached. By
designing a system that can measure joint angles without having the soft actuators attached,
it is also possible to measure joint range of motion, increasing the system’s functionality.
In addition, having a high-cost system can limit how many devices can be provided for
telerehabilitation [25].
The leap motion controller (LMC) is an inexpensive and compact marker-less mea-
surement device for tracking and acquiring the position of hand joints. Although the
accuracy of hand joint angle measurement is not sufficient for rehabilitation purposes [26],
it is possible to improve the measurement accuracy by fitting a regression model of joint
angle [27]. However, the combined use of LMC and soft actuators has not yet been inves-
tigated. Especially, joint modular soft actuators may cause the motion of one single joint,
which is different from most natural hand motions. It is also unclear how the attachment
of the soft actuator affects the LMC and whether or not the measurement accuracy can be
improved in the same way as when the soft actuator is not attached.
In this study, we proposed finger joint stiffness estimation for hand–finger joints using
Modular-SECAs and the LMC, and verified the feasibility of this procedure by performing
stiffness estimation on a mannequin hand and a subject’s MCP joint, respectively. We
designed the Modular-SECA, a joint modular soft actuator incorporating the finger joint
stiffness estimation function. In addition, correction terms (or parameters) were added to
the previous analytical model for the bending analysis of a Modular-SECA supporting a
joint for the purpose of hand rehabilitation. The correction parameters were determined
by comparing the values of the bending angles obtained from the analytical model with
measured angles from a free bending experiment with the Modular-SECAs. Moreover, a
regression model accounting for the actuation for and motion of the neighboring joints was
proposed to correct the direct outputs of LMC.
This paper is organized as follows: First, the design and improved analytical model of
the joint modular soft actuator for stiffness estimation are presented. Next, the regression
model method to improve the accuracy of LMC joint angle measurement is presented.
Finally, validation experiments using a mannequin hand and a subject’s hand are presented
in which stiffness estimation is performed using a combination of these methods.
 2. Materials and Methods
2.1. Joint Modular Soft Actuators for Joint Stiffness Estimation
Robotics 2023, 12, 83 4 of 25
This subsection gives a concise introduction to the design of the joint modular soft
actuator, for the sake of clarity. A more comprehensive explanation can be found in [20].
The analytical model for the SECA and thereby the improved analytical model for the
2. Materials and Methods
joint modular soft
2.1. Joint actuator
Modular are explained.
Soft Actuators for Joint Stiffness Estimation
This subsection gives a concise introduction to the design of the joint modular soft
2.1.1. Design andfor
actuator, Fabrication
the sake of clarity. A more comprehensive explanation can be found in [20].
TheThe analytical modelconsists
Modular-SECA for the SECA
of anand thereby thebody,
elastomer improved analytical model
a semi-circular forchamber
air the joint with
modular soft actuator are explained.
air pockets, a silicone tube, a fiber reinforcement wrapped around its circumference, 3D-
printed rigid connectors,
2.1.1. Design and a torque compensation layer. When pressurized with air
and Fabrication
pressure, theThe fibers wrapped consists
Modular-SECA aroundofthe surface of
an elastomer theaactuator
body, suppress
semi-circular radial
air chamber elonga-
with
air pockets, a silicone tube, a fiber reinforcement wrapped around
tion. The torque-compensating layer limits axial elongation at the bottom, allowing onlyits circumference, 3D-
printed rigid connectors, and a torque compensation layer. When pressurized with air
the top of the actuator to extend and the finger to bend. During decompression, the
pressure, the fibers wrapped around the surface of the actuator suppress radial elongation.
torque-compensating layer provides
The torque-compensating a supplementary
layer limits axial elongationbending moment
at the bottom, to extend
allowing only thethe fin-
ger to itstop
initial
of theposition.
actuator to extend and the finger to bend. During decompression, the torque-
The compensating
silicon partlayer of the actuator
provides was madebending
a supplementary by pouring
moment silicon (Smooth-On,
to extend Dragon
the finger to its
initial position.
Skin 10 Medium) into a 3D-printed mold. A stainless-steel plate (A2 stainless-steel plate)
The silicon part of the actuator was made by pouring silicon (Smooth-On, Dragon Skin
with a thickness of 0.2 mm was used for the torque compensating layer, which was at-
10 Medium) into a 3D-printed mold. A stainless-steel plate (A2 stainless-steel plate) with a
tached tothickness
the bottom of the
of 0.2 mm wassilicon
used for part of the
the torque actuator. Next,
compensating 0.3556
layer, which wasmm diameter
attached to the rein-
forcing fibers
bottommade of Kevlar
of the silicon (Dupont,
® the
part of actuator. Inc.,
Next, Delaware, USA) reinforcing
0.3556 mm diameter were wound fibersaround
made the
®
actuator,ofandKevlar
silicon(Dupont,
tubingInc.,
for Wilmington,
air flow andDE, USA) were connectors
3D-printed wound around madethe actuator, and acid
of polylactic
silicon tubing for air flow and 3D-printed connectors made of polylactic acid (PLA) resin
(PLA) resin were attached. Finally, a thin coat of silicone (Smooth-On, Dragon Skin 10
were attached. Finally, a thin coat of silicone (Smooth-On, Dragon Skin 10 Medium) was
Medium)applied
was applied
over the over
fibers the fibers them
to prevent to prevent them from
from shifting. Figureshifting.
1a shows theFigure 1a shows the
cross-section
cross-section and dimensions
and dimensions of the actuator.
of the actuator.

(a) (b)
Figure 1. (a) Modular-SECA
Figure cross-section
1. (a) Modular-SECA and dimensions;
cross-section (b) (b)
and dimensions; bending angle
bending anglemeasurement of Mod-
measurement of
ular-SECAModular-SECA
in free space.
in free space.

2.1.2. Analytical Model for Energy Conservation of Modular-SECA

2.1.2. Analytical Model for Energy Conservation of Modular-SECA
The analytical model for the free space bending analysis of the SECA is obtained via
The the
analytical model
conservation for the
of energy free
as in space bending analysis of the SECA is obtained via
[17]:
the conservation of energy as in [17]: W +W A L
P= (1)
𝑊 ∆V
+𝑊
𝑃= (1)
∆𝑉
where P, WA , WL and ∆V denote the input air pressure, the bending strain energy stored in
the soft actuator body, the bending strain energy stored in the torque-compensating layer,
where 𝑃,and
𝑊 the
, 𝑊increase
and ∆𝑉 denote the input air pressure, the bending strain energy stored
in volume, respectively. The bending analysis in free space is the bending
in the soft actuator
analysis usingbody, the bending
the actuator only. strain energy stored in the torque-compensating
layer, and theDueincrease in volume,
to the structural respectively.
differences between theThe bending
SECA analysis in free
and the Modular-SECA, space is the
the energy
bending distribution
analysis using in thethe
actuator is different.
actuator only. The rigid connectors attached to the two ends
of each actuator and the neighboring Modular-SECA connected through the connectors
Due to the structural differences between the SECA and the Modular-SECA, the en-
ergy distribution in the actuator is different. The rigid connectors attached to the two ends
of each actuator and the neighboring Modular-SECA connected through the connectors
may cause the change in the ratio of both 𝑊 to 𝑊 , and their sum to 𝑃∆𝑉. Therefore,
Robotics 2023, 12, 83 5 of 25

may cause the change in the ratio of both WA to WL , and their sum to P∆V. Therefore,
the analytical model of the SECA was modified to match the energy distribution of the
Modular-SECA, as shown in Equation (2):

α A WA + α L WL
P= (2)
αV ∆V

where α A , α L and αV are the correction parameters added to Equation (1).

The analytical model for bending analysis in free space was verified with a Modular-
SECA prototype experiment, and a finite element model (FEM). In this study, the Yeoh
3rd-order model was chosen for the strain energy density function of silicone rubber as
follows [28]:
3  i
wm = ∑ Ci λ2 + λ−2 − 2 (3)
i =1

wm is the bending strain energy density function, λ is the principal stretch, and
Ci (i = 1, 2, 3) is the material coefficient. The 3rd-order Yeoh model parameters for the
silicon used (Smooth-On, Dragon Skin 10 Medium) were C1 = 0.04773 MPa,
C2 = 3.42 × 10−4 MPa, and C3 = −1.1942 × 10−6 MPa, respectively. These parame-
ters were calculated by using the material data provided in [29]. The material parameters
of the stainless-steel plates used as torque compensation layers with a Young’s modulus
of 193 GPa, a Poisson’s ratio of 0.318, and a density of 8000 kg/m3 [30]. The bending
angle in free space was measured by a camera-based two-dimensional marker detection
system (Figure 1b). The camera (C930eR, Logicool, Lausanne, Switzerland) recorded the
2D position coordinates of the markers and the angle was calculated. The bending angle,
θ, was obtained using Equation (4) when air pressure was applied from 0 kPa to 80 kPa
in 10 kPa increments. The bending angles obtained from the experiment were compared
with those from the analytical model, and correction parameters were identified using
the least-squares method. A more comprehensive description of FEM can be found in
Appendix A.  
→ →
 a·b 
−1 
θ = cos 
 → →  (4)
a b

2.2. Finger Joint Angle Measurement Using Leap Motion Controller

This subsection describes the performance of LMC and a regression model to correct
joint angles determined using LMC.

2.2.1. Leap Motion Controller

Leap Motion Controller (Ultraleap, Bristol, UK) is a hand motion capture system that
can recognize positions of hand joints within a maximum range of 60–80 cm from the top
of the device with a 140 × 120-degree field of view [31]. LMC is a marker-less motion
capture device that uses reflections from three infrared LEDs on either side of and between
the two infrared cameras to estimate the position of the fingers. The device weighs 32 g
and measures 30 mm × 11.30 mm × 80 mm [32]. LMC uses a proprietary algorithm to
extract hand data from raw sensor data. These data consist of infrared brightness values
and calibration data to remove distortions [25].

2.2.2. Joint Angle Measurement

In this study, a mannequin hand was used instead of a human hand. The mannequin
hand is easy to modify, and it can keep gestures without involuntary motions, making it
suitable for experiments in which the same measurement is repeated. Joint angles were
measured using two methods: LMC and a camera-based 2D marker detection system.
The values from the 2D marker detection system were similar to those measured using a
 𝑊𝑀⃗ ∙ 𝑀𝑃⃗
𝜃 = cos (5)
𝑊𝑀⃗ ∙ 𝑀𝑃⃗

𝑀𝑃⃗ ∙ 𝑃𝐷⃗
Robotics 2023, 12, 83 6 of 25
𝜃 = cos (6)
𝑀𝑃⃗ ∙ 𝑃𝐷⃗
goniometer (Appendix B), so these values were used as the reference values of joint angles
in this study. Figure 2a shows the setup for the joint𝑃𝐷 ⃗ ∙ 𝐷𝑇measurement.
angle ⃗ For the ease of
𝜃 = cos (7)
2D marker attachment, the index finger of the right𝑃𝐷 ⃗ ∙ was
hand 𝐷𝑇 ⃗ used for the measurement.
The 2D markers were attached to the side of the index finger of the mannequin hand, and
theThe anglebetween
distance obtained
LMC viaand
thethe
method usinghand
mannequin the two-dimensional
was set at 15 cm [33].marker is denoted as
The frictional
2D Angle, and the angle obtained by the method using LMC is denoted as the LMC Angle.
resistance of each joint of the index finger of the mannequin hand was negligible.

(a) (b)
Figure
Figure2.2.(a)
(a)Joint
Joint angle measurement
angle measurement setup;
setup; (b)(b) joints
joints andand
jointjoint angles.
angles.

The 2D marker detection system provides the two-dimensional position coordinates

2.2.3. Joint Angle Correction Model
of the marker’s center, while LMC provides the three-dimensional position coordinates of
the Ganguly et al.
finger joints. conducted
Thus, a comparative
the position experiment
coordinates of points W tousing LMC
T shown in and a gold-standard
Figure 2b were
marker-based motion
obtained, and the anglescapture system,
of the distal to measure(DIP)
interphalangeal jointjoint,
angles during
proximal finger flexion and
interphalangeal
(PIP) joint,
extension and MCP joint
movements were
[26]. Thecalculated using the
results showed following
that LMC couldset ofnot
equations
replace[26]:
the gold-stand-
ard system regarding angle measurement.  Therefore, the  output angles need to be cor-
rected to estimate finger joint stiffness using −−→ −−→ al. derived a regression model for
 LMC. Li et 
−1  W M · MP 
each index finger joint to improveθ MCP =thecos accuracy
 → of joint → angle
 (5)
measurement [27]. Although
W M · MP
there was a certain amount of error after joint angle correction, the overall error was im-
proved to within 10 degrees.  
In this study, a regression model is generated
−−→ −−→ for taking into consideration the influ-
ence of joint modular soft actuators · PD of 
MPjoints
θ PIP = costo
−1multiple fingers. Therefore, unlike(6)
the pre-

 
 −−→ −−→ 
vious study, the regression model in this studyMP needs
· PDthe following explanatory variables.
• Angles of other joints:  
As shown in Appendix C, the LMC −−→ −−→
 Angle of one 
PD · DT
joint is affected by the interaction
between joints of the same cos−1 
θ DIP =finger. Therefore,
 −−→ −−→  to estimate
 (7) it is
the angle of one joint,
PD · DT
necessary to provide its current LMC Angle value, but also the angle values of other
joints in the same finger.
The angle obtained via the method using the two-dimensional marker is denoted as
2D Angle, and the angle obtained by the method using LMC is denoted as the LMC Angle.

2.2.3. Joint Angle Correction Model

Ganguly et al. conducted a comparative experiment using LMC and a gold-standard
marker-based motion capture system, to measure joint angles during finger flexion and
extension movements [26]. The results showed that LMC could not replace the gold-
standard system regarding angle measurement. Therefore, the output angles need to
be corrected to estimate finger joint stiffness using LMC. Li et al. derived a regression
model for each index finger joint to improve the accuracy of joint angle measurement [27].
Although there was a certain amount of error after joint angle correction, the overall error
was improved to within 10 degrees.
Robotics 2023, 12, 83 7 of 25

In this study, a regression model is generated for taking into consideration the influence
of joint modular soft actuators to multiple joints of fingers. Therefore, unlike the previous
study, the regression model in this study needs the following explanatory variables.
• Angles of other joints:
As shown in Appendix C, the LMC Angle of one joint is affected by the interaction
between joints of the same finger. Therefore, to estimate the angle of one joint, it is
necessary to provide its current LMC Angle value, but also the angle values of other
joints in the same finger.
• Dummy variables indicating whether one joint modular soft actuator is pressurized or
not:
When joint modular soft actuators are attached to the joints of one finger, the ten-
dency of angle change of each joint differs depending on whether each soft actuator
is pressurized or not (Appendix C). Therefore, if a regression model is generated
from all the data regarding the angular changes of all joints in (a)–(g) in Figure A2,
corresponding to all combinations of the Boolean pressurization (1 = pressurized,
0 = not pressurized) variables, pDIP, pPIP and pMCP, except the case where pDIP = 0,
pPIP = 0, and pMCP = 0, the goodness of fit is likely low. Therefore, by introducing
the Boolean pressurization variables, a single model can be fitted to reflect the different
angular change tendencies caused by different pressurization combinations.
• Terms reflecting the interaction between variables:
Since it is impossible to guarantee that there is no interaction between all the explana-
tory variables, adding interaction terms may improve the correction performance.
Based on the above points, the regression models were generated to correct the LMC
Angles of the three joints, respectively, using the measurement data in Appendix C. LMC
Angles of the three joints of the same finger (DIP, PIP, and MCP) and the dummy pressur-
ization variables were used as explanatory variables (pDIP, pPIP, and pMCP) and their
interaction terms were also added as explanatory variables (Table 1).

Table 1. Candidate explanatory variables for correction models of LMC Angle with the soft actuator.

Candidate Explanatory Variables Number

1. LMC Angle (DIP, PIP, MCP) 3
2. Dummy variable indicating whether the joint is pressurized
(pDIP, pPIP, pMCP) 3

3. Interaction terms of 1 and 2 15

4. Three joint interaction terms
(DIP· PIP· MCP, pDIP· pPIP· pMCP) 2

Total 23

However, explanatory variables that do not contribute to the correction of joint angles
may be included. The exclusion of variables with low contribution would improve the
correction of joint angles. Therefore, a variable selection was performed to the 23 candidate
explanatory variables shown in Table 1 to obtain the final explanatory variables.

2.2.4. Variable Selection

In order to effectively select the explanatory variables closely associated with the
target variable, the forward–backward stepwise selection, one of the most used stepwise
selections, was used. The forward–backward stepwise selection was created based on [34].
The procedure is described below.
1. Initial variable selection:
Among the given candidate explanatory variables, the variable with the largest single
correlation coefficient with the target variable and whose regression contribution rate
Robotics 2023, 12, 83 8 of 25

(regression variance divided by residual variance) is larger than the predetermined

constant FI N is the initial variable.
2. Adding a variable:
By creating a regression model with the current explanatory variables and adding a
new variable from the candidate variables, the variable is included as an explanatory
variable if it has the largest increase in the contribution rate and the increase in
contribution rate is larger than or equal to FI N .
3. Removing a variable:
A regression model is created with the current explanatory variables, t2 is determined
to test the significance of each partial regression coefficient, and the relevant variable
is removed if its minimum is smaller than the predetermined constant FOUT .
4. End of variable selection:
Steps 2 and 3 are repeated in the same manner, and the calculation is terminated when
there are no more variables to be removed because they are smaller than FOUT or
added because they are larger than FI N .
When applying this method, the addition/removal criteria for variables depend on the
values of FI N and FOUT , so how to determine these values is very important. In Ref. [34],
FI N = FOUT = 2.0 is considered appropriate, and this value is also used in this study.

2.3. Finger Joint Stiffness Estimation

This subsection describes how the performance of the proposed stiffness estimation
method is verified.

2.3.1. Analytical Model for Joint Stiffness Estimation

Since stroke patients experience strong resistance to extension due to excessive tension
in the finger flexors, only the joint stiffness upon extending the fingers was targeted in
this study [17]. Using the correction parameters identified in the previous section and [17],
the analytical model for the finger joint stiffness estimation using the Modular-SECA is
as follows:
2(α A WA + α L WL − αV P∆V )
k= , θ ∈ [0, 0.7θ0 ) (8)
( θ − θ0 )2
k, θ and θ0 are the joint stiffness, the joint angle, and the resting angle, respectively. In this
study, the range of possible joint angles was set to θ ∈ [0, 0.7θ0 ) as in the method using the
SECA because we wanted to estimate the stiffness of the joint when it was extended from
its resting angle.

2.3.2. Joint Stiffness Estimation Experiment

Two MCP stiffness estimation experiments were conducted. First, stiff joints were
made for the mannequin hand using torsion springs, and the stiffness estimated using
the proposed methods was compared with the spring constant to decide the estimation
accuracy. Next, the feasibility of the proposed method for joints of multiple fingers was
verified by estimating the stiffness of a subject’s MCP joints.
In the stiffness estimation experiment, the angles of MCP were measured when
air pressure was applied from 0 kPa to 80 kPa at 10 kPa intervals. After the measure-
ments, the stiffness values for each pressure were obtained using Equation (8) in the range
θ ∈ [0, 0.7θ0 ). Finally, these stiffness values were averaged to obtain the final estimated
stiffness value. The resting angle, θ0 , was measured using the 2D marker detection system
or a goniometer.

Joint Stiffness Estimation Using Mannequin Hand

Experiments using a mannequin hand require a target stiffness value for estimation.
The modified Ashworth scale (MAS), an assessment used in rehabilitation, is shown in
Table 2. Based on the MAS scores of healthy subjects and stroke patients and the measured
joint stiffness values [17,18,35], the relationship of the patient’s joint condition to the MAS
 Table 2. Modified Ashworth scale for grading spasticity [36].
Robotics 2023, 12, 83 9 of 25

no increase in muscle tone

slight increase in muscle score tone,andmanifested
the corresponding byreference
a catch andvalues
stiffness release or byinminimal
were defined resistance
two levels (low and
medium) (Table 3). The torsion spring (SAMINI CO., LTD., Shizuoka, Japan) with a spring
range of motion when the affected
constant part/s
equivalent is/are
to the moved
stiffness in flexion
value of MCP of the indexor finger
extensionof the mannequin
slight increase in muscle tone, manifested by a catch, followed by minimal resistance thro
hand (Figure 3) was used.

der (less than half) of theTable ROM 2. Modified Ashworth scale for grading spasticity [36].

more marked
0 : no increase increase in muscle tone through most of the ROM, but affected part(s) easily
in muscle tone
slight increase in muscle tone, manifested by a catch and release or by minimal resistance at the end of the range of motion when
considerable
1 : increase in muscle tone, passive movement is difficult
the affected part/s is/are moved in flexion or extension
affected
1+ : part/s is/are rigid in flexion or extension
slight increase in muscle tone, manifested by a catch, followed by minimal resistance throughout the remainder (less than half)
of the ROM
2 : more marked increase in muscle tone through most of the ROM, but affected part(s) easily moved
3 : considerable increase in muscle tone, passive movement is difficult
4 : Table
affected part/s is/are rigid in 3. Spasticity
flexion and spring constant of the torsion
or extension spring used in the two le

Reference Stiffness
Table 3. Spasticity and spring constant of the torsion spring used in the two levels of stiffness values.
fness Level MAS Score Joint Condition
(N mm/deg)
Reference Stiffness Spring Constant
Stiffness Level MAS Score Joint Condition
w Stiffness 0 Healthy (N mm/deg) 0.52
(N mm/deg)
Low Stiffness 0 Healthy 0.52
dle Stiffness
Middle Stiffness 1+ 1+ Slight spasticity
Slight spasticity 1.40 1.400.552
1.686

Figure 3. Joint with middle stiffness made using a torsion spring with a spring constant of
Figure 3.mm/deg.
1.686 N Joint with middle stiffness made using a torsion spring with a spr
mm/deg. Stiffness values were estimated using different combinations of analytical models (the
SECA model and Modular-SECA model) and joint angles (2D Angle, original LMC Angle,
and corrected LMC Angle), comprising six combinations overall.
Estimation of Stiffness of Joints of Multiple Fingers
Estimation of Stiffness of Joints of Multiple Fingers
The feasibility of estimating the stiffness of joints of multiple fi
The feasibility of estimating the stiffness of joints of multiple fingers using the proposed
posed
methodmethod wastwo
was verified; verified; twoofdifferent
different states joint stiffnessstates of joint
were setup for this stiffness
purpose: we
healthy
pose: and simulated-stiff.
healthy and simulated-stiff.
The stiffnesses of MCP in multiple fingers (index, middle, ring, and little fingers)
of The stiffnesses
a healthy subject’s rightofhand
MCP werein multiple
estimated fingers
using the proposed (index, middle,
method (Figure 4a). ring,
Simulated-stiffness
a healthy subject’s jointsright
with a hand
higher apparent
were stiffness
estimatedvalue were
usingprepared
the byproposed
ban-
daging the healthy joints (Figure 4b). Each of the four fingers was bandaged to increase
Simulated-stiffness
stiffness in the extension joints
direction.with
For eacha higher apparent
state of stiffness, stiffness
estimation value wer
was performed
three times. The resting angle, θ0 , was measured using a goniometer. The thumb was fixed
aging the healthy joints (Figure 4b). Each of the four fingers was b
with a fixture because thumb movement could have affected the measurement data, and
stiffness in between
the distance the extension
the hand anddirection.
LMC was set to For each state of stiffness, estima
15 cm.
three times. The resting angle, 𝜃 , was measured using a goniome
fixed with a fixture because thumb movement could have affected th
and the distance between the hand and LMC was set to 15 cm.
Robotics 2023, 12, 83 10

Robotics 2023, 12, 83 variable; thus, the angle correction model for the index finger cannot be used
10 of 25for the

fingers.

(a) (b)
Figure
Figure4. 4.(a)(a)Setup
Setupofof the experiment
the experiment to to estimate
estimate the stiffness
the stiffness of a healthy
of a healthy subject’ssubject’s MCP join
MCP joint;
(b) stiff joint with a higher apparent stiffness value due to
stiff joint with a higher apparent stiffness value due to bandaging. bandaging.

It should be noted that, in the experiment, when estimating the MCP stiffness of the
3. Results
index finger, both the original LMC Angle and corrected LMC Angle were used, though for
3.1.the middle, ring,
Analytical andof
Model little
thefingers, their MCP stiffness values were estimated using only
Modular-SECA
the original LMC Angle. This is because that, the regression model for the corrected LMC
Theofcorrection
Angle parameters
the index finger identified
could be built by using were 𝛼 = angle
the reference 1, 𝛼 values
= 0.5,obtained
and 𝛼 from= 1.15. Su
tuting these parameters into Equation (2), the analytical model for the
the 2D marker measurement system; however, due to the difficulty to attach 2D markers toModular-SEC
as the other fingers, it was difficult to derive an angle correction model for them.
follows:
Certainly, one option might be making use of the angle correction model of the index
finger for the other fingers. Actually, as shown 𝑊 + 0.5𝑊 D, for the middle and ring
in Appendix
𝑃 =
fingers, the PIP and MCP joints had a good fit (>0.7) 1.15∆𝑉
with the corrected LMC Angle from
the index finger, but the DIP joint had a poor fit. Nevertheless, the angle correction model
forCompared
the MCP jointtoof the SECA
the index analytical
finger model,
also includes theangle
the DIP coefficient of the increase
as an explanatory variable; in cha
volume
thus, the angle correction model for the index finger cannot be used for the other fingers.strain en
relative to the air pressure input increased, and that of the bending
stored in the torque compensation layer decreased in the actuator’s energy distribu
3. Results
Figure 5 shows a graph comparing the actuator-only bending angles in free space
3.1. Analytical Model of the Modular-SECA
difference is seen between the results of the analytical model, experiment, and FEM
The correction parameters identified were α A = 1, α L = 0.5, and αV = 1.15. Substi-
ulation
tuting for the
these Modula-SECA.
parameters However,
into Equation there is model
(2), the analytical a difference between the isresults o
for the Modular-SECA
SECA analytical model and those of the Modular-SECA.
as follows:
WA + 0.5WL
P= (9)
1.15∆V
Compared to the SECA analytical model, the coefficient of the increase in chamber
volume relative to the air pressure input increased, and that of the bending strain energy
stored in the torque compensation layer decreased in the actuator’s energy distribution.
Figure 5 shows a graph comparing the actuator-only bending angles in free space. No
difference is seen between the results of the analytical model, experiment, and FEM simula-
tion for the Modula-SECA. However, there is a difference between the results of the SECA
analytical model and those of the Modular-SECA.

Figure 5. Actuator-only finger bending angle in free space obtained from analytical models,
and experiment.

3.2. Joint Angle Correction Model with Actuator

Regression models created for each of the DIP, PIP, and MCP joints to correc
 stored in the torque compensation layer decreased in the actuator’s e
Figure 5 shows a graph comparing the actuator-only bending angles
difference is seen between the results of the analytical model, experim
Robotics 2023, 12, 83
ulation for the Modula-SECA. However, there is a difference between
11 of 25
SECA analytical model and those of the Modular-SECA.

Figure 5. Actuator-only finger bending angle in free space obtained from analytical models, FEM,
Figure 5. Actuator-only finger bending angle in free space obtained from ana
and experiment.
and experiment.
3.2. Joint Angle Correction Model with Actuator
Regression models created for each of the DIP, PIP, and MCP joints to correct the joint
3.2. Joint Angle Correction Model with Actuator
angles with soft actuators were as follows:
Regression models created for each of the DIP, PIP, and MCP jo
DIP = −6.77D + 13.71P − 12.27M + 92.36pD + 65.56pP − 40pM − 0.01P· M
joint angles with
−78.93pD soft
· pP − actuators
53.97pP were· pM
· pM − 69.07pD as+follows:
69.86pD · pP· pM (10)
+6.9D · pM + P·(4.66pD − 0.56pP − 15.74pM) − M·(4.79pD − 16.41pM)
𝐷𝐼𝑃 = −6.77 𝐷 + 13.71 𝑃 − 12.27 𝑀 + 92.36 𝑝𝐷 + 65.56 𝑝𝑃 − 40 𝑝𝑀
PIP = 2.55D + 1.69M − 3.32pP − 0.04D · P − 0.08D · M − 0.001D · P· M
−1.6pD · pP + 4.54pD · pM + 4.49pD · pP· pM (11)
+ D ·(1.49pP − 1.2pM) − M·(0.25pD + 0.29pM)

MCP = 1.04D + 0.62M + 12.01pP − 0.88pM − 0.01D · P

+1.35pD · pP − 0.67pD · pM − 3.04pP· pM − 1.27pD · pP· pM (12)
− D ·(1.1pP − 0.92pM) − P·(0.12pD − 0.8pP) − M·(1.07pP − 0.42pM)

DIP( D ), PIP( P) and MCP( M ) are LMC Angles, and pDIP( pD ), pPIP( pP) and
pMCP( pM) are dummy Boolean variables indicating whether the soft actuator corre-
sponding to each joint was pressurized or not. The adjusted coefficients of determination
were 0.895 for the DIP joint, 0.992 for the PIP joint, and 0.995 for the MCP joint. The PIP and
MCP joints were above 0.9, and the DIP joint was slightly below 0.9. The model consisted
of around 15 variables, resulting from the variable selection process. As shown, the models
for each joint contain the LMC Angle of other joints.

3.3. Joint Stiffness Estimation Using Mannequin Hand

3.3.1. Regression Model Joint Angle Correction Performance
The 2D angle, original LMC Angle, and corrected LMC Angle (from the proposed
model) of the three joints (DIP, PIP, and MCP) under each pressure from 0 kPa to
80 kPa during the stiffness estimation are shown in Figure 6. To compare the difference in
correction performance between the proposed regression model and the regression model
derived from the method of the previous study [27], third-order regression models were
generated using only the angle of the target joint to be corrected as the explanatory variable
(Equations (13)–(15)). The corrected LMC Angle (from the previous study) is also shown
in Figure 6.
DIP = 0.006DIP3 − 0.32DIP2 + 5.95DIP − 24.96 (13)
 Robotics 2023, 12, 83 12 of 25

PIP = 0.0002PIP3 − 0.02PIP2 + 1.37PIP + 13.95 (14)

Robotics 2023, 12, 83
MCP = 0.002MCP3 − 0.16MCP2 + 5.22MCP − 22.71 (15)

(a) Low Stiffness

(b) Middle Stiffness

Figure 6.Relationship
Figure 6. Relationship between
between pressure
pressure and
and joint joint
angle angle in
in stiffness stiffnessexperiments.
estimation estimation experiments
The correction performance of each regression model was evaluated in terms of the
Table 4. Prediction accuracy of LMC Angle correction model with the actuator (RMSE). Tw
root-mean-square error (RMSE) values between the 2D Angle and LMC Angle in Figure 6
ues are 4).
(Table underlined
As shown,where
exceptthe
theRMSE of the corrected
DIP measurement LMC
on the angle
middle from proposed
stiffness method is l
level (double
than the RMSE of the original LMC angle. Among them, the one whose RMSE
underlined in Table 4), the corrected LMC Angle under the proposed model improved is larger than
RMSE of the corrected
the prediction accuracyLMC angletofrom
compared previousLMC
the corrected studyAngle
is double
from underlined.
the previous study.
The model in the previous study did not include soft actuator pressure information as an
Corrected LMC Angle from Corrected LMC Angl
explanatory variable, so the model had
Stiffness Level Joint Original LMC Angle (deg) to be changed when the combination of joints to be
Proposed
pressured changed. On the other hand, compared Model
with (deg)
original LMC Angle, Previous
the proposedStudy (d
DIP model showed improvement,
15.31 except for the two measurements
5.51 underlined in Table 4. 10.22
Low Stiffness PIP 4.98 4.97 10.94
MCP 16.39 5.07 10.93
DIP 11.74 14.67 8.80
Middle Stiffness PIP 8.34 14.80 23.33
MCP 23.55 8.83 19.66

3.3.2. Joint Stiffness Estimation Accuracy

The results of the estimated MCP joint stiffness of the index finger of the man
hand are shown in Table 5. Comparing Method 1 (the SECA model with 2D Angle
 Robotics 2023, 12, 83 13 of 25

Table 4. Prediction accuracy of LMC Angle correction model with the actuator (RMSE). Two values
are underlined where the RMSE of the corrected LMC angle from proposed method is larger than the
RMSE of the original LMC angle. Among them, the one whose RMSE is larger than the RMSE of the
corrected LMC angle from previous study is double underlined.

Corrected LMC Angle from Corrected LMC Angle from

Stiffness Level Joint Original LMC Angle (deg)
Proposed Model (deg) Previous Study (deg)
DIP 15.31 5.51 10.22
Low Stiffness PIP 4.98 4.97 10.94
Robotics 2023, 12, 83 MCP 16.39 5.07 10.93
DIP 11.74 14.67 8.80
Middle Stiffness PIP 8.34 14.80 23.33
MCP 23.55 8.83 19.66
experiment. Even though the accuracy of the analytical model for the Modular-
better
3.3.2. than
Joint that of
Stiffness the SECA
Estimation model for the reference value of the joint angle (2D A
Accuracy
deviation of of
The results thethejoint angle
estimated estimation
MCP produces
joint stiffness better
of the index fingerperformance for the an
of the mannequin
model
hand are of the SECA
shown in Tablemodel instead
5. Comparing of Modular-SECA,
Method 1 (the SECA model which
with 2Ddeviates from our
Angle mea-
surement) and Method
tion. Therefore, 2 (thethe
despite Modular-SECA
improvement model within
made 2Dboth
Angle measurement),
the it is and L
analytic model
clear that the corrections to the analytical model improved the accuracy of the stiffness
gle measurement, both of them need further enhancement.
estimation. Comparing Method 3 (the SECA model with the original LMC Angle mea-
surement) and Method 5 (the SECA model with the corrected LMC Angle measurement),
Table
and 5. Results
Method 4 (the of stiffness estimation
Modular-SECA experiments
Model with the originalusing
LMC the mannequin
Angle measurement) hand.
andThe smal
6among the six methods
(the Modular-SECA modelis colored
with thered.
corrected LMC Angle measurement), it seems obvi-
ous that the correction of the LMC Angle improves the accuracy of the stiffness estimation.
Estimated Value
However, the comparisons between (NMethods
mm/deg) (Error
3 and = Estimated
4, and Methods 5Value
and 6 −disclose
Targetthat
Value)
Methodof1 stiffness
the accuracy Method
estimation2 deteriorated
Method when
3 theMethod 4
LMC Angle wasMethod
used due 5 to Met
Target Value the correction of the analytical
Modular-SECA model. Figure 7 shows the change
Modular-SECA in the stiffness value Modul
Stiffness Level SECA Model
estimated by the joint angle at a 40 SECA
kPa input Model
air pressure in the low SECA
stiffness Model
experiment.
(N mm/deg) Model Model M
Even though the accuracy of the analytical model for the Modular-SECA is better than that
Original LMC Original LMC Corrected LMC Correc
of the2D
SECA model for2D
Angle the Angle
reference value of the joint angle (2D Angle), a deviation of the
Angle for the analytical
joint angle estimation produces better performance Angle model of Angle
the SECA An
Low Stiffness 0.552 model2.17 (+1.62)
instead 0.87 (+0.32)which−0.13
of Modular-SECA, (−0.68)
deviates −0.31
from our (−0.86) Therefore,
expectation. 0.82 (+0.27)
despite 0.04
the improvement made in both the analytic model and LMC Angle measurement, both of
Middle Stiffness 1.686 4.98 (+3.30) 3.27 (+1.58) −0.09 (−1.78) −0.21 (−1.89) 1.06 (−0.63) 0.38
them need further enhancement.

Figure
Figure Relationship
7. 7. between
Relationship joint angle
between jointand estimated
angle stiffness value
and estimated in low stiffness
stiffness value in condition.
low stiffness co

3.4. Multiple Finger Joint Stiffness Estimation

First, Table 6 shows that the joint angle during rest was larger in the healthy j
the other hand, the joint angle when the soft actuator was attached and not pneum
pressurized showed a flexion angle, while the original LMC Angle showed an ex
angle. Figure A3 shows that when the LMC Angle showed an extension angle in t
joint, the smaller the flexion angle measured with the goniometer was, the larger t
Robotics 2023, 12, 83 14 of 25

Table 5. Results of stiffness estimation experiments using the mannequin hand. The smallest error among the six methods is colored red.

Estimated Value (N mm/deg) (Error = Estimated Value − Target Value)

Method 1 Method 2 Method 3 Method 4 Method 5 Method 6
Target Value
Stiffness Level SECA Model Modular-SECA Model SECA Model Modular-SECA Model SECA Model Modular-SECA Model
(N mm/deg)
2D Angle 2D Angle Original LMC Angle Original LMC Angle Corrected LMC Angle Corrected LMC Angle
Low Stiffness 0.552 2.17 (+1.62) 0.87 (+0.32) −0.13 (−0.68) −0.31 (−0.86) 0.82 (+0.27) 0.04 (−0.52)
Middle Stiffness 1.686 4.98 (+3.30) 3.27 (+1.58) −0.09 (−1.78) −0.21 (−1.89) 1.06 (−0.63) 0.38 (−1.30)
Robotics 2023, 12, 83 15 of 25

3.4. Multiple Finger Joint Stiffness Estimation

First, Table 6 shows that the joint angle during rest was larger in the healthy joint. On
the other hand, the joint angle when the soft actuator was attached and not pneumatically
pressurized showed a flexion angle, while the original LMC Angle showed an extension
angle. Figure A3 shows that when the LMC Angle showed an extension angle in the MCP
joint, the smaller the flexion angle measured with the goniometer was, the larger the LMC
Angle extension angle was. Therefore, the flexion angle at 0 kPa was larger for the stiff
joint. In other words, the stiff joint had stronger resistance in the extension direction, and
the amount of angle change caused by the soft actuator was smaller. Therefore, the stiffness
value of the stiff joint was higher than that of the healthy joint, and bandaging was an
appropriate method to increase the apparent stiffness value.

Table 6. The joint angles of the subject’s index finger during rest and with the actuator attached.

Resting Angle (deg) Joint Angle at 0 kPa (deg)

Goniometer Value LMC Angle LMC Angle
Healthy Joint 44 21.76 −15.03
Stiff Joint 27 21.06 −12.55

The mean and standard deviation of three measurements of the stiffness values esti-
mated for the MCP joints of the fingers are shown in Table 7.

Table 7. Results of an experiment estimating the stiffness of the MCP joint in multiple fingers of a
subject using the analytical model of the Modular-SECA.

Estimated Value (N mm/deg)

Index Middle Ring Little
Original LMC Angle 0.26 ± 0.02 0.20 ± 0.01 0.33 ± 0.03 0.26 ± 0.02
Healthy Joint
Corrected LMC Angle −0.09 ± 0.02 - - -
Original LMC Angle −0.36 ± 0.19 0.01 ± 0.08 0.34 ± 0.14 1.13 ± 0.02
Stiff Joint
Corrected LMC Angle −0.17 ± 0.12 - - -

For the little finger, the stiffness estimates for stiff joints were higher than those for
healthy joints. Though, for the ring finger, it was difficult to distinguish between healthy
and stiff joints from the estimates. For the MCP joint of the index finger and middle finger,
this was even reversed.
Another issue is that of negative estimates of stiffness. As can be seen in Figure 7, the
accuracy of the joint angles was low because the estimated values were negative depending
on the joint angle. The angle correction model for the index finger was generated using
joint angle data obtained from a mannequin hand. Its applicability to the human hand was
low, and the lack of LMC Angle correction may have resulted in negative values.
Moreover, because the fingers were separately restrained by bandaging (Figure 4b),
the strength of the restraint was different for each finger, causing a large difference in
the apparent stiffness for each finger. These issues revealed that stiffness estimation for a
human finger is not so reliable, though the possibility for simultaneous joints of multiple
fingers was shown.

4. Discussion
The results show that the SECA and Modular-SECA have different bending perfor-
mances, and the interaction of joints affected the output of LMC when the Modular-SECA
was attached. Furthermore, it was confirmed that the accuracy of the stiffness estimation
with the Modular-SECA could be improved by correcting the analytical model of stiffness
estimation and the joint angles.
 Robotics 2023, 12, 83 16 of 25

4.1. Analytical Model of Modular-SECA

The free space bending analysis results suggest a difference in the energy distribution
in the actuator between the Modular-SECA and SECA. It was confirmed that the chamber
volume increase, ∆V, relative to the air pressure input was larger for the Modular-SECA
than that for the SECA. The bending strain energy, WL , stored in the torque compensation
layer was smaller for the Modular-SECA. This may have been due to the rigid connectors
attached to both ends of the Modular-SECA, which reduced the energy allocated to WL
within the energy conservation law. In previous studies, WL was modeled as a cantilever
beam subjected to a pure bending moment [17]. However, the Modular-SECA is considered
to be subjected to an additional bending moment in the torque compensation layer. There-
fore, at the same air pressure input, the energy required to bend the torque compensating
layer is reduced, and the volume increase in the chamber is considered to have increased.
From Figure 8, the difference between the 2D Angle and the bending angle obtained
from the Modular-SECA analytical model is smaller than that in the SECA analytical model
results. However, the estimated values for Method 2 (Modular-SECA Model with 2D Angle
ics 2023, 12, 83 measurement) in Table 5 did not reach the target values. This may be due to the lack of
correction in the Modular-SECA analytical model. In particular, the difference between
the experimental and Modular-SECA model bending angles at 0 kPa suggests that the
correction may be insufficient. This difference is thought to have persisted afterward,
preventing the estimation results of Method 2 from reaching the target values. A feature
pressure and that the force in the bending direction is greater. H
of the SECA is that the torque compensation layer provides a supplementary bending
pressure,
moment thetheauxiliary
to return bending
finger to its initial moment
position under reducedprovided by5 shows
pressure. Figure the torque
that the bending angle of Modular-SECA is larger than that of the SECA under applied
the Modular-SECA
pressure and that the force in is thesmaller than isthat
bending direction of However,
greater. the SECA, underand
reducedthe forc
tion may
pressure, be smaller.
the auxiliary (𝜃 −provided
bending moment 𝜃 ) in Equation
by the (8) is alayer
torque-compensating term that ex
of the
Modular-SECA is smaller than that of the SECA, and the force in the extension direction
bending
may moment
be smaller. (θ − θ0 )2 inof the torque
Equation compensation
(8) is a term layer
that expresses the effect onbending
of the the bendi
moment
moment provided
of the by thelayer
torque compensation torque-compensation
on the bending angle. If thelayer
bendingof the Modu
moment
provided by the torque-compensation layer of the Modular-SECA is small, a correction
rection may
parameter parameter
also need to may alsoforneed
be provided to be provided for this term.
this term.

Figure 8. Relationship between pressure and joint angle of MCP in low stiffness estimation (the
Figure 8. Relationship between pressure and joint angle of MCP in low s
2D Angle is the reference value). For the SECA model and Modular-SECA model, the joint angles
Angle
were is the
calculated reference
backward value).
by substituting theFor the
spring SECA
constant model
of the and and
torsion spring Modular-SECA
the input air m
pressure into Equation (8). The upper
calculated backward by substituting the limit angle is 0.7θ 0 , which is the upper limit of the
spring constant of the torsion spstiffness
estimation range θ ∈ [0, 0.7θ0 ).
sure into Equation (8). The upper limit angle is 0.7𝜃 , which is the uppe
mation range 𝜃 ∈ [0, 0.7𝜃 ).

4.2. Finger Joint Angle Measurement using Leap Motion Controller

Robotics 2023, 12, 83 17 of 25

4.2. Finger Joint Angle Measurement Using Leap Motion Controller

First, Appendix E shows that the LMC Angle change trend changes with the attach-
ment of the soft actuator. In Figure A2a,d, the trend of the LMC Angle change around
80 kPa is altered compared to that at the previous pressure point. Briefly, 80 kPa is when
the mannequin hand’s index finger is bent to the point where the soft actuator is visible
from LMC’s camera. Therefore, LMC recognized the soft actuator as a part of the finger,
and the output angle changed.
When the soft actuator was applied, the angle correction model included the angle of
other joints, suggesting that LMC Angle is affected by the interaction between joints within
the same finger. Table 4 shows that the correction performance of the angle correction
model is high for joints with large angle changes when the soft actuator is pressurized.
However, it may be low for joints that are not pressurized and have small angle changes.
This may be due to biased training data for the joints where the soft actuators are not
pressurized, resulting in lower model prediction accuracy. It also suggests that including
the angles of other joints in the explanatory variables may lead to an increase in correction
accuracy but may also lead to a decrease. Correction performance varies with stiffness level
in Table 4 because correction accuracy varied with the angles of other joints. Although the
model of the previous study can improve correction performance by increasing the order
of the regression model, it requires separate models for each combination of joints to which
the soft actuator applies pressure.
Figures 6 and 8 show that the upper limit of the stiffness estimation range using the
LMC Angle is wider than that using the 2D Angle because the LMC Angle is output smaller
than the 2D Angle even after correction. In Figure 8, the stiffness estimation range for
2D Angle is 0 kPa to 50 kPa, while corrected LMC Angle is 0 kPa to 60 kPa; the stiffness
estimation using the LMC Angle includes a greater stiffness value at each pressure in
the final estimate stiffness value than that using the 2D Angle. Therefore, to improve
the estimation performance, it is considered effective not only to improve the correction
accuracy of the LMC Angle but also to change the stiffness estimation range.
When the Modular-SECA was attached to a human hand, MCP joints were constantly
flexed during the experiment, while the LMC Angle showed an extension angle. Since none
of the LMC Angle of the MCPs showed an extension angle in the data when the regression
model was created, the correction performance of the regression model at the extension
angle was insufficient, leading to the under-correction of the LMC Angle.
Since the angle correction model for the index finger could not be applied to the
middle and ring fingers, it is necessary to derive an angle correction model for each finger
separately. However, since it is difficult to measure joint angles using 2D markers for the
middle and ring fingers, it is not possible to derive an angle correction model using the
same method as in this study. Therefore, it is thought that the angle correction model for
the index finger can be applied by excluding from the candidate explanatory variables for
the regression model for the index finger those variables that include the angle of the DIP
joint, which has a low degree of fit with that of the index finger.

4.3. Proposed Method for Finger Joint Stiffness Estimation

The stiffness estimation using LMC and the Modular-SECA showed that the correction
of the analytical model and joint angle improves the stiffness estimation results. However,
the current estimation performance of the proposed method is insufficient due to the
lack of both corrections, and the correction method needs to be improved. The final
stiffness estimate is the average for each pressure in the stiffness estimation range, but
variations exist among the estimates for each pressure. Therefore, the performance of
the proposed method may be improved by weighting the more reliable estimates in the
stiffness estimation range.
Robotics 2023, 12, 83 18 of 25

4.4. Contribution and Limitations

By using joint modular soft actuators, high bending performance can be achieved with
minimal energy for telerehabilitation, and individual differences such as different finger
sizes and different joint stiffness values can be accommodated with minimal effort.
Our contribution includes the following:
1. To the best of our knowledge, for the first time, finger joint stiffness estimation for
using joint modular soft actuators was proposed and implemented. The accuracy
of stiffness estimation needs further improvement, though the possibility has been
shown and the direction for improvement has been made clear. In addition, the
proposed stiffness estimation method enables the simultaneous measurement of
multiple fingers, although the analytical model and the correction method for joint
angles need to be improved.
2. The analytic model was updated from a previous model for the SECA, a whole-finger
soft actuator, with additional parameters for accounting for the features of the joint
modular soft actuator. The model could be regressed from experimental data and
give a better prediction of the behavior of the joint modular actuators.
3. Leap Motion Controller was used for the first time for the angle measurement of
joints of fingers supported by the joint modular soft actuators. A regression model
was proposed to correct measurement results, taking into consideration multiple joint
interaction intensified by wearing the soft actuators, and the influence of the actuation
of the joint modular soft actuators. The effectiveness of the correction was validated
by the experiment results. LMC can be used to measure the angles of finger joints not
only when soft actuators are attached, regardless of whether they are actuated or not,
but also when soft actuators are not attached.
On the other hand, there are several methodological and implementational limitations:
1. The high stiffness of joint cannot be estimated with the Modular-SECA. The refer-
ence value for high stiffness (a MAS (modified Ashworth scale) score of about 3) is
9.60 N mm/deg. Because the extension of the highly stiff finger caused by wearing
the Modular-SECA was small compared with that of the SECA, the joint angle, θ, was
not within the stiffness estimation range, θ ∈ [0, 0.7θ0 ). To estimate high stiffness
as well, it is necessary to design a joint modular actuator that can extend the finger
sufficiently with its intrinsic stiffness, or add another component to actively extend
the joint extension.
2. Only the stiffness estimation of MCP joints of multiple fingers was tested. The stiffness
estimation of DIP and PIP joints of multiple fingers needs to be tested in future.
3. The regression model used to correct the LMC Angle generated using data acquired
from mannequin hands did not apply to human hands. Although mannequin hands
are effective for multiple measurements due to their low data variability, they are
unsuitable for model generation because they do not reproduce human finger move-
ments such as intra-finger constraint. Therefore, if an angle correction model is
generated using data acquired from a human hand, it may apply to both mannequin
hands and human hands. Moreover, this may allow the generated regression model
to be adaptable to the individual difference of hand dimensions.
4. The manually controlled input air pressure affected the human hand’s stiffness estima-
tion. The speed at which the joint moves, which is affected by the input air pressure,
is a factor that influences joint stiffness [37,38]. In the case of manual control, the
rotational speed of the joints differed from measurement to measurement, which may
have increased the variation in data. Thus, automatic pneumatic input control is
necessary to realize the proposed method.
5. In this study, we did not test stiffness estimation in a telerehabilitation setting. In the
near future, it is necessary to verify the effectiveness of the proposed method in such
a setting.
Robotics 2023, 12, 83 19 of 25

5. Conclusions
In this study, we proposed a method combining the Modular-SECA and LMC for
the stiffness estimation of finger joints in a telerehabilitation environment. The SECA
analytical model for stiffness estimation was modified to be applied to the Modular-SECA,
and the LMC Angle was improved by the regression models to improve the accuracy of the
proposed method. It was also shown that using the angle information of other joints and
the dummy variables indicating whether or not the soft actuator is pressurized is effective
for the correction model of the LMC Angle with the soft actuator. Furthermore, the results
suggest that the proposed method could be used in a telerehabilitation environment to
evaluate the spasticity condition in case a more precise correction of the analytical model
and the LMC Angle becomes available. Since the simultaneous estimation of multiple
fingers’ MCP joints is feasible using the proposed method in this study, it is necessary
to make it possible to estimate DIP and PIP joints and joints with high stiffness values.
Moreover, in the near future, it is necessary to verify the effectiveness of the proposed
method in a telerehabilitation setting.

Author Contributions: Conceptualization, F.M. and W.Y.; methodology, F.M., S.K. and P.E.T.V.;
software, F.M.; validation, F.M. and P.E.T.V.; investigation, F.M.; data curation, F.M.; writing—original
draft preparation, F.M.; writing—review and editing, S.K., P.E.T.V., M.-T.K., Y.-H.H., S.Y.H., J.G.-T. and
W.Y.; supervision, W.Y. All authors have read and agreed to the published version of the manuscript.
Funding: This work was supported by Grant-in-Aid for Scientific Research (B), JSPS KAKENHI
Grant Number 22H03450.
Acknowledgments: The authors would like to thank Zhongchao Zhou, Yuxi Lu and Ema Oba for
their support in the experiments.
Conflicts of Interest: The authors declare no conflict of interest.

Abbreviations

LMC Leap Motion Controller

DIP Distal interphalangeal (joint)
PIP Proximal interphalangeal (joint)
MCP Metacarpophalangeal (joint)
2D Angle Joint angle calculated using the two-dimensional marker
LMC Angle Joint angle calculated using LMC
W Midpoint of the wrist joint
M Midpoint of the MCP joint
P Midpoint of the PIP joint
D Midpoint of the DIP joint
T Tip of the finger
SECA soft-elastic composite actuator
Modular-SECA Joint modular soft-elastic composite actuator

Appendix A
The finite element model (FEM) of the Modular-SECA was built on COMSOL Mul-
tiphysics 5.6® . Dragon skin 10 was modeled as a Yeoh third-order hyperelastic material.
The Kevlar fibers, stainless-steel plate and PLA were all modeled as linear elastic materials.
The material parameters used for Dragon Skin 10 and the stainless-steel plate were the
same as those in Section 2.1.2. The Kevlar fibers were modeled with a Young’s modulus
(E-31670 MPa), Poisson’s ratio (v-0.36) and density (ρ-1440 kg/m3 ) [39]. For the PLA the
material parameters used were E-4.4 GPa, v-0.3 and ρ-1240 kg/m3 [40]. A stationary study
for increasing input air pressure of the actuator from 0 kPa to 100 kPa (steps of 10 kPa) was
conducted. Gravity was included in the model. A symmetry constraint was used to take
advantage of the symmetrical conditions of the model and reduce computational times.
 lus (E-31670 MPa), Poisson’s ratio (v-0.36) and density (ρ-1440 kg/m3) [39]. For the PLA
the material parameters used were E-4.4 GPa, v-0.3 and ρ-1240 kg/m3 [40]. A stationary
study for increasing input air pressure of the actuator from 0 kPa to 100 kPa (steps of 10
Robotics 2023, 12, 83
kPa) was conducted. Gravity was included in the model. A symmetry constraint was used
20 of 25
to take advantage of the symmetrical conditions of the model and reduce computational
times. The meshing was carried out using tetrahedral elements. Gravity was incorporated
into
Thethe model.was
meshing Figure A1out
carried shows
usingthe model bending
tetrahedral at Gravity
elements. different input
was air pressures.
incorporated into
the model. Figure A1 shows the model bending at different input air pressures.

(a) P = 0 kPa (b) P = 50 kPa (c) P = 100 kPa

Figure A1. Bending of Modular-SECA’s FEM at different input air pressures (P).
Figure A1. Bending of Modular-SECA’s FEM at different input air pressures (P).
Appendix B
Appendix B
The bending angles of the three joints (DIP, PIP, and MCP) of the mannequin hand’s
The
index bending
finger angles of the
were determined fromthree
0◦ tojoints
90◦ and(DIP, PIP, and
measured. TheMCP) of thewere
joint angles mannequin
measuredhand’s
index fingerusing
separately werea determined
goniometer andfrom 0° marker
a 2D to 90° and measured.
detection system.The
Forjoint angles were meas-
this measurement,
the index
ured fingerusing
separately of the amannequin
goniometer hand
andwith
a 2Dfrictional
markerresistance
detectionwas used For
system. to hold
thisitmeasure-
in
placethe
ment, at aindex
given finger
angle. Each ofmannequin
of the the three joints waswith
hand measured ten times,
frictional and the
resistance difference
was used to hold
it between
in place the
at ajoint angles
given measured
angle. Each ofwiththe the goniometer
three joints wasand that measured
measured with the
ten times, and2Dthe dif-
marker detection system was calculated. Table A1 shows the mean and standard deviation
ference between the joint angles measured with the goniometer and that measured with
of the difference in joint angles measured via the two methods. Three joints had an error of
the 2D marker
around detection
2◦ , indicating that system was
the angles calculated.
were Table A1 shows the mean and standard
equivalent.
deviation of the difference in joint angles measured via the two methods. Three joints had
anTable
error ofThe
A1. around
errors 2°, indicating
between that measured
joint angles the angles were
with equivalent.
a goniometer and 2D marker detection
system (mean ± SD).

Joint Angle Error (deg)

DIP 1.2 ± 0.7
PIP 1.4 ± 0.8
MCP 1.4 ± 0.7

Appendix C
The effect of joint interaction in the same finger on the LMC Angle was examined. Soft
actuators were attached to the mannequin hand, and the soft actuators corresponding to the
DIP, PIP, and MCP joints were selected for each of the seven combinations of measurements
(Figure A2). The 2D Angle and LMC Angle were calculated. A joint modular soft actuator
without a torque compensation layer was used to verify the performance of LMC estimation
over a wide range of bending angles. In some cases, the joint angle of the mannequin hand
showed an extension (the opposite direction of flexion) angle instead of a flexion angle.
Therefore, a flexion/extension judgment was made, and the sign of the joint angle was
set to negative in the case of an extension angle. In the 2D marker detection system, the
decision was based on the sign of the value of the outer product of the vectors (θ MCP is
−−→ −−→
W M × WP ). Since LMC is a three-dimensional coordinate system, the x coordinate of the
−−→ −−→
outer product of the two vectors (θ MCP is W M × MP ) in the measurement method shown
in Figure 2b was the extension angle if it was positive.
The trend of angle change differed depending on the combination of joints to be
pressurized. Even when the 2D Angle did not change, the LMC Angle changed by the
angle of the other joints. This suggests that the interaction between joints in the same finger
affects the LMC Angle.
 (𝜃 is 𝑊𝑀⃗ × 𝑊𝑃⃗). Since LMC is a three-dimensional coordinate system, the x coordi-
nate of the outer product of the two vectors (𝜃 is 𝑊𝑀⃗ × 𝑀𝑃⃗ ) in the measurement
method shown in Figure 2b was the extension angle if it was positive.
The trend of angle change differed depending on the combination of joints to be pres-
Robotics 2023, 12, 83 surized. Even when the 2D Angle did not change, the LMC Angle changed by the angle 21 of 25
of the other joints. This suggests that the interaction between joints in the same finger
affects the LMC Angle.

(a) Pressure on three joints; DIP + PIP + MCP (𝑝𝐷𝐼𝑃: 1; 𝑝𝑃𝐼𝑃: 1; 𝑝𝑀𝐶 : 1)

Pressure on two joints

(b) DIP + PIP (𝑝𝐷𝐼𝑃: 1; 𝑝𝑃𝐼𝑃: 1; 𝑝𝑀𝐶 : 0)

Robotics 2023, 12, 83 20 of 23

(c) DIP + MCP (𝑝𝐷𝐼𝑃: 1, 𝑝𝑃𝐼𝑃: 0, 𝑝𝑀𝐶𝑃: 1)

(d) PIP + MCP (𝑝𝐷𝐼𝑃: 0; 𝑝𝑃𝐼𝑃: 1; 𝑝𝑀𝐶𝑃: 1)

Pressure on one joint

(e) DIP (𝑝𝐷𝐼𝑃: 1; 𝑝𝑃𝐼𝑃: 0; 𝑝𝑀𝐶: 0)

(f) PIP (𝑝𝐷𝐼𝑃: 0; 𝑝𝑃𝐼𝑃: 1; 𝑝𝑀𝐶𝑃: 0)

Figure A2. Cont.
 Robotics 2023, 12, 83 22 of 25

(f) PIP (𝑝𝐷𝐼𝑃: 0; 𝑝𝑃𝐼𝑃: 1; 𝑝𝑀𝐶𝑃: 0)

(g) MCP (𝑝𝐷𝐼𝑃: 0; 𝑝𝑃𝐼𝑃: 0; 𝑝𝑀𝐶𝑃: 1)

Robotics 2023, 12, 83

Figure A2. LMC Angle change due to interaction between joints in the same finger when
ators were attached. 𝑝𝐷𝐼𝑃, 𝑝𝑃𝐼𝑃 and 𝑝𝑀𝐶𝑃 are dummy Boolean variables indicating 1
pressure is input to the soft actuator corresponding to the joint and 0 when it is not.
Figure A2. LMC Angle change due to interaction between joints in the same finger when soft
actuators were attached. pDIP, pPIP and pMCP are dummy Boolean variables indicating 1 when air
Appendix D
pressure is input to the soft actuator corresponding to the joint and 0 when it is not.
With no soft actuator attached, one subject was asked to bend the right hand
Appendix D
middle, ring, and little fingers’ joints and maintain a constant angle. The angle o
these 12 jointsactuator
With no soft attached, one
was measured subject
using LMC wasand
asked to bend the right
a goniometer, hand’s
and thisindex,
was carried
middle, ring, and little fingers’ joints and maintain a constant angle. The angle of each
times. From the measured angle data of the index, an angle correction model of th
of these 12 joints was measured using LMC and a goniometer, and this was carried out
was
ten created
times. Fromusing a third-order
the measured regression
angle data model
of the index, (Figure
an angle A3). When
correction model that
of theangle co
model
index wasofcreated
the index
usingwas applied to
a third-order the other
regression fingers,
model the
(Figure adjusted
A3). When thatcoefficients
angle of
correction modelobtained
nation were of the index was
and applied
these aretoshown
the other
infingers,
Table the
A2.adjusted
The PIP coefficients
and MCP of for the
determination were obtained and these are shown in Table A2. The PIP and MCP for the
and ring fingers had a good fit of 0.7 or better, while the others resulted in a poor
middle and ring fingers had a good fit of 0.7 or better, while the others resulted in a poor fit.
suggests
This that
suggests thateven
evenwhen softactuators
when soft actuatorsareare applied,
applied, the trend
the trend of angular
of angular change maychange m
fromfrom
differ finger to to
finger finger.
finger.

Figure A3. Degree of angular change tendency conformity between the index finger and other fingers.
Figure A3. Degree of angular change tendency conformity between the index finger and othe

Table A2. The goodness of fit between the index finger’s regression model and the othe
angular change trend.

Adjusted R-Square
Middle Ring Littl
DIP 0.149 0.682 0.472
PIP 0.731 0.764 0.065
MCP 0.749 0.894 0.215
 Figure A3. Degree of angular change tendency conformity between the index finger and othe

Robotics 2023, 12, 83 Table A2. The goodness of fit between the index finger’s regression model 23and
of 25the othe

angular change trend.

Adjusted
Table A2. The goodness of fit between the index finger’s regression R-Square
model and the other fingers’
angular change trend. Middle Ring Little
DIP 0.149
Adjusted R-Square 0.682 0.472
PIP Middle 0.731 Ring 0.764 Little 0.065
DIP MCP 0.149 0.749 0.682 0.894 0.472 0.215
PIP 0.731 0.764 0.065
MCP 0.749 0.894 0.215
Appendix E
TheEbending
Appendix of the three joints (DIP, PIP, and MCP) of the mannequin hand
finger,
The 2D Angle
bending andthree
of the LMC Angle
joints (DIP,were measured,
PIP, and respectively.
MCP) of the For this
mannequin hand’s measurem
index
index finger of the mannequin hand with frictional resistance was used to hold it
finger, 2D Angle and LMC Angle were measured, respectively. For this measurement, the
index finger of
at a given the mannequin
angle. hand with frictional
This measurement resistance
data and was used
the joint to hold
angles withit inthe
place at actuato
soft
a given angle. This measurement data and the joint angles with the soft actuator measured
ured in Appendix B are shown in Figure A4. The trend of the LMC Angle chan
in Appendix B are shown in Figure A4. The trend of the LMC Angle change with and
and without
without the soft the soft differs.
actuator actuator differs.

Figure
Figure A4.
A4. Angular
Angular trendtrend comparison
comparison with andwith andsoft
without without
actuator.soft actuator.
References
1. Feigin, V.L.; Forouzanfar, M.H.; Krishnamurthi, R.; Mensah, G.A.; Connor, M.; Bennett, D.A.; Moran, A.E.; Sacco, R.L.; Anderson,
L.; Truelsen, T.; et al. Global and Regional Burden of Stroke during 1990-2010: Findings from the Global Burden of Disease Study
2010. Lancet 2014, 383, 245–254. [CrossRef] [PubMed]
2. Kwakkel, G.; Kollen, B.J.; Van der Grond, J.V.; Prevo, A.J.H. Probability of Regaining Dexterity in the Flaccid Upper Limb: Impact
of Severity of Paresis and Time since Onset in Acute Stroke. Stroke 2003, 34, 2181–2186. [CrossRef] [PubMed]
3. Mouri, T.; Kawasaki, H.; Aoki, T.; Nishimoto, Y.; Ito, S.; Ueki, S. Telerehabilitation for Fingers and Wrist Using a Hand
Rehabilitation Support System and Robot Hand. IFAC Proc. Vol. 2009, 42, 603–608. [CrossRef]
4. Nam, C.; Zhang, B.; Chow, T.; Ye, F.; Huang, Y.; Guo, Z.; Li, W.; Rong, W.; Hu, X.; Poon, W. Home-Based Self-Help Telerehabilitation
of the Upper Limb Assisted by an Electromyography-Driven Wrist/Hand Exoneuromusculoskeleton after Stroke. J. Neuroeng.
Rehabil. 2021, 18, 137. [CrossRef] [PubMed]
5. Placidi, G.; Di Matteo, A.; Lozzi, D.; Polsinelli, M.; Theodoridou, E. Patient-Therapist Cooperative Hand Telerehabilitation
through a Novel Framework Involving the Virtual Glove System. Sensors 2023, 23, 3463. [CrossRef]
6. Arya, K.N.; Pandian, S.; Verma, R.; Garg, R.K. Movement Therapy Induced Neural Reorganization and Motor Recovery in Stroke:
A Review. J. Bodyw. Mov. Ther. 2011, 15, 528–537. [CrossRef]
7. Signal, N.; Martin, T.; Leys, A.; Maloney, R.; Bright, F. Implementation of Telerehabilitation in Response to COVID-19: Lessons
Learnt from Neurorehabilitation Clinical Practice and Education. N. Z. J. Physiother. 2020, 48, 117–126. [CrossRef]
8. Soopramanien, A.; Jamwal, S.; Thomas, P.W. Digital Health Rehabilitation Can Improve Access to Care in Spinal Cord Injury in
the UK: A Proposed Solution. Int. J. Telerehabil. 2020, 12, 3–16. [CrossRef]
9. Auger, L.-P.; Moreau, E.; Côté, O.; Guerrera, R.; Rochette, A.; Kairy, D. Implementation of Telerehabilitation in an Early Supported
Discharge Stroke Rehabilitation Program before and during COVID-19: An Exploration of Influencing Factors. Disabilities 2023, 3,
87–104. [CrossRef]
10. Dietz, V.; Sinkjaer, T. Spastic Movement Disorder: Impaired Reflex Function and Altered Muscle Mechanics. Lancet Neurol. 2007,
6, 725–733. [CrossRef]
11. Sadarangani, G.P.; Jiang, X.; Simpson, L.A.; Eng, J.J.; Menon, C. Force Myography for Monitoring Grasping in Individuals with
Stroke with Mild to Moderate Upper-Extremity Impairments: A Preliminary Investigation in a Controlled Environment. Front.
Bioeng. Biotechnol. 2017, 5, 42. [CrossRef] [PubMed]
Robotics 2023, 12, 83 24 of 25

12. Plantin, J.; Pennati, G.V.; Roca, P.; Baron, J.C.; Laurencikas, E.; Weber, K.; Godbolt, A.K.; Borg, J.; Lindberg, P.G. Quantitative
Assessment of Hand Spasticity after Stroke: Imaging Correlates and Impact on Motor Recovery. Front. Neurol. 2019, 10, 836.
[CrossRef] [PubMed]
13. Tsuji, T.; Ota, T.; Kimura, A.; Chino, N.; Ishigami, S. A Study of Inter-Rater Reliability of the Modified Ashworth Scale (MAS) in
Spasticity in Patients with Stroke. Jpn. J. Rehabil. Med. 2002, 39, 409–415. [CrossRef]
14. Milner, T.E.; Franklin, D.W. Characterization of Multijoint Finger Stiffness: Dependence on Finger Posture and Force Direction.
IEEE Trans. Biomed. Eng. 1998, 45, 1363–1375. [CrossRef]
15. Kamper, D.G.; Zev Rymer, W. Quantitative Features of the Stretch Response of Extrinsic Finger Muscles in Hemiparetic Stroke.
Muscle Nerve 2000, 23, 954–961. [CrossRef]
16. Fiorilla, A.E.; Tsagarakis, N.G.; Nori, F.; Sandini, G. Design of a 2-Finger Hand Exoskeleton for Finger Stiffness Measurements.
Appl. Bionics. Biomech. 2009, 6, 217–228. [CrossRef]
17. Heung, H.L.; Tang, Z.Q.; Shi, X.Q.; Tong, K.Y.; Li, Z. Soft Rehabilitation Actuator With Integrated Post-Stroke Finger Spasticity
Evaluation. Front. Bioeng. Biotechnol. 2020, 8, 111. [CrossRef] [PubMed]
18. Shi, X.Q.; Heung, H.L.; Tang, Z.Q.; Tong, K.Y.; Li, Z. Verification of Finger Joint Stiffness Estimation Method With Soft Robotic
Actuator. Front. Bioeng. Biotechnol. 2020, 8, 592637. [CrossRef]
19. Tarvainen, T.V.J.; Yu, W. Preliminary Results on Multi-Pocket Pneumatic Elastomer Actuators for Human-Robot Interface in
Hand Rehabilitation. In Proceedings of the 2015 IEEE International Conference on Robotics and Biomimetics, IEEE-ROBIO 2015,
Zhuhai, China, 6–9 December 2015; pp. 2635–2639. [CrossRef]
20. Kokubu, S.; Wang, Y.; Tortós Vinocour, P.E.; Lu, Y.; Huang, S.; Nishimura, R.; Hsueh, Y.H.; Yu, W. Evaluation of Fiber-Reinforced
Modular Soft Actuators for Individualized Soft Rehabilitation Gloves. Actuators 2022, 11, 84. [CrossRef]
21. Yun, S.S.; Kang, B.B.; Cho, K.J. Exo-Glove PM: An Easily Customizable Modularized Pneumatic Assistive Glove. IEEE Robot
Autom. Lett. 2017, 2, 1725–1732. [CrossRef]
22. Zhou, C.Q.; Shi, X.Q.; Li, Z.; Tong, K.Y. 3D Printed Soft Robotic Hand Combining Post-Stroke Rehabilitation and Stiffness
Evaluation. In Lecture Notes in Computer Science (Including Subseries Lecture Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics); Springer International Publishing: Cham, Switzerland, 2022; Volume 13457 LNAI, pp. 13–23. [CrossRef]
23. Chong, T.W.; Lee, B.G. American Sign Language Recognition Using Leap Motion Controller with Machine Learning Approach.
Sensors 2018, 18, 3554. [CrossRef] [PubMed]
24. Elgeneidy, K.; Lohse, N.; Jackson, M. Bending Angle Prediction and Control of Soft Pneumatic Actuators with Embedded Flex
Sensors—A Data-Driven Approach. Mechatronics 2018, 50, 234–247. [CrossRef]
25. Kavian, M.; Nadian-Ghomsheh, A. Monitoring Wrist and Fingers Range of Motion Using Leap Motion Camera for Physical
Rehabilitation. In Proceedings of the 2020 International Conference on Machine Vision and Image Processing (MVIP), Qom, Iran,
18–20 February 2020; IEEE: Piscataway, NJ, USA. [CrossRef]
26. Ganguly, A.; Rashidi, G.; Mombaur, K. Comparison of the Performance of the Leap Motion Controllertm with a Standard
Marker-Based Motion Capture System. Sensors 2021, 21, 1750. [CrossRef] [PubMed]
27. Li, X.; Wan, K.; Wen, R.; Hu, Y. Development of Finger Motion Reconstruction System Based on Leap Motion Controller.
In Proceedings of the CIVEMSA 2018—2018 IEEE International Conference on Computational Intelligence and Virtual Environ-
ments for Measurement Systems and Applications, Proceedings, Ottawa, ON, Canada, 12–13 June 2018. [CrossRef]
28. Pagoli, A.; Chapelle, F.; Corrales-Ramon, J.A.; Mezouar, Y.; Lapusta, Y. Review of Soft Fluidic Actuators: Classification and
Materials Modeling Analysis. Smart Mater. Struct. 2022, 31, 013001. [CrossRef]
29. Marechal, L.; Balland, P.; Lindenroth, L.; Petrou, F.; Kontovounisios, C.; Bello, F. Toward a Common Framework and Database of
Materials for Soft Robotics. Soft Robot. 2021, 8, 284–297. [CrossRef]
30. Suzuki, T.; Kanno, I.; Loverich, J.J.; Kotera, H.; Wasa, K. Characterization of Pb(Zr,Ti)O3 Thin Films Deposited on Stainless Steel
Substrates by RF-Magnetron Sputtering for MEMS Applications. Sens. Actuators A Phys. 2006, 125, 382–386. [CrossRef]
31. Pititeeraphab, Y.; Choitkunnan, P.; Thongpance, N.; Kullathum, K.; Pintavirooj, C. Robot-Arm Control System Using LEAP
Motion Controller. In Proceedings of the BME-HUST 2016—3rd International Conference on Biomedical Engineering, Hanoi,
Vietnam, 5–6 October 2016; pp. 109–112. [CrossRef]
32. Leap Motion Controller Datasheet. Available online: https://www.ultraleap.com/datasheets/Leap_Motion_Controller_
Datasheet.pdf (accessed on 10 April 2023).
33. Guna, J.; Jakus, G.; Pogačnik, M.; Tomažič, S.; Sodnik, J. An Analysis of the Precision and Reliability of the Leap Motion Sensor
and Its Suitability for Static and Dynamic Tracking. Sensors 2014, 14, 3702–3720. [CrossRef]
34. Okuno, T.; Kume, H.; Haga, T.; Yoshizawa, T. Tahenryô Kaisekihô (Kaiteiban) [Multivatiate Analysis Methods (Revised Edition)], Revised
Edition; JUSE Press, Ltd.: Tokyo, Japan, 2005.
35. Dionysian, E.; Kabo, J.M.; Dorey, F.J.; Meals, R.A. Proximal Interphalangeal Joint Stiffness: Measurement and Analysis. J. Hand
Surg. 2005, 30, 573–579. [CrossRef]
36. Bohannon, R.W.; Smith, M.B. Interrater Reliability of a Modified Ashworth Scale of Muscle Spasticity. Phys. Ther. 1987, 67,
206–207. [CrossRef]
37. Kamper, D.G.; Hornby, T.G.; Rymer, W.Z. Extrinsic Flexor Muscles Generate Concurrent Flexion of All Three Finger Joints.
J. Biomech. 2002, 35, 1581–1589. [CrossRef]
Robotics 2023, 12, 83 25 of 25

38. Kamper, D.G.; Fischer, H.C.; Cruz, E.G.; Rymer, W.Z. Weakness Is the Primary Contributor to Finger Impairment in Chronic
Stroke. Arch. Phys. Med. Rehabil. 2006, 87, 1262–1269. [CrossRef] [PubMed]
39. Connolly, F.; Walsh, C.J.; Bertoldi, K. Automatic Design of Fiber-Reinforced Soft Actuators for Trajectory Matching. Proc. Natl.
Acad. Sci. USA 2017, 114, 51–56. [CrossRef] [PubMed]
40. Mirkhalaf, S.M.; Fagerström, M. The Mechanical Behavior of Polylactic Acid (PLA) Films: Fabrication, Experiments and Modelling.
Mech. Time Depend Mater. 2021, 25, 119–131. [CrossRef]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.