You are on page 1of 20

This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON ROBOTICS 1

A Compliant Adaptive Gripper and Its Intrinsic


Force Sensing Method
Wenfu Xu , Senior Member, IEEE, Heng Zhang , Han Yuan , Member, IEEE, and Bin Liang , Member, IEEE

Abstract—Grasping unstructured objects and sensing the con- developing a gripper with simultaneous adaptive grasping and
tact force are two vital issues for grippers. However, it is still force sensing abilities. Nevertheless, it is theoretically signifi-
difficult for most existing grippers to realize these two functions
cant and practically valuable to solve the above two problems.
simultaneously. In this article, we revise the traditional fin-ray
finger by inserting a series of rigid nodes into the compliant On the one hand, adaptive grasping is important [11]–[14].
structure and develop an adaptive two-finger gripper. This design Although grasping regular objects in controlled environments is
linearizes the gripper’s deformation-force relationship and enables not difficult, it is valuable to achieve adaptive grasping for most
an intrinsic force sensing ability without any tactile sensor. Experi- common scenarios in daily life, where grippers have to deal with
mental results show that the finger has high accuracy in sensing the free-form objects in unstructured environments. On the other
external force applied at its middle part (average error less than
3%) but much larger errors appear near its two ends. Further hand, it is vital to control the contact forces between grippers and
experiments indicate that the gripper functions well in sensing the objects [2], [10], [15]–[18]. This is because excessive forces may
total grasping force (average error less than 8%). Although larger cause damage to objects, especially for fragile or soft ones, while
errors are observed in estimating the force distribution at each insufficient forces may decrease stability and even lead to task
node, the variation tendency of the sensed force coincides well with failure. The grasping force estimation and control can greatly
the ground truth. Experiments are also carried out on grasping
free-form objects and performing pick-and-place operations to improve the grasping performance. This article aims to develop
further prove the gripper’s adaptive grasping and intrinsic force a gripper that not only can grasp free-form objects adaptively
sensing abilities. but also can sense the grasping forces in real time.
Index Terms—Adaptive grasping, artificial neural network,
compliant gripper, finite element method (FEM), force sensing,
improved fin-ray structure. A. State of the Art
Generally, most current grippers can be divided into two
I. INTRODUCTION categories according to how their main movements are achieved.
The first category consists of rigid grippers whose movements
RASPING is one of the most crucial functions of robots,
G which enables various applications from factories to
homes [1]–[3]. Although substantial achievements have been
are mainly realized by kinematic pairs while the structural de-
formation is negligible. Most transitional grippers widely used
in the industry belong to this category [19], [20], such as the
witnessed in robotic grasping during the past few decades [4]– clamps shown in Fig. 1(a). This type of grippers often contains
[7], grasping objects with free-form shapes and sensing the ac- a couple of rigid links and joints and generally functions by
companying grasping forces remain difficult [8]–[10], let alone translating or rotating a pair of symmetric claws. These grippers
usually have only one degree of freedom (DOF) and benefit
from simple structure, easy control, and good reliability, but
Manuscript received September 30, 2020; revised December 21, 2020; ac- they have weak adaptivity to objects with different shapes due
cepted February 17, 2021. This work was supported in part by the National to the limited DOF. To enhance adaptivity, some grippers adopt
Key R& D Program of China under Grant 2018YFB1304600, in part by the
National Natural Science Foundation of China under Grant 61803125, in part multifinger structures [21], [29]–[31], such as the three-finger
by the National Natural Science Foundation of Guangdong Province China gripper shown in Fig. 1(b). Grippers of this type, still belonging
under Grant 2018A030313247, in part by the Basic Research Program of to the rigid category, are mainly composed of a number of rigid
Shenzhen under Grant JCYJ20190806144416980, and in part by the Natural
Scientific Research Innovation Foundation in Harbin Institute of Technology links and joints. Their DOFs are increased, which results in better
under Grant HIT.NSRIF.2020080. This article was recommended for publication adaptivity and dexterity, but more complexity and less reliability,
by Associate Editor J. Ueda and Editor M. Yim upon evaluation of the reviewers’ compared with the aforementioned clamps. Inspired by human
comments. (Corresponding author: Han Yuan.)
Wenfu Xu, Heng Zhang, and Han Yuan are with the School of Me- hands, another type of rigid grippers is invented that often
chanical Engineering and Automation, Harbin Institute of Technology, Shen- consists of one palm and multiple (usually five) fingers including
zhen, Shenzhen 518055, China (e-mail: wfxu@hit.edu.cn; horacez@126.com; numerous rigid joints and links [22], [32]–[34], such as the
yuanhan@hit.edu.cn).
Bin Liang is with the Department of Automation, Tsinghua University, Beijing DLR-HIT II hand shown in Fig. 1(c). Compared with other rigid
100084, China (e-mail: bliang@mail.tsinghua.edu.cn). grippers, robotic hands have super-dexterity and much better
This article has supplementary material provided by the authors and color adaptivity for unstructured objects but they often suffer from the
versions of one or more figures available at https://doi.org/10.1109/TRO.2021.
3060971. complex structure, low reliability, low payload-to-self-weight
Digital Object Identifier 10.1109/TRO.2021.3060971 ratio, and high cost.
1552-3098 © 2021 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See https://www.ieee.org/publications/rights/index.html for more information.

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON ROBOTICS

Another important issue is grasping force sensing. Unfortu-


nately, many existing grippers, including both rigid and soft
types, have at best very limited sensing ability. This is because
it is a great challenge to find a way to obtain the contact forces
when considering precision, reliability, robustness, stability, and
cost. Current force sensing methods mainly focus on developing
diverse sensors, often force sensors, and integrating them into
grippers to directly measure the contact forces. As a result,
various kinds of sensors have been invented and tested on
grippers [51]–[54]. Most of these sensors are designed based
on force-sensitive resistors or resistive foil stain gauges and
mainly consist of flexible thin films and electrodes [55]–[58].
They can be attached to rigid grippers to directly obtain the
contact forces at desired positions and can also be embedded
in certain soft grippers [59], [60], thanks to their structural
flexibility. However, they are not applicable for grippers with
large strains, especially stretchable deformations. For such soft
grippers, stretchable sensors are developed, based on resistance,
capacitance, or optical structures [61]–[64].
Fig. 1. Typical structures of current robotic grippers. (a) Widely used clamps
from Festo (left) [19] and ROBOTIQ (right) [20]. (b) Three-finger gripper
The above sensor-based force sensing methods are direct and
from ROBOTIQ [21]. (c) DLR-HIT II hand [22]. (d) Pneumatically actuated intuitive but have limitations. First, the mass and volume of most
soft gripper [23]. (e) Hydraulically actuated soft gripper [24]. (f) Grasping by conventional force sensors are large, bringing big challenges to
granular jamming [25]. (g) Soft gripper made of dielectric elastomer [26]. (h)
Soft gripper made of shape-memory material [27]. (i) Fin-ray-based adaptive
the integration with grippers. Although certain sensors could be
gripper from Festo [28]. (a)–(c) belong to the rigid category and (d)–(f) belong designed and fabricated in tiny sizes using some advanced tech-
to the soft category. nologies such as the microelectromechanical system [65], their
accuracy and reliability remain an open question. Second, these
delicate sensors are usually sensitive to environments and have
The second category consists of soft grippers where structural high requirements on temperature, humidity, cleanliness, and
deformations contribute to a major part or all of the grippers’ corrosivity. Thus, it brings many concerns in reliability, stability,
movements. Theoretically, soft grippers can have infinite DOFs and cost, which consequently limit the applications of grippers.
by replacing rigid joints and links with compliant structures that This could also partly explain why very few commercialized
exhibit continuous deformations when interacting with objects. grippers have integrated force sensors. Third, adding external
Therefore, soft grippers demonstrate excellent adaptivity to sensors to grippers can bring difficulties to structural designs
a large range of objects with various shapes. The compliant and fabrications and sometimes can even affect the grippers’
structural deformations can be achieved in numerous manners, physical properties. For example, it is troublesome to arrange
such as inflating the chambers made from elastomers pneumat- the sensor wires, especially when the number is large. The
ically [23], [35]–[38] [see Fig. 1(d)] or hydraulically [24], [39] stiffness and flexibility of the soft grippers can also be changed
[see Fig. 1(e)], vacuuming the granular jamming bags filled with if embedding or adhering the external sensors.
granules [25], [40], [41] [see Fig. 1(f)], applying high voltages Besides adding sensors to directly measure the forces, other
(usually several kilovolts) to dielectric elastomers [26], [42] [see indirect force sensing methods have been proposed based on
Fig. 1(g)], and changing temperatures to induce phase changes of the compliant nature of soft robots, which are also called in-
shape-memory materials [27], [43], [44] [see Fig. 1(h)]. These trinsic force sensing methods [66], [67]. The general ideas of
soft grippers realize deformations in active manners that rely these intrinsic methods are modeling the relationship between
on fluidic sources, high-voltage environments, or temperature- deformations and forces, measuring the robots’ deformations,
adjusting devices, which restrict the applications of these grip- and finally calculating the contact forces. This way, the force
pers to particular scenarios. On the contrary, structural defor- sensing problems can be transferred into static modeling and
mations can also be obtained by passive manners [45], [46]. For deformation measuring issues. The intrinsic force sensing meth-
example, the famous fin-ray-like structure can passively fit its ods have been explored in some previous research works, where
profile to the object’s shape during the contacting process [12], the robots’ deformations are obtained by cameras [66], elec-
[28], [47]–[50] [see Fig. 1(i)]. Moreover, these compliant struc- tromagnetic trackers [68], and fiber Bragg gating [69], and the
tures can be easily actuated by external devices, usually electric robots’ deformation-force relationships are modeled based on
motors, which are independent of the main structures and have a virtual work [70], Newton–Euler iterations [68], Cosserat-rod
wide range for selection. Thus, various mature mechanisms and theory [67], and finite element method (FEM)[71]. However,
control methods can be easily applied, which makes the design, these existing methods have limitations. On the one hand, most
fabrication, and implementation of such compliant grippers of them focus on continuum or soft robotic arms that have
practical and convenient. long thin structures, while applications of such methods on

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

XU et al.: COMPLIANT ADAPTIVE GRIPPER AND ITS INTRINSIC FORCE SENSING METHOD 3

soft grippers have not been reported. On the other hand, the sensor. Furthermore, the proposed gripper has a simple structure
deformation-force models of these robotic arms are usually non- and the whole system can be easily built at a low cost. In
linear and complicated, bringing many concerns on computation addition, no extra sensor is added to the fingers, which enables
time and sensing accuracy. the grasping and sensing with high robustness, good stability,
and reliability. Moreover, the proposed force sensing method is
based on general techniques but realizes high precision and good
B. Contribution real-time capability. It should be noted that the current gripper
The major objective of this research is to develop a gripper is based on our previous work in [72] in which an adaptive
with simultaneous adaptive grasping and force sensing abilities. fin-ray-based gripper was proposed and tested, but the design
For this purpose, we design a compliant gripper based on fin-ray and methodology presented in [72] were quite different from
structure, investigate its intrinsic force sensing method, and this work. Our previous gripper adopted the traditional fin-ray
finally realize free-form object grasping and real-time force structure, which does not contain any inserted rod in the finger.
sensing. The proposed compliant gripper is inspired by the Also, it did not have the intrinsic force sensing ability, while the
traditional fin-ray structure, which has a strong adaptive grasping grasping force curves in [72] were actually measured by pressure
ability for objects with diverse shapes. However, different from sensors.
the most existing fin-ray-based grippers [12], [47]–[49], our
gripper has the extra intrinsic force sensing ability. The key C. Organization
reason is the introduction of nodes. Specifically, we insert a
The rest of this article is organized as follows. First, the
series of rigid rods into the compliant fin-ray structure along the
structural design of the compliant gripper is presented in
finger’s depth direction. Thus, the finger’s stiffness is greatly
Section II. Then, the FEM-based force-to-deformation model
enhanced in the depth direction, while its stiffness remains
and the NNM-based force sensing method are detailed in Sec-
almost unchanged in the plane that is perpendicular to the depth
tions III and IV, respectively. Furthermore, experimental val-
direction. As a result, the finger’s deformations are constrained
idations and analyses are carried out in Section V. Finally,
in a plane. Finally, the relationship between the external forces
discussions are made in Sections VI. Section VII concludes this
and the compliant deformations can be linearized without losing
article.
much accuracy.
To sense the grasping forces, the following procedures are
made. II. STRUCTURAL DESIGN
1) We first obtain the finger’s force-to-deformation model This section presents the structural design of the compliant
using the common FEM and generate numerous sample gripper. The major design objectives are the adaptive grasping
data consisting of node forces and the corresponding node capacity, the potential force sensing ability, and the simple
displacements. actuating mechanism to achieve grasping.
2) Then, we train the finger’s deformation-to-force model, To realize adaptive grasping, the gripper is designed based on
i.e., the force sensing model, based on the general neural the fin-ray structure. Inspired by fish fins, the fin-ray structure
network method (NNM) using the abovementioned sam- has a special deformation feature [48]. If one side of the structure
ple data. is subjected to an external force, the structure tip and base bend
3) Next, we use a hand-eye camera to measure the node toward the opposite direction of the applied force. Therefore,
coordinates and calculate the node displacements. when the fin-ray structure touches an object, it deforms and
4) Finally, the node forces can be estimated by substituting passively adapts to the shape of the contact surface, resulting
the node displacements into the force sensing model. enveloping grasping effect. A number of grippers have been
It should be noticed that steps 1) and 2) are performed of- developed based on the fin-ray structure [47], [48], [50], [72],
fline, while steps 3) and 4) are implemented online that are [73], exhibiting a wide range of applications that realize adaptive
simultaneous with the grasping process. Therefore, the proposed grasping on various objects, including fruits, eggs, biological
force sensing method is real time. It should also be noted that tissues, and articles of daily use. Inspired by these studies, we de-
it is not practical to simply invert the FE model to obtain the sign a compliant gripper containing two identical fin-ray-based
force sensing model because the inputs of the inverted FE model fingers, as shown in Fig. 2. Each finger consists of a base, a
contain all the node deformations including the node rotations V-shape side plate, and several rib plates. Detailed parameters
that are very difficult to accurately measure or estimate. Due to of the finger are shown in the left part of Fig. 2(a).
the limit in deformation measurement, we further establish the Another vital purpose of our structural design is intrinsic
force sensing model based on the general NNM, which avoids force sensing. As explained before, the principle of intrinsic
using such node rotations. force estimation is transferring the direct force sensing prob-
The major contribution of this article is the insertion of rigid lem to the indirect deformation measurement by the structure’s
nodes to the traditional fin-ray structure, which linearizes the force-deformation model. Therefore, it is important to design
structure’s force-deformation relationship and, thus, enables the a structure that has a simple and explicit force-deformation
accurate intrinsic force sensing ability. So far as we know, this relationship, preferably a linear one. However, the traditional fin-
is the first fin-ray-based gripper that simultaneously achieves ray-based grippers usually have a nonlinear force-deformation
adaptive grasping and intrinsic force sensing without any force relationship, such as the gripper in [74], where a nonlinear static

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON ROBOTICS

Fig. 2. Mechanical design of the compliant gripper. (a) Structural parameters of the compliant finger. (b) Detail of the nodes. (c) Actuating and transmitting
mechanism of the compliant gripper.

model was established and the relative difference between the step angle of 1.8◦ . The screw has a pitch of 4 mm and an
experimental results and the model predictions was about 25%. outer diameter of 12 mm. Other parameters of the mechanism
To solve this problem, we modify the traditional fin-ray structure are shown in Fig. 2(c), and its kinematics will be addressed in
by inserting a series of rigid rods along the intersecting lines of Section IV-C.
the rib plate and side plates, as shown in Fig. 2(b). The nominal The compliant fingers are made of thermoplastic polyurethane
diameter of the rod is the same as that of the hole, and a press-fit (TPU). Generally, TPU has a maximum recoverable strain of
is used to tightly embed the rod inside the hole. Thanks to 50% and Young’s modulus ranging from 10 to 1000 MPa [75].
these rods, the gripper’s stiffness along the depth direction is Specifically, the TPU used in this article is provided by Miracll,
enhanced, but its compliance in the plane that is perpendicular China, and the model number is E175, with Young’s modulus
to the depth direction is little affected. Therefore, the gripper’s of about 48 MPa. The fingers are fabricated using injection
overall deformation can be simplified from three-dimensional molding. It should be noted that TPU is chosen due to its good
(3D) to 2-D. Furthermore, the force-deformation relationship elasticity and easy molding character, while the construction
can be linearized and modeled by the traditional FEM, taking method is scalable to other materials that have similar properties,
the rigid rods as the FE nodes and the flexible plates between such as thermoplastic elastomer and rubber. In addition, the
every two adjacent rods as the FE elements. Further details about inserted rods are made by a 3-D printer using a kind of ABS
the FEM-based force-deformation model will be addressed in plastic. Compared with the finger-material TPU, the rod material
Section III. ABS has a similar density but a much higher Young’s modulus,
To achieve the movement needed for grasping, a simple ranging from 1500 to 3000 Mpa. Therefore, the inserted rods
mechanism is designed. As shown in Fig. 2(c), the actuating and can be regarded as rigid with respect to the TPU-made finger.
transmitting system mainly consists of a stepper motor, a screw, It should be noted that the rod diameter should be large enough
and two bilateral symmetric planar slider-rocker mechanisms. to enhance the gripper’s stiffness along the inserting direction,
The motor’s rotation is first transferred to the nut’s locomotion but not too large to affect the overall continuum deformation of
by the screw-nut. Then the nut, which functions as a slider, the flexible finger. Based on our experiences, we suggest a rod
actuates the two slider-rocker mechanisms. Finally, the gripper diameter larger than the thickness of the TPU plate (1.5 mm) but
movement is obtained by employing the finger’s base as the smaller than one-third of the side-plate length (12.75 mm), i.e.,
connecting link of the slider-rocker mechanism. The stepper ranging from 1.5 to 4.25 mm. In this article, the final diameter
motor has a standard torque of 0.35 N · m and a minimum of each rod is 3 mm. It should be noted that the rod dimension

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

XU et al.: COMPLIANT ADAPTIVE GRIPPER AND ITS INTRINSIC FORCE SENSING METHOD 5

and fi = [f1i



, f2i 
, f3i 
, f4i 
, f5i  T
, f6i ] , which are both expressed
in the local frame Ri {xi oi yi }. According to the classical finite
element theory [76], we can define a local stiffness matrix kei to
map the elements’ displacements to forces by fi = kei · qi and
kei is
⎡ EAi ⎤
li 0 0 − EA li
i
0 0
⎢ 0 12EIi 6EIi
0 − 12EI i 6EIi ⎥
⎢ li3 li2 li3 li2 ⎥
⎢ 6EIi 4EIi 6EIi 2EIi ⎥
⎢ 0 0 − ⎥
kei = ⎢ ⎥ (1)
li2 li li2 li
⎢ − EAi 0 0 EAi
0 0 ⎥
⎢ li li ⎥
⎢ 0 − 12EIi − 6EIi 0 12EIi
− 6EIi ⎥
⎣ li3 li2 3
li li ⎦
2
6EIi 2EIi 6EIi 4EIi
0 l2 li 0 − l2 li
i i

where E is Young’s modulus, and Ai , li , and Ii are respectively


the cross-sectional area, length, moment of inertia of the ith
Fig. 3. Schematic of the compliant finger by FEM. (The red lines and red
numbers represent the elements and their sequence; the dark blue points and element. Then, a transformation matrix Ti is formulated to
black-circled number represent the nodes and their sequences.). define the relationship between the local and global frames by
fi = Ti · fi and qi = Ti · qi . Ti can be written as
⎡ ⎤
cos θi sin θi 0 0 0 0
together with other structural parameters, such as the finger size ⎢ − sin θi cos θi 0 0 0 0⎥
⎢ ⎥
and the plate thickness, can further be optimized to improve ⎢ 0 0 1 0 0 0⎥
Ti = ⎢⎢ ⎥ (2)
the grasping performance. However, these issues are beyond
⎢ 0 0 0 cos θi sin θi 0 ⎥ ⎥
the major objective of this article, which could be scheduled in ⎣ 0 0 0 − sin θi cos θi 0 ⎦
future research. 0 0 0 0 0 1

III. FORCE-TO-DEFORMATION MODEL BY FEM where θi is the inclined angle of the local frame Ri with respect
to the global frame R. Then, the global stiffness matrix of the
In Section II, we proposed a modified fin-ray-like finger by in- ith element kei can be obtained using the local stiffness matrix
serting rods into the traditional fin-ray structure. The new finger kei and the transformation matrix Ti
is supposed to have a linear force-deformation relationship. This
section will use the traditional linear finite element theory [76] kei = TTi · kei · Ti . (3)
to present the force-to-deformation model, where the linearity
refers to small nodal displacements and elemental strains. The Third, the global stiffness matrix of the structure can be ob-
validity of the linear model will be experimentally examined in tained by assembling all the elements’ stiffness matrices kei . For
Section V. details, refer to the general FEM [76]. After obtaining the global
First, the FE schematic of the compliant finger with the stiffness matrix, boundary conditions should be implemented.
inserted rigid rods is simplified as a planar frame, which is shown For our finger, nodes 1 and 25 are taken as the constraints because
in Fig. 3, where red lines and dark blue points represent elements these two nodes are located at the finger base that has negligi-
and nodes of the FE model, respectively. Furthermore, all the ble deformation during grasping. Moreover, the coordinates of
nodes and elements are scheduled using black-circled numbers nodes 1 and 25 can be easily obtained according to the gripper’s
and red numbers. It should be noted that here we use a simple kinematics and the motor’s rotation angle. Finally, the stiffness
discretization strategy to get the FE schematics by directly taking matrix K of the compliant finger is obtained, which reveals the
the rigid rods as nodes and the flexible plates between every two force-deformation relationship of the finger
adjacent rods as elements. Nevertheless, it is also possible to use
F=K·Q (4)
other discretization strategies, such as decreasing the element
size and adding more nodes. where F and Q are the external force vector and the deformation
Second, the stiffness matrix of a single element is determined. vector taking all the nodes into account. For the planar beam
For the proposed finger, the beam element is adopted. It should element shown in Fig. 3, there are three components for the ex-
be noted that all elements are beams and there is no link or rod el- ternal forces and three components for the deformations at each
ement since the proposed finger does not contain any hinge, nei- node. Therefore, Q and F are both 69 × 1 vectors, consisting
ther the revolute joint as proposed in [12] and [49] nor the elastic of the external forces applied to the 23 nodes (2–24) and their
hinge as proposed in [28]. As shown in the top enlarged drawing corresponding deformations. K is a 69 × 69 symmetric matrix.
of Fig. 3, each beam element has four displacements and two If noting C as the compliance matrix that is the inverse of K,
rotations under four external forces and two external torques. For the finger’s force-to-deformation model can be written as
the ith element (i = 1, 2, . . ., 35), its displacement vector and
force vector can be written as qi = [q1i  
, q2i 
, q3i 
, q4i 
, q5i  T
, q6i ] Q = C · F = K−1 · F. (5)

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON ROBOTICS

Fig. 4. Validation of the proposed force-to-deformation model. (a) Profiles of the compliant finger calculated by three models (i.e., the proposed force-to-
deformation model, the ANSYS model with 684 nodes, and the ANSYS model with 25 nodes) under different loading conditions (i.e., external forces with various
amplitudes, directions, and locations). (b) Node displacements obtained by the 25-node ANSYS model and the proposed force-to-deformation model and their
differences.

Through (5), the deformation of the compliant finger under a calculating efficiency, which can greatly accelerate the training
given external force can be calculated. process. By contrast, the commercial software ANSYS has a
In order to validate the above force-to-deformation model, we much lower efficiency in preparing massive sample data. Details
use a commercial software ANSYS1 , and build two kinds of 2-D can further be found in Section IV-A. In addition, results also
models. The two models both use the beam element (BEAM3 show that the maximum difference between the two ANSYS
in ANSYS) but with different element sizes. The first frame models with different element sizes is less than 0.26%. This
model has 25 nodes and 35 elements, which are the same as our indicates that the current mesh shown in Fig. 3 has enough
proposed force-to-deformation model [see (1)–(5)], while the precision while decreasing element size is not necessary in terms
second one increases the quantity of the nodes and elements to of accuracy.
684 and 695, respectively. The overall deformation of the finger In order to illustrate the above-mentioned results, Fig. 4 is
is calculated for 100 different loading conditions with various plotted taking three different loading conditions as examples.
magnitudes and locations, using the two ANSYS frame models Fig. 4(a) presents the overall deformed profiles of the finger.
and the proposed model, respectively. For each loading condition, the finger deformations obtained by
Results show that the above three models have almost identi- three models are plotted, where comparisons show that the three
cal results in calculating the node deformations. Specifically, the methods obtain almost identical deformations. Moreover, the
difference between the proposed model and the ANSYS frame displacements at each node are plotted in Fig. 4(b), correspond-
model with 25 nodes is negligible, with a maximum difference of ing to the above three loading conditions. Results further prove
less than 1.5%. This means that these two models are equivalent that the proposed force-to-deformation model and the ANSYS
to each other but we cannot directly use the ANSYS model model have negligible differences.
in the following force sensing method. The major concern is It should be noted that we also build an ANSYS model
calculating efficiency. In the following neural network training, using solid elements (SOLID45) and 3-D meshes for simulation.
massive sample data (20 000 in this article) are required. The
proposed model is in an analytical form and has an excellent 1 ANSYS is software by Ansys Inc.

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

XU et al.: COMPLIANT ADAPTIVE GRIPPER AND ITS INTRINSIC FORCE SENSING METHOD 7

Results show a large difference, compared with the proposed


model and the 2-D ANSYS model. One possible reason is that
the proposed finger structure contains two different materials
due to the inserted rods, which complicates the 3-D meshes.
Meanwhile, the complex interactions between different materi-
als further increase uncertainty in the simulation using the 3-D
ANSYS model. At this point, neither the proposed model nor the
2-D ANSYS model can be the basis of truth, while the behavior
of a real physical system should be investigated to verify the
proposed model. Experimental validation will be presented later
in Section 5.
Fig. 5. Structure of the neural network, which mainly consists of three parts,
IV. INTRINSIC FORCE SENSING METHOD i.e., the input layer, the hidden layers, and the output layer. Generally, 46 network
nodes are set for the input layer, representing the 46 displacements of the 23 FE
Sensing the grasping force without any force sensor is a nodes (nodes 2–24); 150 network nodes are set for the first and third hidden
significant feature of the proposed gripper. This section presents layers H1 and H3 ; 200 network nodes are set for the second hidden layer H2 ;
the intrinsic force sensing method. First, the deformation-to- finally, m network nodes are set for the output layer, representing m external
forces, where m depends on specific loading conditions.
force model is obtained using the general NNM. Then, the node
coordinates are measured by a hand-eye camera. Finally, node
deformations are calculated and inputted into the deformation-
to-force model to calculate the grasping force.
the above ranges will vary with the material and the structural
parameters of the finger. However, a general rule is that the
A. Deformation-to-Force Model by NNM applied forces should not cause a deformation that exceeds the
In Section III, a simple deformation-to-force model is already structural elastic boundary. Within the above ranges, we generate
presented by (4). According to (4), the applied external force F 20 000 groups of sample data containing the external forces and
can be directly calculated if the node deformation Q is given, the corresponding node displacements. It should be noted that
but this method is not practical in real applications due to the the validity of the FEM-based force-to-deformation model will
following reasons. Q is a displacement vector that contains further be presented in Section V, where physical experiments
not only the node displacements but also the node rotations. are performed at the above boundary of the range.
Meanwhile, the node displacements are relatively easy to mea- Then, we establish a backpropagation NNM, which is a
sure, but the node rotations are difficult to obtain accurately. common multilayer feed-forward network [77]. As illustrated
Honestly, we have tried to estimate the node rotations through in Fig. 5, this model mainly consists of three parts, i.e., the
structural deformation and then directly use (4) to calculate input layer, the hidden layers, and the output layer. For the input
the external forces, but the inaccuracy in rotation estimation layer, generally, we set 46 network nodes that represent the 46
could lead to unacceptable errors in force sensing. Furthermore, displacements of the 23 FE nodes (nodes 2–24 shown in Fig. 3)
we have also tried to regard both node rotations and external since each node has two displacements without considering the
forces as unknowns and use least-square optimization to solve rotation. Nevertheless, the number of the network nodes can
this problem, but the force sensing accuracy is poor and the be less than 46 if the displacements of certain FE nodes are not
convergence of the optimization algorithm is highly sensitive to available, but fewer input nodes could decrease the force sensing
the node displacement errors. Therefore, this simple FEM-based accuracy. Thus, it is necessary to involve more input nodes as
deformation-to-force equation cannot be directly used in the long as their displacements are available. For the hidden layers,
intrinsic force sensing of our gripper. we set H1 , H2 , and H3 , totally three layers. In this article, H1
Due to the abovementioned limitations, we further establish and H3 both have 150 network nodes and H2 has 200 network
the deformation-to-force model by NNM. As well known, the nodes. It should be noted that the structure of the hidden layers
NNM is a kind of data-driven method, which does not need is not unalterable. Both the layer number and the node number
explicit mathematical models but maps the inputs and the outputs can be changed as long as they are within the general framework
through training using numerous sample data. Therefore, if a of NNM [77].
large number of displacement-force samples can be obtained and The output layer consists of the external forces applied to
used to train the NNM, it is possible to establish the mapping the finger and the number of the independent force coordinates
from displacement to force without the rotation information. depends on how many finger nodes have physical contact with
For this purpose, we first generate the sample data for train- the object. For our gripper, since only one side of the finger
ing. According to the force-to-deformation model proposed in contacts objects, there are at most 12 nodes that may have
Section III [see (1)–(5)], it is easy to obtain the node displace- contact, i.e., from node 13 to node 24 (see Fig. 3). For each
ments by inputting any external force. Specifically, for the TPU contacting node, the external force vector contains at most three
finger designed in this article, the range of the external force components, i.e., two forces and one torque. Therefore, the
at a single node is set as 0–5 N, and the range of the resultant output layer herein contains 36 independent force coordinates at
force considering all contacting nodes is set as 0–10 N. In fact, most. However, the number of the independent force coordinates

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON ROBOTICS

Fig. 6. Calculation of the node displacements. (a) Measurement of the node coordinates by a hand-eye camera. (b) Relationship between the undeformed and
deformed finger shapes (dark blue for the deformed, light gray for the undeformed). (c) Kinematics of the slider-rocker mechanism.

should be much less than 36 because only a few finger nodes red. Then, edge detection and extraction are made. A series
could have a contact for most real grasping. Moreover, the of ellipses are obtained in the image plane. After that, ellipse
number of the unknown force coordinates can further be reduced fitting is made using the detected edges and parameters of each
considering other constraints in real grasping. This issue will ellipse are calculated. Finally, the node coordinates in the camera
further be discussed in Section VI. Finally, the neural network coordinate frame are acquired by combining the elliptic equation
is trained and the deformation-to-force model, i.e., the force and the calibrated camera parameters.
sensing model, is established. It should be noted that the force In this article, we use a simple monocular camera provided by
sensing model should be retrained if either the input layer or the DATU company and the model is DYSMT805. This camera has
output layer is changed. a 1/3.2-in colorful CMOS and its resolution is 1600×1400. The
accuracy of the proposed measuring system is experimentally
validated using the compliant finger with known deformations.
B. Node Coordinate Measurement by Hand-Eye Camera Specifically, the actual deformations of the finger are obtained
As shown in Fig. 6, the two compliant fingers, driven by by a group of customized moulds. Each mould has a serial of
the slider-rocker mechanism, will have passive deformations holes with a diameter of 3 mm, which is the same with that
during the grasping process. In fact, these passive deformations of the inserted rods. Meanwhile, the locations of these holes
can be defined by the difference between the deformed and the are calculated by the force-to-deformation model presented in
undeformed finger shapes. Therefore, we need to obtain the node Section III. Therefore, these holes in the mould are very close
coordinates of the deformed and the undeformed fingers in real to the rods in the deformed finger. Consequently, we can easily
time. This section presents how to obtain the node coordinates insert the rod tips into the mould holes, thanks to the compliance
of the deformed finger. of the TPU-made finger. This way, the rod coordinates to be
The general idea for obtaining the node coordinates is to use measured are the same as the mould holes. Since these holes
the noncontact machine vision. Since all the FE nodes of the are fabricated by the machine tool, they have a high precision
finger coincide with the center of the insert rods, it is equivalent (position errors less than 0.01 mm) and can be used as the ground
to measure the coordinates of the rod centers. As shown in truth. It should be noted that we fabricated six moulds that corre-
Fig. 6(a), we fix a small monocular camera on the gripper’s spond to different deformed finger shapes under various loading
base. The camera’s image plane is perpendicular to the axes conditions. Experimental results show that the maximum error
of the inserted rods and the distance between the camera and of the proposed vision-based measuring system is less than
the finger is 170 mm. Through this arrangement, the camera 0.2 mm. It should be noted that such high accuracy could be
can obtain a real-time image containing the rods’ end-surfaces. obtained by a simple monocular camera because the measuring
To recognize the rods’ coordinates, the following procedures targets are cooperative, i.e., the radius of each rod is known,
are made according to the general machine vision method [78]. and the distance between the camera and the finger is fixed.
First, features of the rods’ end-surfaces are segmented using Nevertheless, other methods could be implemented to further
hue, saturation, and value color space segmentation. In order improve the accuracy and robustness of the vision algorithm,
to increase robustness, we paint the rods’ end-surfaces using such as pasting a reflective tape and using an infrared camera,
colors that have distinct differences with the environment and which could be useful for real applications, especially when the
other parts of the finger. For our case, we use yellow and illumination is not good.

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

XU et al.: COMPLIANT ADAPTIVE GRIPPER AND ITS INTRINSIC FORCE SENSING METHOD 9

Another important issue is the rate of the vision-based mea-


surement method. In this article, the maximum frame rate of the
camera at 1600×1400 is 30 frames per second (FPS). Mean-
while, the computing time of the proposed vision recognition
algorithm is about 40 ms, i.e., 25 cycles per second. Therefore,
the maximum measurement rate is 25 Hz. This rate is vital
for the real-time capacity of the force sensing method. Further
discussion on this issue will be made in Section VI.

C. Node Displacement Calculation and Force Sensing


Besides measuring the node coordinates of the undeformed
finger by the hand-eye camera, we also need to know the Fig. 7. Experimental setup for the single finger tests: the compliant finger is
undeformed finger shape to obtain the node displacements. fixed on the frame and loaded by standard weights through a thin wire, and its
node coordinates are measured by the monocular camera in real time.
For this purpose, we first present the kinematics of the slider-
rocker mechanism, then calculate the node deformations by
homogeneous coordinate transformations, and finally achieve
the real-time grasping force sensing. Through (10), we can transfer the node coordinates of the
The kinematic schematic of the slider-rocker mechanism is undeformed finger from the local finger frame Rfng to the global
shown in Fig. 6(c). Since the slider is driven by a screw, the camera frame Rcam for any given motor angle θM .
slider’s height l3 is a variable. It depends on the rotation angle So far, the node coordinates of the deformed and undeformed
of the screw that is equivalent to the motor angle θM . Note the fingers are all expressed in the same frame, i.e., the camera
screw pitch as ps and the initial height of the slider as l03 , then frame. Thus, the nodes’ displacements can be easily obtained.
the slider’s height l3 can be calculated as Finally, the contact force during the whole grasping process can
θM be calculated in real time by substituting the nodes’ displace-
l3 = l03 + · ps . (6) ments into the NNM-based force sensing model.

Besides, the kinematics of the slider-rocker mechanism can be
expressed as V. EXPERIMENTAL VALIDATION AND ANALYSIS
l4 = l2 · cosφ2 − l1 · cosφ1 (7) This section will present the experimental validation to verify
the proposed design and sensing method. For this purpose, we
l3 = l1 · sinφ1 + l2 · sinφ2 (8) first calibrate Young’s modulus of the finger, which is a vital
where l1 , l2 represent the link lengths, l4 gives the location of the material parameter for the following force sensing experiments.
fixed joint, and φ1 and φ2 give the links’ incline angles. l1 , l2 , and Second, we perform experimental validations on a single finger
l4 are structural parameters that can be found in Fig. 2, while and investigate the sensing accuracy at different locations along
φ1 and φ2 are variables associated with the motor angle θM . the finger. Third, we carry out three groups of grasping exper-
According to (7) and (8), φ1 and φ2 can be solved analytically. iments on the whole gripper and examine the sensing accuracy
Then, we can calculate the node coordinates of the unde- for different topologies of the contact points. Finally, the gripper
formed finger. To do this, we first establish three frames, i.e., is tested by grasping experiments on real-world objects with
the finger frame Rfng , the slider-rocker frame Rmech, and the free-form shapes and by pick-and-place operations with the help
camera frame Rcam , as shown in Fig. 6(c). Noting Rmech TRfng of a robotic arm.
as the transformation matrix from Rfng to Rmech , we can obtain
Rmech
TRfng through a rotation Rz (φ2 ) about the z-axis and a A. Young’s Modulus Calibration
translation POfng
 Young’s modulus is an important parameter of our force
Rmech Rz (φ2 ) POfng sensing model. As explained in the structural design part, our
TRfng = (9)
0 1 finger is made of TPU whose Young’s modulus depends on the
material composition and also relates to fabrication procedures.
where POfng = [−l1 · cosφ1 , l1 · sinφ1 , 0]T . In addition, the
Therefore, the exact value of Young’s modulus needs calibration
transformation from the slider-rocker frame Rmech to the cam-
after the finger is fabricated.
era frame Rcam can be obtained by a translation POmech =
Here, we design a simple device to perform Young’s modulus
[−wcam , −hcam , lcam ]T , where lcam is the distance between the
calibration. As shown in Fig. 7, we fix the compliant finger on
camera and the nodes’ plane, and wcam and hcam are the hor-
the frame and apply a vertical force to a node by hanging a group
izontal and vertical distances between the origin points Omech
of standard weights through a thin wire. Meanwhile, we use a
and Ocam , respectively. Therefore, the transformation matrix
Rcam monocular camera to measure the coordinates of the colored
TRfng from Rfng to Rcam can be obtained by
 rods in real time. The magnitude of the applied force can be
Rcam Rz (φ2 ) POfng + POmech adjusted by hanging different standard weights. The calibration
TRfng = . (10)
0 1 procedures are as follows.

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON ROBOTICS

1) An image of the finger is captured at the initial state 4) Steps 2) and 3) are repeated six times by changing the
without hanging any weight and the node coordinates loading location, i.e., hanging the standard weight at nodes
(nodes 2–24) of the undeformed finger are recognized. 16, 18, 20, 22, and 24, respectively.
2) A standard weight is hung at node 17 and the weight Through the above experiments, we obtain the force sensing
varies from 50 to 300 g with an increment of 50 g each performance of a single finger under different loading condi-
time. Thus, there are totally six different standard weights. tions. Results are listed in Table I. The average relative errors
For each standard weight, an image of the deformed at the six nodes, from node 14 to node 24 (i.e., from tip to
finger is captured after static equilibrium to recognize the base), are respectively 56.3%, 12.1%, 1.4%, 1.5%, 2.4%, and
node coordinates (nodes 2–24). Consequently, the node 20.0%. We also gray the errors that exceed 10% and darken
displacements are obtained by comparison with the node the errors that exceed 25%. A clear tendency is found that the
coordinates of the undeformed finger in step 1). It should sensing accuracy is high when the external force is applied at
be noted that the experiment is repeated three times for the middle part of the finger. For example, the average sensing
each standard weight and the average value is used to error at nodes 18, 20, and 22 is less than 3%, and even the
reduce random errors. maximum error at these three nodes is less than 5%. However,
3) The same structural parameters and loading conditions are the sensing accuracy tends to be worse when the contacting point
set in the FEM-based force-to-deformation model to cal- moves toward the finger tip or base. For example, the maximum
culate the node displacements (nodes 2–24). It should be sensing error at nodes 16 and 24 reaches 18% and 24.2%, while
noted that Young’s modulus E in the FEM-based model is the average errors are 12.1% and 20.0%. Moreover, when the
an unknown parameter to be calibrated, thus the resulting contacting point is very close to the finger tip such as node 14,
node displacements are the function of E. the sensing error is large where almost all errors exceed 50%.
4) The above steps are repeated by changing the hanging To sum up, the proposed fin-ray-based finger has good force
position to node 19 and node 21, respectively. sensing accuracy when contacting at its middle part but has
Finally, Young’s modulus E can be calibrated by minimizing worse sensing accuracy when contacting near the tip or base,
the total difference of all the node displacements between the especially the tip.
measured value and the FEM-calculated value, which is mathe- Further analysis shows that the sensed forces at nodes 14,
matically expressed as 16, and 24 are all larger than the ground truths. Meanwhile, the
variance of the relative errors at the same node is not large. For
example, although nearly all errors at node 14 exceed 50%, their
E = min ei,j,ki = 2 ∼ 24, j = 1 ∼ 6, k = 1 ∼ 3
variation is small, which ranges from 49.6 to 63.7%. Therefore,
(11) these errors are not random errors but systemic errors, which
where ei,j,k is the node’s Euclidean distance between the mea- are caused by the nonlinearity of the compliant finger. In fact,
sured value and the FEM-calculated value for the ith node, when the real finger deformations are larger than the model-estimated
the jth standard weight is hung at the kth node. Following the deformations at these nodes. Consequently, the real equivalent
above process, the calibrated Young’s modulus of our finger is Young’s modulus at the tip and base should be lower than the
4.7976 × 107 Pa. above-calibrated value, which only uses the data of the three
middle nodes 17, 19, and 21. However, recalibration using
more nodes is not recommended because this will decrease
B. Force Sensing Experiments on a Single Finger the sensing accuracy at the middle part, although it could
In order to validate the proposed force sensing method, ex- improve the accuracy at the tip and base. Actually, for most
periments are performed first on a single finger. The basic idea applications, it is vital to guarantee the force sensing accuracy
is to apply a given force to the finger and use the proposed at the middle since the grasping operation of fin-ray-based
sensing method to estimate the force, and then compare the grippers usually has better stability and reliability when the
sensed value with the given value. For this purpose, a standard contact occurs at the middle part of the finger. Nevertheless,
weight is hung on the finger to precisely generate the required if high sensing accuracy at the finger tip is required in particular
force using the same device shown in Fig. 7. Experiments are situations, the above errors are also possible to be well compen-
carried out according to the following procedures. sated by various methods, such as using nonconstant Young’s
1) The coordinates of all nodes (nodes 2–24) are measured modulus.
by the camera when zero load is applied and taken as the
initial value of the undeformed finger.
2) A standard weight of 50 g is hung at node 14. The node C. Force Sensing Experiments on the Compliant Gripper
coordinates (nodes 2–24) at static equilibrium are mea- Besides validation on a single finger, we also carry out
sured by the camera and inputted into the force sensing grasping experiments on the whole gripper to further verify the
model to calculate the external force. This step is repeated proposed force sensing method.
for three times to get an average value of the sensed force. The experimental setup is shown in Fig. 8. The two-finger
3) Step 2) is repeated six times, and for each time, a different gripper is fixed on the base and actuated by a stepper motor that
standard weight is hung. The six weights range from 50 is controlled by a slave computer (STM32 board). A hand-eye
to 300 g with an increment of 50 g. camera is fixed in front of the fingers to measure the deformation

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

XU et al.: COMPLIANT ADAPTIVE GRIPPER AND ITS INTRINSIC FORCE SENSING METHOD 11

TABLE I
RESULTS OF THE FORCE SENSING EXPERIMENTS ON A SINGLE FINGER: COMPARISON BETWEEN THE SENSED VALUE AND THE GROUND
TRUTH AT DIFFERENT LOADING POSITIONS

which the analog voltage signal is converted to the digital signal,


then mapped to the force value, and finally sent to the host
computer. It should be noted that the structural parameters of the
proposed sensor holders are carefully designed with the purpose
of making the fingers’ nodes correspond well to the sensors’
centers. Therefore, the grasping forces at certain nodes of the
finger can be precisely measured by the corresponding force
sensors.
Using the above experimental system, three groups of tests are
respectively performed using the modules shown in Fig. 8(a)–
(c), when a single node, two adjacent nodes, and two nonadjacent
nodes of the finger contact the sensor module. Meanwhile,
each group contains several grasping tests at different contact
Fig. 8. Experimental setup for the force sensing tests on the compliant gripper.
positions. Experiments are performed according to the following
(a) Overall layout: the two-finger gripper equipped with a hand-eye camera is procedures.
fixed on the frame base; a 3-D printed object with embedded pressure sensors is At the beginning, the sensor module is placed on the vertical
placed but not fixed on top of a vertical support; the force sensors are only
used to get the ground truth for validation purpose; sensors are calibrated
support, and the two fingers are opened to embrace the object but
before embedded. (b) Sensor module for single-node contacting tests. (c) without physical contact, referring to the left part of Fig. 8. Then,
Sensor module for two-adjacent-node contacting tests. (d) Sensor module for the fingers are actuated by the motor and gradually grasp the
two-nonadjacent-node contacting tests.
sensor module. It should be noted that the driving motor rotates
discontinuously during the grasping process. Specifically, the
motor is first rotated by 22.5◦ , then it pauses. At this moment,
during the whole grasping process in real time. The object to be the grasping forces are measured by the force sensors, and si-
grasped is placed on top of a vertical support but not fixed on it to multaneously calculated by the proposed force sensing method.
avoid interference during grasping. Both the slave computer and After this, the motor continues its rotation (22.5◦ every time),
the camera are connected to a host computer (Intel NUC) where and the force measuring and sensing procedures go on. Through
the proposed force sensing algorithm and the system control the above discretization, the sensed grasping force can be easily
programs are running through robot operating system. compared with the measured ground truth, which facilitates the
An important issue of the experimental validation is how data analysis. Nevertheless, the grasping operation can be done
to accurately obtain the real forces during the whole grasping continuously and force sensing can be achieved in real time.
process. In this article, we use a common method to get the Examples can be found later in Fig. 16.
ground truth, i.e., directly measuring the real contact force by After one test is finished, the gripper is reset and the height of
force sensors. However, this method is different from those the sensor module is adjusted to perform another test at different
in some previous literature works where the force sensors are contact positions. After one group of tests is finished, the force
simply placed on or embedded in the grippers. In this article, module is changed to perform another group of tests. Results of
we integrate the force sensors into the object while not changing the three groups of tests are shown in Figs. 9– 11. Moreover, the
the gripper to prevent affecting its physical properties. Moreover, average errors of all tests are calculated and listed in Table II.
due to the large flexibility of the finger, the traditional rigid First, as shown in Fig. 9, the sensed grasping forces (blue
sensors are difficult to attach and flexible sensors are hard to lines with stars) are compared with the measured value (red
achieve high accuracy. lines with circles), when the contacting node is 21, 19, and 17,
As shown in Fig. 8(b)–(d), we design a group of sensor mod- respectively. Moreover, the absolute errors (blue texts and lines)
ules that are composed of four force sensors, a frame to hold the and relative errors (red texts and lines) of the force sensing model
sensors, and several caps. These force sensors are provided by are also plotted. Generally, the sensed forces coincide well with
Simbatouch, China, and the model is SBT674-10. Each sensor the ground truths. Specifically, as listed in Table II, the overall
has a measuring range of 0–10 N and an accuracy of 0.01 N. average absolute error is 0.07 N when the grasping forces vary
The force sensor is connected to the slave computer through from 0 to 3 N, and the average relative error is 6.5% when the

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

12 IEEE TRANSACTIONS ON ROBOTICS

Fig. 9. Experimental validation of the proposed force sensing method on the gripper, when a single node of the finger contacts the sensor module. (a) Contact at
node 21. (b) Contact at node 19. (c) Contact at node 17.

grasping forces vary from 0.3 to 3 N. It should be noted that Second, experiments are further performed when two adjacent
small grasping forces (less than 0.3 N here) are not considered nodes contact the object. Results are shown in Fig. 10(a) and (b)
when calculating the relative errors. This is because, for small when nodes 19, 20 and 16, 17 respectively contact the sensor
grasping forces, a small absolute force sensing error may cause module. It is found that the force sensing errors are large if
an extremely large relative error. looking at any one of the two contacting nodes, where the

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

XU et al.: COMPLIANT ADAPTIVE GRIPPER AND ITS INTRINSIC FORCE SENSING METHOD 13

Fig. 10. Experimental validation of the proposed force sensing method on the gripper, when two adjacent nodes of the finger contact the sensor module.
(a) Contact at nodes 19 and 20. (b) Contact at nodes 16 and 17.

TABLE II
AVERAGE FORCE SENSING ERRORS OF THE GRASPING EXPERIMENTS: DIFFERENCES BETWEEN THE MEASURED AND SENSED FORCES

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

14 IEEE TRANSACTIONS ON ROBOTICS

Fig. 11. Experimental validation of the proposed force sensing method on the gripper when two nonadjacent nodes of the finger contact the sensor module.
(a) Contact at nodes 18 and 20. (b) Contact at nodes 16 and 20.

coincidence between the sensed forces and the measured value when multinodes contact the object, but it cannot give an ex-
is not good (see the blue and red lines). Specifically, as listed in act distribution of the grasping force at each node. To further
Table II, the average relative error is 47.5%, even not considering investigate this issue, a group of simulations is made on the
the cases when grasping forces are less than 0.3 N. However, if compliant finger when contact occurs at different locations. For
the resultant force of the two contacting nodes is considered, each simulation, the distribution of the grasping forces at each
much better coincidence can be found (see the black lines), node varies, but the resultant remains constant. The deformed
where the average relative error reduces to 7.4%. shapes of the finger are shown in Fig. 12. It is found that
Third, two more tests are carried out when two nonadjacent when the two contacting nodes are close to each other [see
nodes contact the sensor module. Results are shown in Fig. 11, Fig. 12(a)], the overall deformation of the finger remains nearly
where (a) presents the result of nodes 18 and 20 and (b) for nodes unchanged although the force distribution varies from (1 N,
16 and 20. Similar to the tests in Fig. 10, comparisons are made 4 N) via (2.5 N, 2.5 N) to (4 N, 1 N). However, the overall
for each of the two nodes and their resultant. It is found that deformations begin to have obvious differences as the force
the overall average relative error is 29.8% if only considering distribution changes if the distance of the two contacting nodes
a single node. Although this error is much smaller than that increases. In addition, the farther the distance is, the larger the
of the adjacent-node contacting tests (47.5%), it is still large. differences become, as shown in Fig. 12(b)–(d). Therefore, if
Nevertheless, the average relative error dramatically decreases external forces are applied to certain nodes that are close to each
to 7.6% when calculating the resultant force of the two nodes. other, various force distributions could cause quite similar fin-
According to the results in Figs. 10 and 11, it is indicated ger shapes. Consequently, our deformation-based force sensing
that the proposed force sensing method can precisely estimate model has poor accuracy in estimating different force distri-
the grasping forces when a single node contacts the object butions in this situation. Nevertheless, the overall deformation
and it also has high accuracy in predicting the resultant force of the finger can well reflect the resultant forces, and thus, the

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

XU et al.: COMPLIANT ADAPTIVE GRIPPER AND ITS INTRINSIC FORCE SENSING METHOD 15

Fig. 12. Deformed shapes of the compliant finger when the force distribution
varies but the resultant remains constant. (a) Contact at nodes 18 and 19. (b)
Contact at nodes 18 and 20. (c) Contact at nodes 17 and 20. (d) Contact at nodes
17 and 21.

proposed model has good accuracy in sensing the total grasping


force.
From the above three groups of tests, it is also found that the
measured and sensed forces of the left finger are quite close
Fig. 13. Deformations of the compliant finger when a group of external forces
to those of the right finger since the two fingers have similar with different directions is applied to the same node. (a) Contact at node 21. (b)
deformations when grasping objects with bilaterally symmetric Contact at node 17.
shapes. In addition, even though there exist relatively large
sensing errors for the two-node contacting tests, the proposed
force sensing method can well predict the variation tendency To further investigate the grasping force direction, simulations
of the grasping forces. For example, as shown in Fig. 11(a), the are performed. As shown in Fig. 13(a) or (b), a group of external
grasping force at node 18 is close to zero and the force at node 20 forces with different directions is applied to the same node. It
gradually grows when the motor angle is less than about 500◦ . is found that although the force directions and amplitudes are
Then, the force at node 18 begins to decrease and the force at 20 different, the deformations of the finger are quite similar. This
begins to increase, and their cross occurs when the motor angle means that even though the finger deformation is determined,
is about 700◦ . there may exist more than one solution of the external force.
It should be noted that in the above three groups of tests, the Thus, the proposed deformation-based force sensing method
grasping force direction at each node is assumed to be known, will not work if both the force direction and amplitude are
which coincides with the axis of the cylindrical force sensor. unknown. Generally, the direction of the grasping force is
Moreover, the force direction is assumed to be unchanged during perpendicular to the contacting surface, which can be easily
the whole grasping process. Meanwhile, the torque caused by detected by vision, while the amplitude of the grasping force
grasping at each node is neglected. With the above assumptions, is usually the unknown parameter to be sensed. Therefore, the
the sensed forces can be directly compared with the measured direction of the grasping force should be given or measured
forces. In fact, the error caused by these assumptions is relatively before applying the proposed force sensing method.
small. On the one hand, the node has a small diameter, which
is only 3 mm in this article, thus the moment arm is small and
D. Applications of the Compliant Gripper
the resulting torque at each node is also negligible. On the other
hand, it is true that the grasping force direction will change In this section, the proposed gripper is applied to perform
with the movement and deformation of the finger. Generally, grasping tasks on real-world objects.
the force direction should be perpendicular to the contacting First, grasping tests are performed on irregular-shape objects.
surface, i.e., along the normal direction of the surface. For our Results are shown in Fig. 14, where the grasping forces at nodes
gripper, the maximum angle between the normal direction and 16–20 are estimated and plotted during the whole grasping
the sensor-axis direction is about 12.5◦ , when the flare angle of process. For irregular-shape objects, it is difficult to integrate
the gripper reaches maximum. Therefore, the maximum error force sensors. Thus, the ground truth of the grasping force is
caused by assuming the grasping force direction to be coincident not available. However, it is still possible to partly verify the
with the sensor-axis direction is only about 1-cos(12.5◦ )=2.4%, force sensing method through analyzing the curves of the sensed
which is negligible. These assumptions are also validated by the forces. According to Fig. 14, for each test, the total sensed
above experiments. grasping forces of the left finger (black lines with circles) are

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

16 IEEE TRANSACTIONS ON ROBOTICS

Fig. 15. Experimental setup of the pick-and-place applications. (a) Overall


arrangement. (b) Group of objects with diverse shapes.

Second, the compliant gripper is attached to the end of a


robotic arm and pick-and-place operations are performed on
various objects with diverse shapes. As shown in Fig. 15, the
proposed gripper and a hand-eye camera are attached to the
end of a robotic arm (UR 5e).2 A group of objects with diverse
shapes is tested in the pick-and-place operations, as shown in
Fig. 15(b). The trajectories of the gripper are planned offline
using the teach-repeat mode of the robotic arm. In the teaching
process, any interference between the camera and the environ-
ment must be avoided. In fact, the current camera arrangement
brings an obvious concern in terms of interference, especially
for applications in other unstructured environments. This issue
will be discussed later in Section VI.
During the whole grasping process, the finger deformations
are captured by the hand-eye camera in real time and then
inputted into the proposed force sensing model to obtain the
grasping forces. It should be noted that the driving motor rotates
continuously at a speed of 30 ◦ /s, and the camera captures the
finger deformation at a speed of 15 FPS. Therefore, we can
obtain a sensed force value when the motor rotates every two
degrees.
For sake of concision, results of the ball and the mouse
Fig. 14. Grasping and sensing experiments on different objects with irregular are given as examples in Fig. 16, where the grasping force
shapes: variation of the grasping forces during the whole grasping process. (a)–
(e) Five ordinary objects with free-form shapes and (e) is an egg that belongs to variation is plotted with the motor rotation angle. Results show
fragile objects. that the grasping forces at all nodes are bilaterally symmetric
(0◦ –900◦ versus 900◦ –0◦ ), which respectively correspond to the
pick and place procedures. Meanwhile, the total grasping forces
very close to that of the right finger (black lines with stars), of the left finger are almost the same with that of the right finger,
which corresponds to the fact that the object is in quasi-static which accords with the fact that the objects are in quasi-static
equilibrium during the grasping process. Moreover, the proposed equilibrium during the grasping process. It is noteworthy that
sensing method allows the contact force at a certain node to when grasping the mouse, the grasping forces at the left node
always being zero. Usually, this zero force can be exactly pre- 19 are always zero since there is no physical contact, which
dicted, such as the right node 17 (red line with stars) in Fig. 14(c) can be well predicted by the proposed force sensing method.
and node 19 (right lines) in Fig. 14(d). Sometimes, the estimation More details of the above pick-and-place experiments can also
of zero force may have small errors (less than 0.3 N), such as be found in the attached video.
the left node 19 (blue line with circles) in Fig. 14(a). It is worth Overall, these experiments indicate that our compliant gripper
noting that the egg in Fig. 14(e) belongs to not only irregular has a good adaptive grasping ability on different objects with
but also fragile objects. Through monitoring the contact forces
during the whole grasping process, the proposed method helps 2 UR 5e is a robotic arm by UNIVERSAL ROBOT Incorporation. https://
a lot to decrease the risk of breaking fragile objects. www.universal-robots.com/products/ur5-robot/

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

XU et al.: COMPLIANT ADAPTIVE GRIPPER AND ITS INTRINSIC FORCE SENSING METHOD 17

time to train the NN model offline, the calculating time of the


trained NNM-based sensing model is negligible. Therefore, the
force sensing rate mainly depends on how fast the vision-based
measurement is. As introduced in Section IV-B, the maximum
measurement rate in this article is 25 Hz considering the con-
straints of the camera and the algorithm. Consequently, the
maximum force sensing frequency herein cannot exceed 25 Hz.
In addition, the grasping speed of the gripper should be limited.
The first reason should be the limited rate of the vision-based
force sensing method as the sensed forces will have delays under
high grasping speed. The second reason may be explained by
the hysteresis of the TPU material. If the grasping speed is
too high, the deformation of the finger might be inadequate (or
delayed), which finally results in force sensing errors. It should
be noted that the maximum allowable grasping speed is difficult
to qualify. In this article, the real-time capacity of the proposed
sensing method is verified by the grasping experiments shown
in Fig. 16 under a force sensing rate of 15 Hz. In the future, the
Fig. 16. Variation of the grasping forces during the pick-and-place process. sensing rate and grasping speed can be improved by increasing
(a) For the ball. (b) For the mouse.
the cameras’ frame rate, accelerating the vision algorithm, and
compensating the material hysteresis.
3) On the number of the unknown force coordinates in the
free-form shapes. Moreover, the gripper can sense the grasp-
sensing model. As explained in Section IV-A, this number should
ing forces during the whole process in real time. Sensing and
be 3n theoretically, where 3 represents the external load at a
controlling the grasping forces play an important role in real
single node including 2 forces and 1 torque, and n represents
applications. It is possible to guarantee a sufficient force for
the number of the nodes that contact objects. However, for real
stably grasping the object, and meanwhile to avoid an excessive
grasping, the torque at each node can be neglected due to the
force that could cause physical damage to the gripper or the
small node diameter and the grasping force direction at each
object.
node is also given or can be measured by vision. Consequently,
the number of unknown force coordinates can be finally reduced
VI. DISCUSSION to n, i.e., the amplitude of the grasping force at the nodes that
Previous sections present the design, modeling, and experi- may contact the object. For example, in the experiments shown
mental validation of the compliant gripper, where its abilities in Figs. 9 and 10, there are respectively one and two nodes
in terms of adaptive grasping and intrinsic force sensing are that have contact, thus n = 1 and 2. For the tests in Fig. 11,
proved. Before the end of this article, we think that the following there is one node at first and then two nodes that have contact,
discussions are necessary to enhance comprehension. thus n = 2. This variation can be well predicted by the curves
1) On the arrangement of the hand-eye camera. In this re- shown in Fig. 11 using the proposed sensing method. To sum
search, the camera is installed on the gripper in a way shown in up, although the proposed model can theoretically have many
Figs. 6(a) and 15, where the image plane of the camera is per- unknown force coordinates, the unknowns are much fewer in real
pendicular to the inserted rods on the fingers. This arrangement grasping. Specifically, the number of unknowns is suggested to
facilitates the vision measurement but brings a potential risk of be equal to that of the contacting nodes. It should be noted that
interference with other objects in the environment. Nevertheless, this requirement is not difficult in real applications since the
it should be noted that the major objective of this article is to contacting locations on the finger can be easily detected by the
propose and validate the intrinsic force sensing method. Further hand-eye camera.
improvement can be made in future work. We plan to cooperate
with some industrial partners and integrate the camera inside
the gripper’s base and capture the finger’s deformations from VII. CONCLUSION
an inclined view. Although this could increase the complexity In this article, a compliant gripper was developed with the
of the vision recognition algorithm, the measuring accuracy can purpose of achieving simultaneous adaptive grasping and in-
be guaranteed since the target remains cooperative. With these trinsic force sensing. To this aim, the traditional fin-ray structure
efforts, the potential interference in real applications could be was employed but modified by inserting a series of rigid rods to
avoided as much as possible. linearize the relationship between the finger deformation and the
2) On the sensing rate and grasping speed. The sensing rate is external force. Meanwhile, the force-to-deformation model was
vital for the real-time performance of the proposed force sensing formulated by the classic FEM. However, due to the difficulty
method and also has an important influence on the grasping in accurately measuring the node rotation, the force sensing
speed in real applications. As explained before, the proposed model could not be obtained by simply inverting the FEM-based
sensing method is based on NNM. Although it takes a long force-to-deformation model but was achieved by neural network

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

18 IEEE TRANSACTIONS ON ROBOTICS

training. Finally, the intrinsic force sensing was realized by [10] V. Lippiello, B. Siciliano, and L. Villani, “A grasping force optimization
measuring the finger deformation using a hand-eye camera in algorithm for multiarm robots with multifingered hands,” IEEE Trans.
Robot., vol. 29, no. 1, pp. 55–67, Feb. 2013.
real time and substituting the calculated node displacements into [11] A. M. Dollar and R. D. Howe, “The highly adaptive SDM hand: Design
the force sensing model. and performance evaluation,” Int. J. Rob. Res., vol. 29, no. 5, pp. 585–597,
Experiments were performed to validate the proposed design 2010.
[12] R. Chen et al., “Bio-inspired shape-adaptive soft robotic grippers aug-
and methods. Results showed that the compliant finger had mented with electroadhesion functionality,” Soft Robot., vol. 6, no. 6,
high sensing accuracy (average error less than 3%) when the pp. 701–712, 2019.
external force was applied at the middle part of the finger but [13] C. C. Kessens and J. P. Desai, “Versatile passive grasping for manipula-
tion,” IEEE/ASME Trans. Mechatron., vol. 21, no. 3, pp. 1293–1302, Jun.
had poor accuracy (error more than 20%) near the finger tip or 2016.
base, especially the tip. Moreover, further tests also indicated [14] C. Li, X. Gu, and H. Ren, “A cable-driven flexible robotic grasper with
that the gripper can precisely predict the total grasping force lego-like modular and reconfigurable joints,” IEEE/ASME Trans. Mecha-
tron., vol. 22, no. 6, pp. 2757–2767, Dec. 2017.
(average error less than 8%) but become less accurate (error [15] P. K. Allen, A. T. Miller, P. Y. Oh, and B. S. Leibowitz, “Integration of
more than 30%) when estimating the force distribution at each vision, force and tactile sensing for grasping,” Int. J. Intell. Mach., vol. 4,
node. Nevertheless, the variation tendency of the sensed forces pp. 129–149, 1999.
[16] M. C. Carrozza, B. Massa, S. Micera, R. Lazzarini, M. Zecca, and
always coincides well with the ground truth. In addition, appli- P. Dario, “The development of a novel prosthetic hand-ongoing research
cations of the gripper were demonstrated in terms of grasping and preliminary results,” IEEE/ASME Trans. Mechatron., vol. 7, no. 2,
free-form objects and performing pick-and-place tasks, where pp. 108–114, Jun. 2002.
[17] M. T. Francomano, D. Accoto, and E. Guglielmelli, “Artificial sense
the adaptive grasping and real-time force sensing abilities were of slip-a review,” IEEE Sens. J., vol. 13, no. 7, pp. 2489–2498,
further proved. Jul. 2013.
It should be noted that the proposed method required knowing [18] A. Ikeda, Y. Kurita, J. Ueda, Y. Matsumoto, and T. Ogasawara, “Grip force
control for an elastic finger using vision-based incipient slip feedback,”
the grasping force direction to perform force sensing. Mean- in Proc. IEEE/RSJ Int. Conf. Intell. Robots Syst., Sendai, Japan, 2004,
while, to improve accuracy, it is better to minimize the number pp. 810–815.
of unknown forces in the sensing model, i.e., detecting the nodes [19] Festo, “The parallel gripper DHPS.” [Online]. Available: https://www.
festo.com/cat/de_de/products_010804, Accessed on: 2020.
that possibly contact the object, and making the forces at the [20] Robotiq, “The 2F series grippers.” [Online]. Available: https://robotiq.
rest nodes zero. These requirements will not bring additional com/products/2f85-140-adaptive-robot-gripper, Accessed on: 2020.
difficulty since both the force directions and the contacting nodes [21] Robotiq, “The 3-finger adaptive robot gripper.” [Online]. Available:
https://robotiq.com/products/3-finger-adaptive-robot-gripper, Accessed
can be easily recognized by vision using the hand-eye camera. on: 2020.
To sum up, the gripper presented in this research realized not [22] Z. Chen, N. Y. Lii, T. Wimboeck, S. Fan, and H. Liu, “Experimental
only the adaptive grasping but also the intrinsic force sensing, evaluation of cartesian and joint impedance control with adaptive friction
compensation for the dexterous robot hand DLR-HIT II,” Int. J. Humanoid
which enabled a number of new applications from industrial Robot., vol. 8, no. 4, pp. 649–671, 2011.
to daily life, where adaptivity and force control were required. [23] F. Ilievski, A. D. Mazzeo, R. F. Shepherd, X. Chen, and G. M. Whitesides,
While our current gripper is still in the early stage, some aspects “Soft robotics for chemists,” Angew. Chem. Int. Ed., vol. 50, no. 8,
pp. 1890–1895, 2011.
can further be improved in the future, such as the finger structure [24] K. C. Galloway et al., “Soft robotic grippers for biological sampling on
optimization, the hand-eye camera arrangement, and the vision deep reefs,” Soft Robot., vol. 3, no. 1, pp. 23–33, 2016.
algorithm enhancement. [25] J. Amend, N. Cheng, S. Fakhouri, and B. Culley, “Soft robotics com-
mercialization: Jamming grippers from research to product,” Soft Robot.,
vol. 3, no. 4, pp. 213–222, 2016.
[26] I. A. Anderson, T. A. Gisby, T. G. McKay, B. M. O’Brien, and E. P. Calius,
REFERENCES “Multi-functional dielectric elastomer artificial muscles for soft and smart
machines,” J. Appl. Phys., vol. 112, no. 4, 2012, Art. no. 041101.
[1] R. Hodson, “How robots are grasping the art of gripping,” Nature, vol. 557, [27] M. Behl, K. Kratz, J. Zotzmann, U. Nöchel, and A. Lendlein, “Reversible
no. 7706, pp. S23–S23, 2018. bidirectional shape-memory polymers,” Adv. Mater., vol. 25, no. 32,
[2] A. M. Dollar, L. P. Jentoft, J. H. Gao, and R. D. Howe, “Contact sensing pp. 4466–4469, 2013.
and grasping performance of compliant hands,” Auton. Robot., vol. 28, [28] Festo, “The MultiChoiceGripper.” [Online]. Available: https://www.festo.
no. 1, p. 65–75, 2010. com/group/en/cms/10221.htm, Accessed on: 2020.
[3] D. Song, C. H. Ek, K. Huebner, and D. Kragic, “Task-based robot grasp [29] K. Telegenov, Y. Tlegenov, and A. Shintemirov, “A low-cost open-source
planning using probabilistic inference,” IEEE Trans. Robot., vol. 31, no. 3, 3-D-printed three-finger gripper platform for research and educational
pp. 546–561, Jun. 2015. purposes,” IEEE Access, vol. 3, pp. 638–647, 2015.
[4] K. B. Shimoga, “Robot grasp synthesis algorithms: A survey,” Int. J. Rob. [30] S. B. Backus and A. M. Dollar, “An adaptive three-fingered prismatic
Res., vol. 15, no. 3, pp. 230–266, 1996. gripper with passive rotational joints,” IEEE Robot. Autom. Lett., vol. 1,
[5] J. Bohg, A. Morales, T. Asfour, and D. Kragic, “Data-driven grasp no. 2, pp. 668–675, Jul. 2016.
synthesis—A survey,” IEEE Trans. Robot., vol. 30, no. 2, pp. 289–309, [31] E. Martin, A. L. Desbiens, T. Laliberté, and C. Gosselin, “SARAH hand
Apr. 2014. used for space operation on STVF robot,” in Proc. Int. Conf. Intell. Manip.
[6] E. Mattar, “A survey of bio-inspired robotics hands implementation: New Grasping. Genova, Italy, 2004, pp. 279–284.
directions in dexterous manipulation,” Robot. Auton. Syst, vol. 61, no. 5, [32] Shadow-Robot-Company, “The Shadow Hand.” [Online]. Avail-
pp. 517–544, 2013. able: https://www.shadowrobot.com/products/dexterous-hand/,Accessed
[7] U. Gupta, L. Qin, Y. Wang, H. Godaba, and J. Zhu, “Soft robots based on: 2020.
on dielectric elastomer actuators: A review,” Smart Mater. Struct., vol. 28, [33] J. Butterfaß, M. Grebenstein, H. Liu, and G. Hirzinger, “DLR-Hand II:
no. 10, 2019, Art. no. 103002. Next generation of a dextrous robot hand,” in Proc. IEEE Int. Conf. Robot.
[8] J. Shintake, V. Cacucciolo, D. Floreano, and H. Shea, “Soft robotic Autom., Seoul, South Korea, 2001, pp. 109–114.
grippers,” Adv. Mater., vol. 30, no. 29, 2018, Art. no. 1707035. [34] C. Della Santina, C. Piazza, G. Grioli, M. G. Catalano, and A. Bicchi,
[9] F. Cini, V. Ortenzi, P. Corke, and M. Controzzi, “On the choice of grasp “Toward dexterous manipulation with augmented adaptive synergies: The
type and location when handing over an object,” Sci. Robot., vol. 4, no. 27, Pisa/IIT SoftHand 2,” IEEE Trans. Robot., vol. 34, no. 5, pp. 1141–1156,
2019, Art. no. eaau 9757. Oct. 2018.

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

XU et al.: COMPLIANT ADAPTIVE GRIPPER AND ITS INTRINSIC FORCE SENSING METHOD 19

[35] M. Zhu, Y. Mori, T. Wakayama, A. Wada, and S. Kawamura, “A fully multi- [60] K. Elgeneidy, N. Lohse, and M. Jackson, “Bending angle prediction and
material three-dimensional printed soft gripper with variable stiffness for control of soft pneumatic actuators with embedded flex sensors—A data-
robust grasping,” Soft Robot., vol. 6, no. 4, pp. 507–519, 2019. driven approach,” Mechatronics, vol. 50, pp. 234–247, 2018.
[36] R. Deimel and O. Brock, “A novel type of compliant and underactuated [61] D.-H. Kim et al., “Materials and noncoplanar mesh designs for integrated
robotic hand for dexterous grasping,” Int. J. Rob. Res., vol. 35, no. 1/3, circuits with linear elastic responses to extreme mechanical deformations,”
pp. 161–185, 2016. Proc. Nat. Acad. Sci. USA, vol. 105, no. 48, pp. 18 675–18 680, 2008.
[37] B. Mosadegh et al., “Pneumatic networks for soft robotics that actuate [62] M. Amjadi, K.-U. Kyung, I. Park, and M. Sitti, “Stretchable, skin-
rapidly,” Adv. Funct. Mater., vol. 24, no. 15, pp. 2163–2170, 2014. mountable, and wearable strain sensors and their potential applications: A
[38] R. V. Martinez, A. C. Glavan, C. Keplinger, A. I. Oyetibo, and G. M. review,” Adv. Funct. Mater., vol. 26, no. 11, pp. 1678–1698, 2016.
Whitesides, “Soft actuators and robots that are resistant to mechanical [63] C. Larson et al., “Highly stretchable electroluminescent skin for optical
damage,” Adv. Funct. Mater., vol. 24, no. 20, pp. 3003–3010, 2014. signaling and tactile sensing,” Science, vol. 351, no. 6277, pp. 1071–1074,
[39] N. R. Sinatra, C. B. Teeple, D. M. Vogt, K. K. Parker, D. F. Gruber, and 2016.
R. J. Wood, “Ultragentle manipulation of delicate structures using a soft [64] H. Zhao, K. O’Brien, S. Li, and R. F. Shepherd, “Optoelectronically
robotic gripper,” Sci. Robot., vol. 4, no. 33, 2019, Art. no. eaax 5425. innervated soft prosthetic hand via stretchable optical waveguides,” Sci.
[40] E. Brown et al., “Universal robotic gripper based on the jamming Robot., vol. 1, no. 1, 2016, Art. no. eaai 7529.
of granular material,” Proc. Natl. Acad. Sci. USA, vol. 107, no. 44, [65] S. Yang and Q. Xu, “A review on actuation and sensing techniques
pp. 18809–18814, 2010. for MEMS-based microgrippers,” J. Micro-Bio Robot., vol. 13, no. 1-4,
[41] J. R. Amend, E. Brown, N. Rodenberg, H. M. Jaeger, and H. Lipson, pp. 1–14, 2017.
“A positive pressure universal gripper based on the jamming of granular [66] K. Xu and N. Simaan, “An investigation of the intrinsic force sensing
material,” IEEE Trans. Robot., vol. 28, no. 2, pp. 341–350, Apr. 2012. capabilities of continuum robots,” IEEE Trans. Robot., vol. 24, no. 3,
[42] Y. Wang, U. Gupta, N. Parulekar, and J. Zhu, “A soft gripper of fast speed pp. 576–587, Jun. 2008.
and low energy consumption,” Sci. China Technol. Sci., vol. 62, no. 1, [67] V. A. Aloi and D. C. Rucker, “Estimating loads along elastic rods,” in Proc.
pp. 31–38, 2019. IEEE Int. Conf. Robot. Autom., Montreal, QC, Canada, 2019, pp. 2867–
[43] Q. Ge, A. H. Sakhaei, H. Lee, C. K. Dunn, N. X. Fang, and M. L. Dunn, 2873.
“Multimaterial 4D printing with tailorable shape memory polymers,” Sci. [68] H. Yuan, P. W. Y. Chiu, and Z. Li, “Shape-reconstruction-based force
Rep., vol. 6, 2016, Art. no. 31110. sensing method for continuum surgical robots with large deformation,”
[44] C.-C. Lan, C.-M. Lin, and C.-H. Fan, “A self-sensing microgripper module IEEE Robot. Autom. Lett., vol. 2, no. 4, pp. 1972–1979, Oct. 2017.
with wide handling ranges,” IEEE/ASME Trans. Mechatronics, vol. 16, [69] R. Xu, A. Yurkewich, and R. V. Patel, “Curvature, torsion, and force
no. 1, pp. 141–150, Feb. 2011. sensing in continuum robots using helically wrapped FBG sensors,” IEEE
[45] D. Petkovic, N. D. Pavlovic, S. Shamshirband, and N. B. Anuar, “Devel- Robot. Autom. Lett., vol. 1, no. 2, pp. 1052–1059, Jul. 2016.
opment of a new type of passively adaptive compliant gripper,” Ind. Rob., [70] A. Bajo and N. Simaan, “Hybrid motion/force control of multi-backbone
vol. 40, no. 6, pp. 610–623, 2013. continuum robots,” Int. J. Rob. Res., vol. 35, no. 4, pp. 422–434, 2016.
[46] C.-H. Liu, G.-F. Huang, C.-H. Chiu, and T.-Y. Pai, “Topology synthesis [71] Z. Zhang, A. Petit, J. Dequidt, and C. Duriez, “Calibration and external
and optimal design of an adaptive compliant gripper to maximize output force sensing for soft robots using an RGB-D camera,” IEEE Robot. Autom.
displacement,” J. Intell. Robot Syst., vol. 90, no. 3-4, pp. 287–304, 2018. Lett., vol. 4, no. 3, pp. 2356–2363, Jul. 2019.
[47] M. H. Ali, A. Zhanabayev, S. Khamzhin, and K. Mussin, “Biologically [72] W. Xu, H. Zhang, N. Zheng, and H. Yuan, “Design and experiments of a
inspired gripper based on the fin ray effect,” in Proc. Int. Conf. Control, compliant adaptive grasper based on fish fin structure,” in Proc. IEEE Int.
Autom. Robot., Beijing, China, 2019, pp. 865–869. Conf. Robot. Biomim.. Kuala Lumpur, Malaysia, 2018, pp. 293–298.
[48] W. Crooks, S. Rozen-Levy, B. Trimmer, C. Rogers, and W. Messner, [73] L. Kölle and O. Schwarz, “Bionic forceps for the handling of sensitive
“Passive gripper inspired by Manduca sexta and the fin ray effect,” Int. tissue,” Curr. Directions Biomed. Eng., vol. 2, no. 1, pp. 91–93, 2016.
J. Adv. Robot. Syst., vol. 14, no. 4, pp. 1–7, 2017. [74] W. Crooks, “Soft robotic manipulators inspired by the fin ray effect,” Dept.
[49] M. Anwar, T. A. Khawli, I. Hussain, D. Gan, and F. Renda, “Modeling and Mech. Eng., Ph.D. dissertation, Tufts Univ., Medford, MA, USA, 2017.
prototyping of a soft closed-chain modular gripper,” Ind. Robot, vol. 46, [75] Covestro, “Mechanical Properties of TPU.” [Online]. Available:
no. 1, pp. 135–145, 2019. https://solutions.covestro.com/en/highlights/articles/theme/product-
[50] W. Crooks, G. Vukasin, M. O’Sullivan, W. Messner, and C. Rogers, “Fin technology/mechanical-properties-tpu, Accessed on: 2020.
ray effect inspired soft robotic gripper: From the robosoft grand challenge [76] T. J. Hughes, The Finite Element Method: Linear Static and Dynamic Finite
toward optimization,” Front. Robot. AI, vol. 3, no. 70, pp. 1–9, 2016. Element Analysis. North Chelmsford, MA, USA: Courier Corporation,
[51] H. Kawasaki, T. Komatsu, and K. Uchiyama, “Dexterous anthropomorphic 2012.
robot hand with distributed tactile sensor: Gifu hand II,” IEEE/ASME [77] M. H. Hassoun, Fundamentals of Artificial Neural Networks. Cambridge,
Trans. Mechatronics, vol. 7, no. 3, pp. 296–303, Sep. 2002. MA, USA: MIT Press, 1995.
[52] S. C. Jacobsen, J. E. Wood, D. Knutti, and K. B. Biggers, “The UTAH/MIT [78] R. C. Gonzalez, R. E. Woods, and S. L. Eddins, Digital Image Processing
dextrous hand: Work in progress,” Int. J. Robot. Res., vol. 3, no. 4, Using MATLAB. New Delhi, India: Pearson Education India, 2004.
pp. 21–50, 1984.
[53] H. Yousef, M. Boukallel, and K. Althoefer, “Tactile sensing for dexterous
in-hand manipulation in robotics—A review,” Sens. Actuator A. Phys.,
vol. 167, no. 2, pp. 171–187, 2011.
[54] Y. She, C. Li, J. Cleary, and H.-J. Su, “Design and fabrication of a soft
robotic hand with embedded actuators and sensors,” J. Mech. Robot, vol. 7,
no. 2, 2015, Art. no. 021007.
[55] J. Hashizume, T. M. Huh, S. A. Suresh, and M. R. Cutkosky, “Capacitive
sensing for a gripper with gecko-inspired adhesive film,” IEEE Robot.
Autom. Lett., vol. 4, no. 2, pp. 677–683, Apr. 2019.
[56] A. Saudabayev and H. A. Varol, “Sensors for robotic hands: A. Survey of
state of the art,” IEEE Access, vol. 3, pp. 1765–1782, 2015.
[57] S. Aoyagi, M. Suzuki, T. Morita, T. Takahashi, and H. Takise, “Bellows Wenfu Xu (Senior Member, IEEE) received the Ph.D.
suction cup equipped with force sensing ability by direct coating thin film degree in control science and engineering from the
resistor for vacuum type robotic hand,” IEEE/ASME Trans. Mechatron., Harbin Institute of Technology, Harbin, China, in
vol. 25, no. 5, pp. 2501–2512, Oct. 2020. 2007.
[58] K. Koyama, H. Hasegawa, Y. Suzuki, A. Ming, and M. Shimojo, “Pre- He was a Research Associate with the Depart-
shaping for various objects by the robot hand equipped with resistor ment of Mechanical Engineering and Automation,
network structure proximity sensors,” in Proc. IEEE/RSJ Int. Conf. Intell. Chinese University of Hong Kong, Hong Kong. He
Robots Syst., Tokyo, Japan, 2013, pp. 4027–4033. is currently a Professor with the Harbin Institute of
[59] A. Koivikko, E. S. Raei, M. Mosallaei, M. Mäntysalo, and V. Sariola, Technology, Shenzhen, China. His research interests
“Screen-printed curvature sensors for soft robots,” IEEE Sens. J., vol. 18, include bionic robots, space robots, flexible robots,
no. 1, pp. 223–230, Jan. 2018. and intelligent control.

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

20 IEEE TRANSACTIONS ON ROBOTICS

Heng Zhang received the B.E. degree in mechanical Bin Liang (Member, IEEE) received the B.S. degree
design manufacturing and automation from Yanshan in control engineering from Northwestern Polytech-
University, Hebei, China, in 2017. He is currently nical University, Xi’an, China, in 1991, and the Ph.D
working toward the Postdoctorate degree with the degree in precision instrument and mechanology from
Harbin Institute of Technology, Shenzhen, China. Tsinghua University, Beijing, China, in 1994.
His research interests include soft robotics and He is a Professor with the Department of Au-
machine vision. tomation, Tsinghua University, Beijing. His research
interests include space robots, manipulators, and in-
telligent control.

Han Yuan (Member, IEEE) received the Ph.D. degree


in mechanical engineering from Institut National des
Sciences Appliques de Rennes, Rennes, France, in
2015.
He was a Research Scientist with IFMA, Clermont-
Ferrand, France, from 2015 to 2016 and a Postdoc-
toral Fellow with the Chinese University of Hong
Kong, Hong Kong, from 2016 to 2017. He is currently
an Assistant Professor with the Harbin Institute of
Technology, Shenzhen, China. His research interests
include soft robots and cable-driven manipulators.

Authorized licensed use limited to: San Francisco State Univ. Downloaded on June 20,2021 at 06:18:28 UTC from IEEE Xplore. Restrictions apply.

You might also like