Sensors and Actuators A 285 (2019) 519–530

Contents lists available at ScienceDirect

Sensors and Actuators A: Physical

journal homepage: www.elsevier.com/locate/sna

Continuum robot shape estimation using permanent magnets and

magnetic sensors
Hao Guo, Feng Ju, Yanfei Cao, Fei Qi, Dongming Bai, Yaoyao Wang, Bai Chen ∗
School of Mechanical and Electrical Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing, Jiangsu Province, China

a r t i c l e i n f o a b s t r a c t

Article history: Continuum robots have attracted more attention owing to their safety and flexibility for confined and
Received 4 August 2018 complex environment. In order to control the flexible continuum robot more accurately, it is necessary to
Received in revised form 19 October 2018 track the shape of the continuum robot itself in real time. The feedback of the control system is precisely
Accepted 19 November 2018
controlled to ensure the safety of the motion of the robot in the unknown environment. This paper
Available online 19 November 2018
proposes to integrate small permanent magnets and magnetic sensors into continuum robot to estimate
the shape change. The sensors detect the posture change of magnets, the computer collects data through
Keywords:
serial ports. By calculating the relative bending angle of magnets, the data can be imported into Matlab
Continuum robot
Shape estimation
to fit the Quadratic Bézier curve to reconstruct the shape in 2D plane. Considering the interference of
Permanent magnet multiple magnets, the simulation software Maxwell is used to help arrange the magnets and magnetic
Magnetic sensor sensors reasonably to decrease the significant interference in the sensed magnetic flux density. The
Shape reconstruction experimental results show that the method is applicable to bending angles of 0◦ –90◦ . In this range, the
maximum and minimum error of the terminal position is 9.339 mm and 0.227 mm respectively, which
verifies certain accuracy and feasibility of this method.
© 2018 Elsevier B.V. All rights reserved.

1. Introduction [3]. This method acquires the image of a robot through a high-
speed camera, and then uses the image processing technology to get
Continuum robots have attracted more attention owing to its the shape of the robot. However, the camera acquires low quality
safety and flexibility for confined and complex environment [1]. images in low-light environment, and the image processing takes
Compared to the traditional prosthetic robots, the continuum a long time and low resolution. Binocular vision [4,5], trinocular
robots can be better adapted to the narrow, unknown working vision [6] and magnetic resonance imaging (MRI) are also applied
environment by changing its shape, such as industrial inspecting, in continuum robot shape sensing [7]. The intelligent materials are
post-earthquake relief and minimally invasive surgery applica- used in [8]. The main kind of materials are polyvinylidene fluoride
tions. However, due to its flexibility, the bending shape and (PVDF) and ion exchange polymerized metal material (IPMC). There
terminal position for continuum robots cannot be estimated. In are also some other materials used in this filed, such as stretch sen-
order to control the flexible continuum robot more accurately, it sor [9] and rubber diffraction grating sheet [10]. In the way of [11],
is necessary to track the shape of the continuum robot itself in real the corresponding strain can be deduced by detecting the voltage
time. The feedback of the control system is precisely controlled transformation when the flexible beam is deformed, so as to real-
to ensure the safety of the motion of the robot in the unknown ize the 2-D shape reconstruction. The approach of [12] uses IPMC
environment. to build the experimental system. IPMC is combined with a can-
The technology of continuum shape sensing has been mainly tilever beam to calculate the bending change of the beam through
focused on medical image technology, such as X-ray, CT, MRI. Intel- the host computer. IPMC has high sensitivity, the output voltage is
ligent materials, electromagnetic tracker and FBGs have also been at milli voltage level and the deformation range of the material is
used in recent years. As explained in [2], monocular vision is used large. Both approaches can be integrated in the continuum robots,
for shape sensing. Further research is in progress by Chitrakaran but it will lead to the complex configuration of the robot. It also
has not yet been focused on the shape sensing of continuum robots
in the 3-D space. Rone and Ben-Tzvi [13] propose a shape sensing
method using passive cable displacement embedded in the robot
∗ Corresponding author. itself to achieve closed-loop control, but due to the friction between
E-mail address: [email protected] (B. Chen).

https://doi.org/10.1016/j.sna.2018.11.030
0924-4247/© 2018 Elsevier B.V. All rights reserved.
520 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530

the line and the guide hole, the measurement is easy to be affected the bottom of each section and the other is installed in the middle
by the external force. Ozel [14] presents a systematic approach to position of each section. The magnetic sensor have been placed at
implement the feedback control of FFA-based soft robotic bending certain distance from the magnet. The magnetic field of the magnet
modules by using commercial flex bend sensors, but it only senses can be detected by the sensor array when it moves along with the
2-D shape at this stage and has poor performances at high dynam- motion of the continuum robot (Fig. 1(c)). Each magnetic sensor can
ics and drift effect [15,16]. Davis M.A. uses the wavelength division detect the relative bending angle of its corresponding magnet. The
multiplexing technology (WMD) of fiber Bragg grating to detect magnets must be installed on the connecting disk which will bend
the deformation of the cantilever beam [17]. Since then, Roesthuis along with the NiTi core column. But the need to pay attention to
integrates the fiber prague grating (FBG) sensor into the continuum is that this method is used under the condition that the continuum
robots [18,19]. The strain of the robot’s bending deformation is cal- robot is not affected by external forces, the bending angles of each
culated by measuring the change of the wavelength of reflected NiTi core column are the same. This method will be used to make
light. The curvature of the central skeleton of the continuum robot a preliminary study of 2D position of the continuum robot shape
is derived by the measured strain. This method has advantages of sensing.
high accuracy and high real-time but the measuring range is rel- Define the function of the relative bending angle from the tri-
atively small. It can only be used for small curvature deformation axial magnetic sensors as follows
of continuum robots. Shuang Song team proposes the use of elec-
tromagnetic tracker [20] and magnetic sensor array [21] to track i = /2-arctan2(Bxi , Byi ) (1)
the continuum robot shape, then reconstructs the shape by fitting where i is the i-th relative bending angle calculated by this method
the quadratic Bézier curves [20] and cubic Bézier curve [22]. This and Bi = (Bxi , Byi , Bzi )T is the measured magnetic field from the i-th
method has contact-free sensing capability and the high working sensor. By using the relative bending angle, the position of magnetic
frequency, ensuring high accuracy in a large range of dynamic con- sensors, magnets and the terminal position of the single section will
ditions [15]. However, this scheme does not integrate sensors into be get, which can be calculated as follows
continuum robot and needs external detection systems. ⎧
According to the result of previous studies, Hall-effect sen- ⎪ xs1 = 0, ys1 = 0;
⎪
⎪
sors are compact and inexpensive, they have been used for ⎪
⎪
⎪
⎪ xm1 =
l1
−
l1
cos(1 ), ym1 =
l1
anthropomorphic robotic hand [23]. Due to the magnetic sen- ⎪
⎪   
sin(1 );
sors’ contact-free sensing capability and high accuracy, the paper ⎪
⎪ 1 1 1
⎪
⎪ s1 s1
proposes to integrate small permanent magnet and magnetic sen- ⎪
⎪ xs2 = − cos(1 + ˇ1 ),
⎪
⎪ ˇ1 ˇ1
sor (HMC5883L) into continuum robot to measure the shape. ⎪
⎪
⎪
⎪ s
Compared with other magnetic-sensor-based methods, no other ⎪ ys2 = 1 sin(1 + ˇ1 );
⎪
external detection platforms are arranged outside the contin- ⎨ ˇ1
uum robot and the degree of integration is improved. Considering − (2)
⎪
⎪ xm2 = xs2 + la2 cos(
2
− 1 − ˇ1 ),
the interference of multiple magnets, the magnets and magnetic ⎪
⎪ 2
⎪
⎪
sensors are arranged reasonably to decrease the significant inter- ⎪
⎪  − 2
ference in the sensed magnetic flux density. ⎪ ym2 = ys2 + la2 sin(
⎪ − 1 − ˇ1 );
⎪
⎪ 2
The paper is organized as follows: in section II, firstly, the struc- ⎪
⎪
⎪
⎪
ture of the continuum robot (Fig. 1) and the principle are described. ⎪ xt = xm2 + sa2 cos(  − ˇ2 − 1 − ˇ1 − 2 ),
⎪
⎪
⎪ 2
Secondly, in order to determine where the magnets and sensors ⎪
⎪
are installed, the magnetic flux densities of single magnet and two ⎪
⎩ y = y + s sin(  − ˇ2
t m2 a2 −  − ˇ −  );1 1 2
magnets are simulated and analyzed, including the effect of the 2
distance between magnets and sensors. The error caused by the
where li is the absolute distance between the i-th magnetic sensor
superposition principle of magnetic field is also analyzed. The above
and corresponding magnet, lai is the relative distance between the
work helps to decrease the significant interference in the sensed
i-th magnetic sensor and corresponding magnet. s1 is the absolute
magnetic flux density. At last, the shape estimation method will
distance from the Magnet-1 to Sensor-2, sa1 is the relative distance
be explained. In section III, the test platform is introduced to do
from the Magnet-1 to Sensor-2; s2 is the absolute distance from the
experiment and reconstruct the continuum robot shape. In view
Magnet-2 to the terminal position, sa2 is the relative distance from
of the interference of geomagnetic field, we propose correspond-
the Magnet-2 to the terminal position. (xsi , ysi ) is the i-th sensor
ing compensation methods to improve the experimental accuracy
position, (xmi , ymi ) is the i-th magnet position. The terminal position
Conclusion and future work will be drawn in Section IV.
of the single section is described as (xt , yt ) (Fig. 2). The relative
distance can be represented with the following formula
⎧   
2. Methods ⎪
⎪ li
2
⎪
⎨ lai = 2 1 − cos(i )
i
2.1. Structure and principle    (3)
⎪
⎪
⎪
⎩ sai = si 2
2 1 − cos(ˇi )
In general, the traditional continuum robots have twisting prob- ˇi
lems in their central located flexible backbone. To minimize this
twisting angle [24,25], propose a novel structure of the continuum ˇi represents the other two bending angles which cannot be directly
robot. This paper proposes to integrate the permanent cylindri- calculated by the magnets. The value is related to the number of the
cal magnet and magnetic sensor into this structure. As shown in disks and i , ˇi can be re represented with the following formula
Fig. 1(b), the single section is made up of 20 connecting disks. Each
s1 s2
connecting disk is a circle with a diameter of 30 mm and a thick- ˇ1 = 1 , ˇ2 = 2 ; (4)
l1 l2
ness of 5 mm, which is connected by two 5-mm-long NiTi core
columns. The length of the single section is 195 mm. The single These five positions help to realize the 2D shape estimation in
section includes two sets of detection systems (each system con- the following part. The method of the shape estimation is to fit a
sists of a magnetic sensor and a small magnet). One is installed at curve to reconstruct the single section shape in the 2D plane.
 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530 521

Fig. 1. (a) Concept design of the continuum robot. (b) Overall sketch of the 2-DOF (rotating and bending) single section (include two cylindrical permanent magnets and two
three-axis magnetic sensors). The single section is made up of 20 connecting disks. Each connecting disk is a circle with a diameter of 30 mm and a thickness of 5 mm, which
is connected by two 5-mm-long NiTi core columns. The length of the single section is 195 mm. (c) Sketch for the single section from resting state to bending state. The figure
indicates that the method used for shape estimation by measuring the relative position between each magnet and sensor.

Fig. 4. The method of calculating control points of Quadratic Bézier curve (a) The
position of four control points, (b) The relationship between the length of tangent
line in each arc.

Fig. 2. Kinematic model for the single section. i is the i-th relative change calculated
by the method, ˇi represents the i-th angle which is related to the number of the
disks and i . the single section of the continuum robot is divided into four parts,
each part corresponds to a section of curve.
quadratic Bézier curve. The curve starts at start point going toward
control point and arrives at end point.
The Quadratic Bézier curve need to be fitted for four times (each
part corresponds to a section of curve, which can be seen in Fig. 2),
so This method needs four control points, which are P1.1 , P1.2 , P1.3 ,
P1.4 (Fig. 4). The control point is the intersection point of the tangent
lines of the starting point and the end point for each part. Their
coordinates can be calculated by the following formula

⎧
⎪
⎪ xP1.1 = xm1 − lb1 sin(˛1 ), yP1.1 = ym1 − lb1 cos(˛1 );
⎪
⎨
Fig. 3. Quadratic Bézier curve. P0 is the start point, and P2 is the end point, P1 is the xP1.2 = xs2 − lb2 sin(˛2 ), yP1.2 = ys2 − lb2 cos(˛2 );
(6)
control point. The curve starts at start point(P0 ) going toward control point(P1 ) and ⎪
⎪
arrives at end point(P2 ). ⎪
⎩ xP1.3 = xm2 − lb3 sin(˛3 ), yP1.3 = ym2 − lb3 cos(˛3 );
xP1.4 = xt − lb4 sin(˛4 ), yP1.4 = yt − lb4 cos(˛4 );

A quadratic Bézier curve is shown in Fig. 3. P0 is the start point,

and P2 is the end point; P1 is the control point. The explicit form of where (xP1.i , yP1.i ) is the i-th control point position (Fig. 4). ˛i is the
the curve is included angle between the tangent line of the i-th end point and
the vertical direction. lbi is the length of the tangent line between
B(t) = (1 − t)2 P0 + 2t(1 − t)P1 + t 2 P2 (5) the i-th control point and end point. We assume that the single
section of the continuum robot is not affected by external forces
therefore, the curve fitting can be used based on (3). A single section and the bending angles of each NiTi core column are the same.
can be divided into four part (Fig. 2), and each part needs to fit the Under the above condition, each arc is equal curvature, the length
522 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530

Fig. 5. The size of the magnet model used in the simulation. The diameter is 4 mm
and the height is 3 mm. The direction of magnetization is axial magnetization. The
material is NdFe35.

of the two tangents of each part is equal, which can be represented

with the following formula
⎧
⎪ l1 1 l2 2
⎨ lb1 =  tan 2 , lb3 =  tan 2 ;
1 2
(7)
⎪
⎩ s1 ˇ1 s2 ˇ2
lb2 = tan , lb4 = tan ;
ˇ1 2 ˇ2 2
According to the above formulas, the curve can be fitted in MATLAB
through the control points and the corresponding start points and
the end points.
This part explains the basic structure and principles. Two sets
of detection systems are integrated into a single section of the
continuum robot, each detection system includes a magnet and a
magnetic sensor. We assume that the single section of the con-
tinuum robot is not affected by external forces and the bending
angles of each NiTi core column are the same. By detecting the rel-
ative bending angle of the magnet in each detection system, the
absolute positions of the two magnets in the coordinate system are
estimated. There are five absolute position points and four abso-
lute control points in all. These five absolute position points divide
the single section into four parts, each part needs to fit a Quadratic
Fig. 6. (a)The simulation method of single sensor and single magnet (b)The simula-
Bézier curve to reconstruct its shape in the 2D plane. The simulation tion of the single magnet. The magnetic flux density of the magnet is resolved to three
of the reasonable arrangement and simulation of shape estimation components along the coordinate axis, any magnet flux density of the components
will appear in the following part. can be obtained at any distance from the center of the magnet. (c)The relation-
ship between Y-axis magnetic flux density and the distance from the center of the
magnet.
2.2. Simulation of reasonable arrangement

Due to the mutual interference of multiple magnets, in this

chapter, the simulation software Maxwell is used to analyze affec- sensors and magnets are installed along Y axis in the structure we
tion of the distance between single magnet and single sensor. Then, mentioned above, Y-axis magnetic flux density is chosen as our cri-
the effect of interaction between two magnets on the sensors will teria for where the magnetic sensor should be installed away from
also be analyzed. This work helps arrange the magnets and mag- the magnets to work properly. Then, we design three installation
netic sensors reasonably to decrease the significant interference schemes for sensors and magnets based on the continuum robot
in the sensed magnetic flux density, and verify that the change of structure we have and analyze the influence of mutual interference
magnetic flux density follows some regularity, when the posture of of magnets in these schemes on the magnetic flux densities col-
the magnets changes. lected by sensors. Finally, we analyze the error in three schemes to
Fig. 5 shows the size of the magnet model used in the simulation. use superposition principle of magnetic field to determine scheme.
The magnets with 4 mm diameters are used as research targets. Fig. 6(a) and (b) shows the simulation method of single Sensor-
The height of this magnet is 3 mm. The direction of magnetization 1 and single Magnet-1. By using simulation software Maxwell, The
is axial magnetization. The cylindrical magnet in this method is magnetic flux density of the magnet is resolved to three compo-
made of NdFe35. Hc = 890,000 A/m and Br = 12,300 G. By importing nents along the coordinate axis, any magnet flux density of the
the above material parameters into Maxwell, the magnetic field components can be obtained at any distance from the center of
simulation of permanent magnet can be get. Since this method is the magnet. The relationship between Y-axis magnetic flux density
only used to get the 2D shape of the continuum in X–Y plane, Bx and distance from the center of the magnet is obtained, as shown in
and By are magnet flux density vector components in different (x, Fig. 6(c). A preliminary conclusion can be made from Fig. 6: The dis-
y) positions with respect to the magnet. tance between the magnetic sensor and the magnet must be over
Firstly, we use finite element analysis (FEA) to obtain the mag- 10 mm. Since the magnetic sensor used in this method (HMC5883L)
netic flux density vector data. We will analyze single magnet for its can achieve 0.005 G resolution in ± 8 G fields, the Y-axis magnetic
magnetic flux density at any distance from the center of it. Since flux density of the magnet within the range of ± 10 mm is too high.
 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530 523

4 × 3 magnet, the sensor must be installed more than 20 mm away

from the center of this magnet.
Based on the above simulation and the existing single section
of the continuum robot structure, three schemes of sensors and
magnets installation are designed, which can be seen in Fig. 8. The
magnets need to be installed on the connecting disks which will
bend along with the NiTi center-column. The magnets and sensors
in the two detection systems adopt the same installation scheme,
so the distance between the Magnet-1 and the Magnet-2 is 100 mm.
The connecting disks that can be used for installing Magnet-1 are
the forth, the sixth and the eighth one. The distance between the
center of each disk is 10 mm. In Scheme-1, the forth disk is chosen
to install the Magnet-1. The distance between the Sensor-1 and the
center of the Magnet-1 is 30 mm, the distance between the Sensor-
2 and the center of the Magnet-2 is also 30 mm. In Scheme-2, the
forth sixth is chosen to install the Magnet-1. The distance between
the Sensor-1 and the center of the Magnet-1 is 50 mm, the distance
between the Sensor-2 and the center of the Magnet-2 is also 50 mm.
In Scheme-3, the eighth disk is chosen to install the Magnet-1. The
distance between the Sensor-1 and the center of the Magnet-1 is
70 mm, the distance between the Sensor-2 and the center of the
Magnet-2 is also 70 mm.
Because the detection method uses two magnets, it is necessary
to consider the effect of the interaction between the two magnets
on the data collected by the sensor. Fig. 9(a) and (b) shows the sim-
Fig. 7. The Y-axis magnetic flux density of the magnet outside the range of ± 10 mm ulation method of single Sensor-1 and two magnets. Firstly, the
away from the center of the magnet. The origin location (0,0) is the position of magnetic flux density of single Magnet-1 is collected at 30 mm
magnet. The red dotted line is the reference line of 8guss (For interpretation of the
references to colour in this figure legend, the reader is referred to the web version
on the left side of center of the Magnet-1 along Y axis. Then, the
of this article). Magnet-2 is placed at 100 mm on the right side of the center of
the Magnet-1 along Y axis. The Magnet-2 is chosen as the interfer-
ence term. The magnetic flux density of two magnets is collected
again at 30 mm on the left side of center of the Magnet-1. Compare
In order to determine more accurate installation position of the the magnetic flux density of two magnets with the magnetic flux
magnet, the magnetic flux density of the magnet outside the range density of single magnet. The error can be seen in Table 1. When
of ± 10 mm needs to be conducted further study. The result can be the distance between the Sensor-1 and the Magnet-1 is 30 mm,
seen in Fig. 7. The origin location (0, 0) is the position of magnet. The the error is −0.0087 G. When the distance between the Sensor-1
red dotted line is the reference line of 8guss (in Fig. 7). Fig. 7 shows and the Magnet-1 is 50 mm, the error is −0.0007 G. When the dis-
the simulation of the 4 × 3 magnet outside the range of ± 10 mm tance between the Sensor-1 and the Magnet-1 is 70 mm, the error is
away from the center of the magnet. The point of intersection of −0.0002 G. The error caused by interaction of two magnets accounts
the red dotted line and the curve of the magnetic flux densities for 0.26%, 0.10% and 0.09% of the magnetic flux density of single
shows the reasonable position of the sensor installation. For the Magnet-1 respectively. From the above result, when the distance

Fig. 8. Three schemes of sensors and magnets installation. The distance between the Sensor-1 and the center of the Magnet-1 is 30 mm, 50 mm and 70 mm respectively. The
magnets and sensors in the two detection systems adopt the same installation scheme. The distance between the Magnet-1 and the Magnet-2 is 100 mm.
524 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530

Fig. 10. (a) The simulation method of single Sensor-2 and two magnets (b)
The distance between the Magnet-1 and the Magnet-2 is 100 mm. Considering
the mutual interference between the two magnets, we use simulation software
to analyze two magnets to collect magnetic flux density of two magnets at
30 mm/50 mm/70 mm(respectively) away from the center of Magnet-1, and com-
pare these dates with the magnetic flux density of the single Magnet-1 here.

Fig. 9. (a) The simulation method of single Sensor-1 and two magnets (b) The dis-
the Sensor-2 and Magnet-2 is 50 mm, the error is −0.5532 G. When
tance between the Magnet-1 and the Magnet-2 is 100 mm. The simulation software
is used to analyze the mutual interference between the two magnets to collect mag- the distance between the Sensor-2 and the Magnet-2 is 70 mm,
netic flux density of Magnet-1 and Magnet-2 at 30 mm/50 mm /70 mm (respectively) the error is −2.0458 G. From the above result, the error in Scheme-
away from the center of Magnet-1, and compare these dates with the magnetic flux 1 is the smallest. But in the latter two schemes, the errors are too
density of the single Magnet-1 here. large. In Scheme-3, the Magnet-1 is close to Sensor-2, so the Sensor-
2 mainly collects the magnetic flux density of the Magnet-1. In
Table 1 Scheme-2, the magnetic flux density of two magnets is 1.1280 G,
Error caused by interaction of two magnets (By1).
which is almost double the data (0.5748 G) of single Magnet-2. The
L = 30 mm/G L = 50 mm/G L = 70 mm/G Sensor-2 is in the middle position between two magnets, so the
Single Magnet 3.3976 0.6762 0.2234 Sensor-2 is affected by the superposition principle of magnetic field
Two Magnets 3.4063 0.6769 0.2236 of two magnets. However, the superposition principle of magnetic
Error −0.0087 −0.0007 −0.0002 field not only interferes with the Y-axis magnetic flux density, but
Proportion 0.26% 0.10% 0.09% also interferes with the X-axis magnetic flux density.
The influence of the superposition principle of magnetic field
Table 2 needs to be analyzed together with the bending process of the sin-
Error caused by interaction of two magnets (By2). gle section of the continuum robot. UG software is used to build
simple 3D models of the bending single section, the magnets will
L = 30 mm/G L = 50 mm/G L = 70 mm/G
change posture along with the bending of the continuum robot.
Single Magnet 3.3967 0.5748 0.2623
Then, the models are imported into Maxwell software for analysis.
Two Magnets 3.4592 1.1280 2.3081
Error −0.0625 −0.5532 −2.0458 The details are shown in Fig. 11, There are seven kinds of bend-
ing angles: 0◦ , 15◦ , 30◦ , 45◦ , 60◦ , 75◦ , 90◦ . Red cylinders represent
magnets. Blue cylinders represent Sensors. The light blue curves
between the two magnets is 100 mm, the interference between the represent the centerline of the single section of the continuum
magnets has little effect on the Sensor-1.
Since the Sensor-1 is on the left side of the Magnet-1, the
Magnet-2 has little effect on Sensor-1. However, the Sensor-2 is
between the Magnet-1 and the Magnet-2, and it will be affected by
both magnets at the same time. In order to understand the effects
of this condition, the simulation of single Sensor-2 and two mag-
nets is carried out. Fig. 10 show the simulation method of single
Sensor-2 and two magnets. The Magnet-1 is chosen as the inter-
ference term. The Magnet-1 is place at 100 mm on the left of the
Magnet-2 along Y axis in each scheme. In Scheme-1, the Sensor-2
is placed at 30 mm on the left side of the center of the Magnet-2
along Y axis. In Scheme-2, the Sensor-2 is placed at 50 mm on the
left side of the center of the Magnet-2 along Y axis. In Scheme-3,
the Sensor-2 is placed at 70 mm on the left side of the center of
Fig. 11. Simple 3D models of seven kinds of bending angles for the single section(0◦ ,
the Magnet-2 along Y axis. The analyzing method is the same as
15◦ , 30◦ , 45◦ , 60◦ , 75◦ , 90◦ ). Red cylinders represent magnets. Blue cylinders repre-
above. Compare the magnetic flux density of two magnets with the sent Sensor. The light blue curves represent the centerline of the single section of
magnetic flux density of single Magnet-2. The error can be seen in the continuum robot(l1=l2 = 30 mm/50 mm/70 mm). (For interpretation of the ref-
Table 2. When the distance between the Sensor-2 and the Magnet- erences to colour in this figure legend, the reader is referred to the web version of
2 is 30 mm, the error is −0.0625 G. When the distance between this article).
 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530 525

Fig. 12. The simulation method of two sensors and two magnets. The Y-axis and X-axis magnetic flux densities of Sensor-1 and Sensor 2 will be extracted under the states
of seven bending angles.

robot. Three schemes are all built simple 3D models. Simulation and Sensor 2 will be extracted under the states of seven bending
can help to find how the X-axis and Y-axis magnetic flux density angles. The relations between magnetic flux densities and bending
will change and how the superposition principle of magnetic field angles can be seen in Fig. 13. Fig. 13(a) shows the relation between
affect the data collected by the Sensors when the magnets posture the magnetic flux densities of two magnets and bending angles
changes. in Scheme-1. Fig. 13(b) shows the relation between the magnetic
Fig. 12 shows the simulation method of two sensors and two flux densities of two magnets and bending angles in Scheme-2.
magnets. The X-axis and Y-axis magnetic flux density of Sensor-1 Fig. 13(c) shows the relation between the magnetic flux densities of

Fig. 13. (a)The relation between the magnetic flux densities of two magnets and the bending angles of the single section in Scheme-1. (b)The relation between the magnetic
flux densities of two magnets and the bending angles of the single section in Scheme-2. (c)The relation between the magnetic flux densities of two magnets and the bending
angles of the single section in Scheme-3. The Blue represents the magnetic flux density of the Magnet-1 collected by the Sensor-1. The Red represents the magnetic flux
density of the Magnet-2 collected by the Sensor-2. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article).
526 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530

Fig. 15. The final structure. Each magnet and sensor is arranged at an equal dis-
tance. The distance between Magnet-1 and Sensor-1 is 50 mm, the distance between
Fig. 14. The superposition principle of magnetic field. B1 and B2 must be the mag- Magnet-1 and Magnet-2 is 100 mm.
netic flux density produced at the point of the Sensor-2. B1 is not the value produced
at the point of the Sensor-1.
Sensor-2. B1 is not the value produced at the point of the Sensor-1.
Through the above formula, we can basically remove the effect of
two magnets and bending angles in Scheme-3. The histogram can mutual interference of two magnets on Sensor-2 to obtain the true
more directly reflect the influence of the superposition principle of magnetic flux densities of the Magnet-2 collected by the Sensor-2.
magnetic field on the X-axis and Y-axis magnetic flux densities. The In Scheme-1, the distance between the Magnet-1 and Senso-1
Blue represents the magnetic flux density of the Magnet-1 collected is 30 mm and the distance between the Magnet-1 and Sensor-2 is
by the Sensor-1. The Red represents the magnetic flux density of 70 mm. According the above analysis, we have known that the mag-
the Magnet-2 collected by the Sensor-2. From the above diagram, netic flux density at 30 mm away from the center of the Magnet-1
we can see that the X-axis and Y-axis magnetic flux densities of the is different from the magnetic flux density at 70 mm away from
three schemes are greatly influenced by the superposition principle the center of the Magnet-1.If we use Scheme-1, we will not acquire
of magnetic field. In Scheme-3, Since the Magnet-1 is close to the the magnetic flux density of the Magnet-1 produced at the point of
Sensor-2, The Red data is the magnetic flux density of the Magnet-1 the Sensor-2. So Scheme-1 is unacceptable. In Scheme-2, the dis-
collected by the Sensor-2. The Scheme-3 is the most disturbing and tance between the Magnet-1 and Senso-1 is 50 mm and the distance
undesirable. between the Magnet-1 and Sensor-2 is 50 mm. The magnetic flux
As shown in Fig. 14, this is the superposition principle of mag- density of the Magnet-1 produced at the point of the Sensor-1 is
netic field in the above scheme. The magnetic flux density collected almost equal to the magnetic flux density of the Magnet-1 pro-
by Sensor-2 can be expressed in the following formula: duced at the point of the Sensor-2. The final structure uses the
Scheme-2, which can be seen in Fig. 15. The relation between the
Bx = Bx1 cos(1 + ˇ1 ) + Bx2 − By1 sin(1 + ˇ1 )
(8) true magnetic flux densities of Magent-1 collected by the Sensor-1
By = By1 cos(1 + ˇ1 ) + By2 and bending angle can be seen in Fig. 16(a), the relation between the
true magnetic flux densities of Magent-2 collected by the Sensor-2
where Bx1 and By1 are the components of B1 , Bx2 and By2 are the and bending angle can be seen in Fig. 16(b). The diagram shows
components of B2 .What we need to pay attention to is that B1 and that the change of magnetic flux density follows some regularity,
B2 must be the magnetic flux density produced at the point of the when the posture of the magnets changes.

Fig. 16. (a)The relation between the true magnetic flux densities of Magent-1 collected by the Sensor-1 and the bending angles of the single section (b)The relation between
the true magnetic flux densities of Magent-2 collected by the Sensor-2 and the bending angles of the single section.
 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530 527

Fig. 18. The error between simulation result of curves with the single section and
theoretical result of curves with the single section. The red curves are fitted by the
theoretical data which is measured by the 3D model. The blue curves are fitted by
the simulating data which is collected in simulation software (For interpretation
Fig. 17. The comparison between theoretical value and simulating value, (a) The of the references to colour in this figure legend, the reader is referred to the web
bending angle 1 of Magnet-1 relative to the Sensor-1 (b) The bending angle 2 of version of this article).
Magnet-2 relative to the Sensor-2. Theoretical value is the relative bending angle
measured by the 3D model. Simulating value is the relative bending angle calculated
by the simulating data collected by sensors in simulation software. controlled at 2.560 mm, which is an acceptable range of errors. The
simulation verifies the accuracy of the method in this range.
This part has shown the influence of the self-characteristics of This part has shown the comparison between simulating rel-
the magnet. The simulation software is used to help to reasonably ative angle and theoretical relative angle, and the error between
arrange the distance between the magnets and the sensors. The simulation curve shape and theoretical curve shape. The results
superposition principle of magnetic field and the mutual interfer- show that the errors are within acceptable range and demonstrate
ence between magnets are the main problems when we determine the accuracy of the method in the ideal range. However, simulation
the final structure. It is verified that the change of magnetic flux is carried out without interference from other factors, the effect in
density follows some regularity, when the posture of the magnets the real environment will be verified in the following experiments.
changes. In the next part, the simulation data will be imported into
Matlab to estimate the shape in the following part. 3. Results and discussion

2.3. Simulation of shape estimation 3.1. Test platform

According to the Formula (2) mentioned above, the simulating The test platform include 2 parts: the single section prototype
values are substituted into the Formula (2) to get the simulating and the contrastive experiment platform. As shown in Fig. 19(a),
angles (1 and 2 ) of the magnets relative to the sensors. Theoret- a contrastive experiment platform is designed to calibrate seven
ical value is the bending angle measured in the three-dimensional bending angles, which are 0◦ , 15◦ , 30◦ , 45◦ , 60◦ , 75◦ and 90◦ . The
model. Simulating value is the bending angle calculated by the data red lines are the reference lines. The circular holes in the picture
obtained by simulation. The results are shown in the Fig. 17. The- are used to install the fixed bars to determine the single section
oretical value is the relative bending angle measured by the 3D bending angle. This method is more clear and intuitive in com-
model. Simulating value is the relative bending angle calculated by parison experiments, and it doesn’t need to use other detection
the simulating data collected by sensors in simulation software. The equipment to judge the bending angle of the continuum robot. As
maximum error of 1 in simulation is 1.1939◦ , the minimum error shown in Fig. 19(b), this is our single section prototype of the con-
of 1 in simulation is 0.1039◦ ; the maximum error of 2 in simu- tinuum robot integrated with magnetic sensors and magnets. The
lation is 1.462◦ , the minimum error of 2 in simulation is 0.1684◦ . test platform we built are shown in Fig. 19(c). The sensor detects
These errors are within the acceptable range. Fig. 18 shows the error the posture change of magnets, the computer collects data through
between simulation result of curves with the single section and the- serial ports. By calculating the relative bending angle of magnets,
oretical result of curves with the single section. The red curves are the data can be imported into Matlab to fit the curve.
fitted by the theoretical data which is measured by the 3D model. This part has shown the test platform, which includes the sin-
The blue curves are fitted by the simulating data which is collected gle section prototype and the contrastive experiment platform. The
in simulation software. The error between curves is increasing specific experimental results and steps will show in the next part.
when the bending angle increases gradually. The maximum error
appear at a bending angle of 90◦ , the error of the terminal posi- 3.2. Experiment and result
tion is 2.696 mm, the minimum error appear at a bending angle
of 0◦ , the error of the terminal position is 0.077 mm. Because this In section II, we propose to use the superposition principle of
method is to use relative bending angles (1 and 2 ) to calculate magnetic field to modify the magnetic flux densities collected by
the fitting positions. With the increasing of the bending angle of Sensor-2. This is a compensation method which we will use in the
the single section, the errors of relative bending angles will lead experiment. Fig. 20 shows the process of our experiment, which
to the increasing of the deviation of fitting positions. These errors includes seven kinds of bending angles. During the experiment, no
are constantly amplified in the fitting curves. The average error is other metal products or electronic equipment are found within 2 m
528 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530

Fig. 19. (a)A contrastive experiment platform, (b)The single section prototype of
the continuum robot integrated with magnetic sensors and magnets, (c)The test
platform.

Fig. 20. The process of the experiment, which includes 7 kinds of bending angles.
of the test platform to reduce the interference from the surrounding During the experiment, no other metal products or electronic equipment are found
environment. The relative bending angles can be obtained by exper- within 2 m of the test platform to reduce the interference from the surrounding
environment.
iment, and compare them with theoretical value and simulating
value.
. In order to verify the effectiveness of the compensation
method, we firstly carry out experiments without compensation
method. The results are shown in Fig. 21. The relative angles calcu-
lated by the experiment have a great deviation from the theoretical
value and the simulating value in the range of 45◦ to 90◦ . This is due
to the influence of the surrounding environment on it, including
the effect of the geomagnetic field, the magnetic material and the
electrical conductors. The maximum error of 1 is 6.514◦ , the min-
imum error of 1 is 0.0725◦ ; the maximum error of 2 is 5.213◦ , the
minimum error of 2 is 0.0545◦ . The error is beyond the acceptable
range. The error in experiment is very large without compensation.
However, the error includes the effect caused by the interference
of electronic instruments and geomagnetic field. For the interfer-
ence of electronic instruments and geomagnetic field, we use single
HMC5883 L to collect the magnetic flux density of the surrounding
environment when we do experiment to modify the data collected
by sensors through the superposition principle of magnetic field.
The magnetic flux density of surrounding environment is about
0.0499 G. Then we carry out the next experiment with compen-
sation. The result can be seen in Fig. 22. By using the compensation
method, the error is obviously reduced. The experimental results
are not as perfect as those in the simulation environment. The con-
necting disk of the single section is made of resin material. The
connecting column is made of NiTi material. Neither of these two Fig. 21. The comparison among theoretical values and experimental values (with-
out compensation). (a) The bending angle 1 of Magnet-1 relative to the Sensor-1.
materials will be magnetized. Although HMC5883 L has certain
(b) The bending angle 2 of Magnet-2 relative to the Sensor-2. Theoretical value is
anti-geomagnetic interference ability, it cannot completely elim- the relative bending angle measured by the 3D model. Experimental value is the
inate the interference of electronic instruments and geomagnetic relative bending angle calculated by the experimental data collected by sensors.
field.
 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530 529

in the fitting curves. The experiment verifies certain accuracy of the

method in the range of 0–90◦ , especially in the range of 0–45◦ .
This part has shown the error between experimental relative
bending angle and theoretical relative bending angle, and the error
between experimental curve shape and theoretical curve shape.
The experimental results show that the method is applicable to
bending angles of 0◦ –90◦ . In this range, the maximum and min-
imum error of the terminal position is 9.339 mm and 0.227 mm
respectively, which verifies certain accuracy and feasibility of this
method.

4. Conclusion and future work

In this paper, a new method is proposed to estimate the con-

tinuum robot shape. Firstly, two sets of detection systems are
integrated into a single section of the continuum robot, each detec-
tion system includes a magnet and a magnetic sensor. We assume
that the single section has equal curvature when it bends. By detect-
ing the relative bending angle of the magnet in each detection
system, the absolute positions of the two magnets in the coordi-
nate system are estimated. There are five absolute position points
Fig. 22. The comparison among theoretical values and experimental values (with
and four absolute control points in all. These five absolute position
compensation). (a) The bending angle 1 of Magnet-1 relative to the Sensor-1. (b)
The bending angle 2 of Magnet-2 relative to the Sensor-2. Theoretical value is the points divide the single section into four parts, each part needs to
relative bending angle measured by the 3D model. Experimental value is the relative fit a Quadratic Bézier curve to reconstruct its shape in the 2D plane.
bending angle calculated by the experimental data collected by sensors. Secondly, considering the influence of the self-characteristics of
the magnets on the installation position of the magnets, the simu-
lation software Maxwell is used to help to determine the distance
between the magnets and the sensors. The superposition principle
of magnetic field and the mutual interference between magnets
are the main problems when we determine the final structure. It
is verified that the change of magnetic flux density follows some
regularity when the posture of the magnets changes. We analyze
the error between simulating relative bending angle and theo-
retical relative bending angle, and the error between simulation
curve shape and theoretical curve shape. The results show that the
errors are within acceptable range. However, simulation is carried
out without interference from other factors, the effect in the real
environment will be verified in the following experiments.
Finally, the test platform is built, which includes the single
section prototype and the contrastive experiment platform. The
superposition principle of magnetic field is used to modify the
magnetic flux densities collected by sensors, which can reduce the
interference of electronic instruments and geomagnetic field. The
experimental results show that the method is applicable to bend-
ing angles of 0◦ to 90◦ . In this range, the maximum and minimum
Fig. 23. The error between experimental result of curves with the single section and
error is 9.339 mm and 0.227 mm respectively, which verifies certain
theoretical result of curves with the single section. The red curves are fitted by the
theoretical data which is measured by the 3D model. The black curves are fitted by accuracy and feasibility of this method.
the experimental data which is collected in experiment (For interpretation of the The continuum robot shape estimation using magnets and mag-
references to colour in this figure legend, the reader is referred to the web version netic sensors will be made a further study to introduce the detection
of this article).
methods from 2D space to 3D space. We have used the superpo-
sition principle of magnetic field to compensate the interference
Fig. 23 shows the error between experimental result of curves term of electronic instruments and geomagnetic field. But this com-
with the single section and theoretical result of curves with the sin- pensation method is very simple, and cannot completely eliminate
gle section. With the increasing of the bending angle of the single the environmental impact. The more advanced compensation algo-
section, the error of the terminal position at 0◦ is 0.227 mm. The rithm will be studied in the future. By reducing the error and making
error of the terminal position at 15◦ is 1.207 mm, the error of the the shape estimation method more accurate, this method will be
terminal position at 30◦ is 2.081 mm, the error of the terminal posi- applied to multi section of the continuum robot.
tion at 45◦ is 4.582 mm, the error of the terminal position at 60◦ is
6.774 mm, the error of the terminal position at 75◦ is 7.733 mm, the
error of the terminal position at 90◦ is 9.339 mm. The reason for the Acknowledgements
increasing of the fitting curves’ error is as follows: This method uses
relative bending angles (1 and 2 ) to calculate the fitting positions. This work is supported by National Natural Science Founda-
With the increasing of the bending angle of the single section, the tion of China (No. 51575256), “the Fundamental Research Funds
errors of relative bending angles will lead to the increasing of the for the Central Universities, No. NS2018033”, and also funded by
deviation of fitting positions. These errors are constantly amplified Open Foundation of the State Key Laboratory of Fluid Power and
530 H. Guo et al. / Sensors and Actuators A 285 (2019) 519–530

Mechatronic Systems (GZKF-201711), and the National Key R&D [21] S. Song, C. Zhang, L. Liu, M. Meng, Preliminary study on magnetic
Program of China (2018YFC0309102). tracking-based planar shape sensing and navigation for flexible surgical
robots in transoral surgery: methods and phantom experiments, Int. J.
Comput. Assist. Radiol. Surg. 13 (2) (2018) 241–251.
References [22] S. Song, Z. Li, H. Ren, H. Yu, Shape reconstruction for wire-driven flexible
robots based on Bézier curve and electromagnetic positioning, Mechatronics
[1] Z. Li, R. Du, M.C. Lei, S.M. Yuan, Design and analysis of a biomimetic 29 (August) (2015) 28–35.
wire-driven robot arm, in: ASME 2011 International Mechanical Engineering [23] J. Butterfass, M. Grebenstein, H. Liu, G. Hirzinger, DLR-hand II: next
Congress and Exposition, 2011, pp. 191–198, American society of mechanical genera-tion of a dexterous robot hand, Proceedings 2001 ICRA. IEEE
engineers. InternationalConference on Robotics and Automation vol. 1 (2001) 109–114.
[2] M.W. Hannan, I.D. Walker, Real-time shape estimation for continuum robots [24] X. Dong, M. Raffles, S.C. Guzman, et al., A novel snake arm robot using
using vision, Robotica 23 (5) (2005) 645–651. twin-pivot compliant joints: design, modeling and validation, J. Mech. Robot.
[3] V.K. Chitrakaran, A. Behal, D.M. Dawson, et al., Setpoint regulation of 8 (2) (2016).
continuum robots using a fixed camera, in: American Control Conference, [25] X. Dong, D. Axinte, D. Palmer, et al., Development of a slender continuum
2004. Proceedings of the 2004, 2004, pp. 1504–1509. robotic system for on-wing inspection/repair of gas turbine engines, Robot.
[4] C.L. Chen, C.L. Tai, Y.F. Lio, Virtual binocular vision systems to solid model Comput. Integr. Manuf. 44 (C) (2017) 218–229.
reconstruction, Int. J. Adv. Manuf. Technol. 35 (3–4) (2007) 379–384.
[5] J.M. Croom, D.C. Rucker, J.M. Romano, et al., Visual sensing of continuum Biographies
robot shape using self-organizing maps, 2010 IEEE International Conference
on Robotics and Automation (ICRA) (2010) 4591–4596.
[6] C.V. Stewart, C.R. Dyer, The trinocular general support algorithm: a three
Hao Guo received the Bachelor’s degree from Nanjing Forestry University, Nanjing,
-camera stereo algorithm for overcoming binocular matching errors, Second
China in 2017. He is currently a Master’s degree candidate at the College of Mechan-
International Conference on Computer Vision (1989) 134–138.
ical and Electrical Engineering, Nanjing University of Aeronautics and Astronautics,
[7] Y.L. Park, S. Elayaperumal, B. Daniel, S.C. Ryu, M. Shin, J. Savall, R.J. Black, B.
Nanjing, China. His current research interests include continuum robots and surgical
Moslehi, M.R. Cutkosky, Real-time estimation of 3-d needle shape and
robots.
detection for mri-guided interventions, IEEE/ASME Trans. Mechatron. 15 (6)
(2010) 906–915. Feng Ju received the PhD degree from Nanyang Technological University, Singapore.
[8] Y. Shapiro, G. Kosa, A. Wolf, Shape tracking of planar hyper-flexible beams via He is currently an Associate Professor in the School of Mechanical and Electrical
embedded PVDF deflection sensors, IEEE/ASME Trans. Mechatron. 19 (4) Engineering, Nanjing University of Aeronautics and Astronautics. His research inter-
(2014) 1260–1267. ests include piezoelectric sensors and actuators, robotic tactile sensing and motion
[9] M. Cianchetti, F. Renda, A. Licofonte, et al., Sensorization of continuum soft control for robotics.
robots for reconstructing their spatial configuration, IEEE Ras & Embs
International Conference on Biomedical Robotics and Biomechatronics (2012) Yanfei Cao received the Bachelor’s degree from Nanjing University of Aeronautics
634–639. and Astronautics, Nanjing, China in 2016. He is currently a Master’s degree candi-
[10] K. Suzumori, M. Mihara, S. Wakimoto, Beautiful flexible microactuator date at the College of Mechanical and Electrical Engineering, Nanjing University of
changing its structural color with variable pitch grating, IEEE International Aeronautics and Astronautics, Nanjing, China. His current research interests include
Conference on Robotics and Automation (2011) 2771–2776. surgical robots and cable-driven robots.
[11] Z. Chen, Y. Shen, N. Xi, X. Tan, Integrated sensing for ionic polymer-metal
composite actuators using PVDF thin films, Smart Mater. Struct. 16 (2007) Fei Qi received the Master’s degree from Jiangnan University, Wuxi, China in 2016.
262–271. He is currently a PhD degree candidate at the College of Mechanical and Electrical
[12] Y. Bahramzadeh, M. Shahinpoor, Dynamic curvature sensing employing Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing, China.
ionic-polymer-metal composite sensors, Smart Mater. Struct. 20 (2011), His current research interests include surgical robots and robot kinematics.
094011.
Dongmin Bai received the Master’s degree from Shanghai University, Shanghai,
[13] W.S. Rone, P. Ben-Tzvi, Multi-segment continuum robot shape estimation
China in 2015. He is currently a PhD degree candidate at the College of Mechani-
using passive cable displacement, in: 2013 IEEE International Symposium on
cal and Electrical Engineering, Nanjing University of Aeronautics and Astronautics,
Robotic and Sensors Environments (ROSE), Washington, DC, 2013, pp. 37–42.
Nanjing, China. His current research interests include surgical robots and haptic
[14] G. Gerboni, A. Diodato, G. Ciuti, et al., Feedback control of soft robot actuators
feedback.
via commercial flex bend sensors, IEEE/ASME Trans. Mechatron. 22 (4) (2017)
1881–1888. Yaoyao Wang received the Bachelor’s degree from Southeast University, Nanjing,
[15] S. Ozel, N.A. Keskin, D. Khea, C.D. Onal, A precise embedded curvature sensor China in 2011 and the PhD degree from Zhejiang University, Hangzhou, China in
module for soft-bodied robots, Sens. Actuators A Phys. 236 (2015) 349–356. 2016.He is currently a Lecturer in the College of Mechanical and Electrical Engi-
[16] L.K. Simone, D.G. Kamper, Design considerations for a wearable monitor to neering, Nanjing University of Aeronautics and Astronautics, Nanjing, China. His
measure finger posture, J. Neuroeng. Rehabil. 2 (1) (2005) 5. current research interests include robust control and adaptive control of underwater
[17] M.A. Davis, A.D. Kersey, J. Sirkis, et al., Shape and vibration mode sensing vehicle-manipulator system, design and control of cable-driven manipulators.
using a fiber optic Bragg grating array, Smart Mater. Struct. 6 (1996) 759–765.
[18] R.J. Roesthuis, M. Kemp, J.J.V.D. Dobbelsteen, et al., Three-dimensional needle Bai Chen received the Bachelor’s and PhD degrees from Zhejiang University,
shape reconstruction using an array of fiber Bragg grating sensors, IEEE/ASME Hangzhou, China, in 2000 and 2005, respectively. He is currently a Professor and
Trans. Mechatron. 19 (4) (2014) 1115–1126. PhD candidate supervisor at the College of Mechanical and Electrical Engineer-
[19] R.J. Roesthuis, S. Misra, Steering of multisegment continuum manipulators ing, Nanjing University of Aeronautics and Astronautics, Nanjing, China. His current
using rigid-link modeling and FBG-based shape sensing, IEEE Trans. Robot. 32 research interests include design and control of surgical robots, cable-driven robots
(2) (2017) 372–382. and sperm-like swimming micro robots.
[20] S. Song, Z. Li, Q.H. Meng, et al., Real-time shape estimation for wire-driven
flexible robots with multiple bending sections based on quadratic Bézier
curves, IEEE Sens. J. 15 (11) (2015) 6326–6334.