This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access

Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/[Link] Number

Innovative Human-Like Dual Robotic Hand

Mechatronic Design and Its Chess-Playing
Experiment
Ming-Bao Huang and Han-Pang Huang, Member, IEEE
Department of Mechanical Engineering, National Taiwan University, Taipei 10617, TAIWAN (R.O.C.)

Corresponding author: Han-Pang Huang (e-mail: hanpang@[Link]).

This work was partially supported by HIWIN Technologies Corp., R.O.C., under grant number 104-S-A13.

ABSTRACT This paper describes how a standard HI-tech WINner (HIWIN) Corporation articulated six
degrees of freedom (DOFs) industrial robot arm and a National Taiwan University five-finger robotic hand
were equipped with additional hardware and software to enable the resulting machine to play chess with a
human chess player. For automatic interaction with the human player, moves were recorded by a webcam
and automatically analyzed. No manual (keyboard) input was necessary. The chess-playing robotic system
has been provided. An innovative humanoid robotic hand with 12 DOFs and 19 joints were designed, and
distributed tactile sensor arrays with 376 detecting points on its surface were developed. The hand can
communicate with the external through controller area networks bus. For performance evaluation of the
designed robotic hand, we analyzed the workspace, intersection volume, and manipulability. First, several
anatomical analyses were conducted prior to a decision on the kinematic design. Optimization procedures
were then developed in this study to improve the parameters of the design and structure of the mechanism.
The concepts of a series elastic actuator (SEA) and an under-actuated mechanism were also employed to
give the robotic hand a compliant property and high dexterity; hence, a humanoid robotic hand was devised.
Since the robotic hand is as small and dexterous as a human hand, it can be conveniently used for a wide
range of applications. Finally, the system has been proved to be executable in a complex environment.

INDEX TERMS Chess-playing robotic system, cable-driven, compliant, multi-fingered robotic hands,
series elastic actuator (SEA), under-actuated mechanical hand.

I. INTRODUCTION situations nor perform fine manipulation. To increase the

applications of a robotic hand, such as for medical robots and

I N recent years, numerous multi-fingered robotic hands

have been developed. The design concept of robotic
hands can be classified into two categories. One is to design
service robots, it has become important to design a robotic
hand that has sophisticated capabilities similar to those of the
human hand [2]. Indeed, the performance of robotic hands
a robotic hand that is easily controlled for a reliable grasp. may be enhanced by more completely imitating the human
The other approach is to design a humanoid robotic hand hand, as an ideal model. But a human-like robotic hand
that can perform not only grasping but also fine usually has complex and expensive mechanisms. Hence, this
manipulation. The mechanics of the human hand have been study focused on the use of design methods and concepts to
studied as a model for mechanical devices since the late simplify the mechanism of the robotic hand and control
18th century [1]. Initially, the focus was primarily on system while still having most of the features of the human
prostheses and telemanipulation systems. As the demand hand to accomplish most of the same functions. Some studies
for assembly production has increased, reliable industrial [3], [4], [5] have used similar approaches, but this paper
mechanical grippers have become very important devices in proposes some different ones. Moreover, the robust control
many industrial applications. method and compact hardware used in this study were crucial
However, these types of robotic hands were built for to the design. Finally, the robotic hand system was designed
specific environments, and they neither adapt to unstructured to demonstrate the possibilities of a simplified mechanism

2169-3536 © 2019 IEEE. Translations and content mining are permitted for academic research only.
VOLUME XX, 2019 Personal use is also permitted, but republication/redistribution requires IEEE permission. 1
See [Link] for more information.
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

that can achieve the functions of the human hand. The two
main focuses of this study were mentioned in the following.
First, the control input was designed to resist unknown
disturbances. By analyzing the object dynamics model
without knowing its corresponding inertia and Coriolis
matrices of the object, the designed contact force can
accomplish the predefined object trajectory, and the
manipulated object is stable and robust under certain
degrees of uncertainty. We proposed a simple rule to check
the stability of the grasped object and select the appropriate
contact force. Furthermore, we proposed a strategy for
grasping and manipulating unknown objects.
Second, a slippage avoidance strategy was constructed.
While distributing the designed contact force to each
fingertip, every contact force normal to its contact surface
has to be minimized and constrained within the friction
cone, and slippage can be easily avoided. A simple
parameter was designed to confine the resultant contact
forces to lie in the interior of their friction cones while
proceeding the optimization problems.
This paper discusses the mechanism design of a robotic
hand according to a practical design flow. Before the design
concept and requirements were decided, hand anatomy,
including skeleton and finger motion, were analyzed and then FIGURE 1. Grasping conditions.

the kinematic configuration was depicted. Two optimization

problems were confronted, and the optimal parameters were
selected by solving the optimization problems. The
mechanism was designed according to the optimal results
and the design requirements; then the mechanism with the
chess-playing system was developed.
This paper is organized as follows: Section II describes the
analysis of anatomy and kinematic design; Section III
describes the optimal design for digit length and angular rate;
Section IV presents an overview of the National Taiwan
University (NTU) five-finger robotic hand mechanism and
(a) Abduction/Adduction. (b) Finger flexing motion.
the hardware and software controller architecture; Section V
presents the experimental results; and Section VI addresses
the conclusions and possibilities for future studies.

II. ANALYSIS OF ANATOMY AND KINEMATIC DESIGN

A. HAND ANATOMY
As shown in Fig. 1 [6], except for the thumb, fingers are
composed of three joints and three digital bones, also called
phalanges. The three joints are the distal interphalangeal
(DIP) joint, proximal interphalangeal (PIP) joint, and
metacarpophalangeal (MCP or MP) joint. The thumb has
only two phalanges and two joints between the phalanges,
but it also has an extra joint, called the basilar joint, which
provides the thumb holds the functions that could be
provided by the MCP joint for the other fingers. If both the
metacarpal bone and the basilar joint are considered to be a
(c) Thumb motion.
part of the thumb, then all five fingers will have similar
structures. FIGURE 2. Finger motions of human hand [6], [7].

2 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

TABLE I
STATE OF THE ART IN THE FIELD OF ARTIFICIAL GRASPING SYSTEMS: A COMPARATIVE ANALYSIS
Size hand/ No. of No. of Fingertip Under-actuated
Hand No. of actuators Weight (g)
human hand fingers joints force (N) method
Developed 1.2 +
5 19 12 built-in 1,065 6 Tendon
robotic hand controllers
Human hand
1 5 22 38 built-in + ext. about 400 >300 Tendon
[6], [7]
DLR/HIT
1.5 + controllers 5 20 15 built-in about 1,500 10 Tendon
hand II [8]
Gifu hand III
1.6 + controllers 5 20 16 built-in 1,400 1.8 Mechanism
[9]
Shadow hand
1.2 + controllers 5 21 20 ext. 4,000 x Tendon
[10], [11]
NTU hand
1.6 + controllers 5 17 17 built-in 1,569 x Mechanism
[12]
Utah/MIT
2 + controllers 4 16 32 ext. x 31 Tendon
hand [13]
Barrett hand Mechanical
1.2 3 8 4 built-in 1,180 20
[14], [15] breakaway clutch

In addition, the ratio of phalanges is determined by

measuring several human hands, as shown in Fig. 3, which
the ratio of phalanges is decided as (9 : 6 : 5) from proximal
phalange to distal phalange. By observing our hands, it
becomes apparent that the joints of the hand have compliance,
and some studies [16], [17], [18] have also demonstrated that
the stiffness of finger joints varies with posture and force.
This section highlights the related prior research on under-
actuated hands. Table I shows a comparative analysis of
some of the artificial hands discussed in the body of recent
FIGURE 3. Measuring the ratio of phalanges.
robotics literature and summarizes some famous under-
actuated hands from studies published so far. The parameters
Analyses of finger motions and hand postures play an
selected by those researchers served to compare different
important role in related mechanism design with anatomy.
hand concepts, and given for each prototype are the number
Fig. 2(a) shows the abduction/adduction motion. The middle
of DOFs, the number of actuators, the size, the weight, and
finger and the thumb do not move, and muscles make the
the resulting force. In comparison between the artificial hand
index finger, ring finger, and small finger adduct and abduct.
and the natural human hand, which represents the ideal
The flexing motion of a single finger is shown in Fig. 2(b).
model, it is obvious that current available knowledge and
The rotating range of the MCP joint can be measured, as
technology produce hand prototypes are far from mimicking
shown in the upper image of Fig. 2(b). The bottom image of
a similar performance as the natural hand.
Fig. 2(b) shows how, on flexing the DIP joint, a ligament
Despite this, recent achievements and signs of progress
called the retinacular ligament becomes taut and pulls the PIP
indicate a positive trend in making better hand concept and
joint into flexion. Similarly, on extending the PIP joint, the
prototypes. However, due to technical limitations, it is
DIP joint is pulled by the retinacular ligament into extension,
necessary to make some compromise in the ability to imitate
too [7]. Hence, it can be concluded that the PIP joint and the
the human hand, which requires defining priority
DIP joint rotate together in a given angular velocity ratio. Fig.
specification according to the intended application such as,
2(c) shows a clear diagram of the motion of the thumb.
for example, when a greater resulting force is obtained, it
Although the DIP joint and the MCP joint of the thumb are
causes heavier weight and larger size. Based on the results of
similar to the DIP joint and the PIP joint of the rest of the
this analysis, the concept of the innovative hand is presented.
fingers, the basilar joint has a special rotating range because
According to the prior work, a five fingers artificial hand has
the basilar joint has two DOFs leading to anteposition
a total of 12 DOFs and 19 controlled joints, as shown in Fig.
(opposition), as shown on the right hand side of Fig. 2(c).
4(b).

VOLUME XX, 2019 3

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

B. KINEMATIC DESIGN III. OPTIMAL DESIGN FOR DIGIT LENGTH AND

ANGULAR RATE
Fig. 4 shows the kinematic configuration of the robotic hand. Before commencing the detailed design, some parameters of
Each of the joints is identified with two numbers and the robotic hand had to be determined. Given that the fingers
described as symbol J n , m , where the first suffix indicates the are a modularized design, it was reasonable to take just one
finger number and the second suffix indicates the joint finger into consideration. Thus, an optimization problem was
number of the finger [19]. This kinematic configuration is modeled to solve the parameters of the mechanism.
arranged according to the proposed motion analysis and
design concept. There are two bars between joint J n ,2 and
joint J n ,3 in each finger, which represent the under-actuated
mechanism that provides a coupling between the two joints.
Therefore, there is an angular velocity ratio of joint J n ,2 to
joint J n ,3 . This ratio will be discussed later.

J 2,3 J 3,3 J 4,3 J 5,3

FIGURE 5. Schematic of geometry for a robotic finger.
J 2,2 J 3,2 J 4,2 J 5,2

J 2,1 J 3,1
The geometry for the robotic finger design is shown in Fig.
J 4,1 J 5,1
J1,3 5. The kinematic parameters that had to be determined were
the phalange lengths ( L1 , L2 , L3 ) and the angular velocity
J1,2 J 2,0 J 4,0
J 5,0
ratio (R) of the DIP joint implants to the PIP joint implants.
J1,1 The design was concerned with weight, payload during
J1,0 power grasp, and manipulability. Those issues have to be
solved in order to realize efficient, flexible, and stable grasp
(a) Kinematic structure.
(b) Tactile sensor arrays [20] and operation, and hence a multi-objective optimization problem
hydrocolloid dressing. will be formed. After the problem was formulated, the solver
FIGURE 4. NTU five-finger robotic hand. provided by MATLAB Optimization Toolbox™ was used to
solve the problem, and the parameters were obtained [21].
Joints J 2,0 , J 4,0 , and J 5,0 are arranged to accomplish
abduction and adduction. Since these three joints are also A. OPTIMALITY CRITERIA
coupled, the index finger, ring finger, and little finger will 1) WEIGHT
open and close together. Joint J1,0 plays the role of the The weight of the robotic hand must be as light as possible.
basilar joint, which allows the robotic hand to make the Since all components are necessary and cannot be omitted,
anteposition (opposition) motion. Consequently, the the volume and the material have significant influence over
innovative humanoid robotic hand has a total of 19 DOFs, the weight of the robotic hand. Here, the emphasis is on the
seven of which are passive, leaving 12 independent DOFs. volume.
The member joints for each independent degree of freedom
(DOF) are listed in Table II along with the arrangement of
2) PAYLOAD DURING POWER GRASP
the simplest essential set. Based on this design, a compact
and functional robotic hand was developed.
1 3
TABLE II
THE INDEPENDENT DOFS AND MEMBER JOINTS N1
1
N3
4
2
DOF Member joints DOF Member joints 2
N4
1 J1,0 7 J 3,2 , J 3,3 N2

2 J1,1 8 J 4,1 W
N5 5
Motor
3 J1,2 , J1,3 9 J 4,2 , J 4,3 (a) Power grasp of a human
(b) Free-body diagram of power grasp.
hand.
4 J 2,1 10 J 5,1
FIGURE 6. Schematic of power grasp.
5 J 2,2 , J 2,3 11 J 5,2 , J 5,3
6 J 3,1 12 J 2,0 , J 4,0 , J 5,0 Fig. 6(a) shows the power grasp postured by a human hand.
A power grasp is characterized by multiple points of contact
between the object being grasped and the surfaces of the
fingers and palm. Because the grasp is highly stable as a

4 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

result of form closure [22], the designed robotic hand has Sampling point in workspace
80
also to be powerful during the power grasp. (To grasp an
object by hand requires constraining the object completely by
using the fingers and/or the palm of the hand. A fixture is in 60

form closure if any possible movement of the object is

resisted by a non-penetration constraint.) To compute the 40
payload of the robotic hand, a free-body diagram of the

Y-axis (mm)
human grasp was drawn, as shown in Fig. 6(b). The motors
applied torque directly at the joints marked  1 ,  3 , and  4 .
20

Because of the coupled motion between the PIP joint and the
DIP joint,  2 has a relationship with  1 of 0

1
2  1 , (1) -20
1 R
where R is the angular velocity ratio, and  4 has the same -40
relation to  5 . The contact forces N1 , N2 ,  N5 were 0 10 20 30 40
X-axis (mm)
50 60 70 80

derived from the torques and the phalange lengths.

The payload can be written as FIGURE 7. Evaluation of manipulability at specific points.
5
W    ( Ni cos(i )   Ni sin(i )) , (2) B. PROBLEM FORMULATION
i 1
The result is that the ratio of the phalanges was determined to
where  i is the angle between the contact force N i and the be (9 : 6 : 5) from the proximal phalange to the distal
vertical, as shown in Fig. 6(b), and  is the friction phalange. There was a constraint on the lengths of the robotic
coefficient. The second term on the right side is friction. fingers, and the parameters L1 , L2 , and L3 were rewritten as
Except for the torques  1 ,  3 ,  4 , and the friction coefficient,
the variables are influenced by the phalange lengths and the  L1  9t

angular velocity ratio.  L2  6t . (5)
 L  5t
 3
3) MANIPULABILITY
Manipulability evaluates the ability to move and apply forces Therefore, the design variables are (t , R) , where t is the
in arbitrary directions [23], [24]. It has been used in task length of the robotic finger optimized parameters set of
planning, control, and robot design [25], and it can be 70 mm to 100 mm in order to be similar to the length of
evaluated by the manipulability measure. In general, human hands, as shown in Fig. 7. To have similar human
manipulability M can be defined as [26], [27], [28], [29]: motion, R is the angular velocity ratio of 0.5 to 1.5.
Furthermore, a human-like outline and movement are
M  det( JJ T ) , (3) preferred, so constraints are imposed upon the design
variables:
where det( JJ T ) denotes the determinant of the matrix JJ T .
Therefore, if J is a square matrix, M will become 70  20t  100

M  det( J ) . Manipulability can also be expressed as 0.5  R  1.5
M  12  m  1 2   m , (4)  . (6)
3.5  t  5.0 (length rang)
where 1  2    m  0 are eigenvalues of JJ T , and 
1   2     m  0 are singular values of the m  n 0.5  R  1.5 (motion constraint)
Jacobian matrix J. Thus, maximum manipulability of the Finally, the mathematical model of this optimization
robotic hand was expected. problem is summarized as follows:
Next, the manipulability of the robotic finger was taken
into account. Due to the local property of manipulability, the min F1 (t , R)  9t  6t  5t  20t
overall behavior cannot be guaranteed by evaluating only one max F2 (t , R) 
point in the workspace. Therefore, the average value of 5

manipulability at different points in the workspace was   ( N i (t , R) cos(i (t , R))   N i (t , R) sin(i (t , R))) (7)
i 1
computed, and these points are displayed as asterisks in Fig. n
7. This allows the manipulator to exhibit a rich set of control  M j (t , R)
j 1
properties that are responsive to the world. max F3 (t , R)  M (t , R) 
n

VOLUME XX, 2019 5

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

 3.5  t  5.0
subject to  ,
0.5  R  1.5
where M j denotes the manipulability evaluated at a certain
point in the workspace of the fingertip, as shown in Fig. 7.
The parameter n indicates the number of total sampling
points, and the friction coefficient  is set as 0.3.

C. MULTI-OBJECTIVE OPTIMIZATION
The three objective functions in (7) have different tendencies
to optimization, and the optimized parameter set of FIGURE 9. Open-loop positioning response of models with springs.
(t , R)  (4.32 , 0.82) was selected by using the multi-
objective optimizing method, minimax. When controlling position, the shorter settling time was
The phalange lengths ( L1 , L2 , L3 )  (38.9 , 25.9 , 21.6) better. The open-loop positioning was applied to the model.
were then decided, and the angle relation 3  0.822 was The position and angular velocity responses of the normal
determined as well. model are shown in Fig. 8. In addition, Fig. 9 shows the
responses of the model with the spring, where the finger with
D. OPTIMAL DESIGN FOR SPRING POSITION the spring in the DIP joint settled down more quickly.
As described in Section II, it is apparent that human fingers
have compliance, which is an important capability for 2) SHOCK ABSORPTION PROPERTY
protecting bones. Therefore, many mechanisms with
compliance [30], [31], such as a series elastic actuator (SEA)
[32], [33], [34] and the force sensing compliant actuator
(FSC) [35], were designed. The idea is to integrate flexible
components between the motor and the end effector, or
motor and ground. The parameters of the phalanges and the
joints were set according to the result achieved in Section III
C.
With mechanical compliance, impact shocks to the robotic
finger are dampened by the flexible components, protecting FIGURE 10. Shock absorption properties given an impact force on the
the motor gearbox from damage. These types of shocks are fingertip. The endured torques of the motors were measured.
commonly used in robotic hands and seriously limit its
longevity. The passive compliance also allows a finger to It is expected that the motor endures less shock when it
better conform to an object through fine-grained adjustments receives an impact. Therefore, the shock absorption property
of posture. Moreover, by measuring the deflection and was tested, and a 10 N impact was applied to the fingertip as
Hooke’s law, the torque applied by the motor can be known. shown in Fig. 10. Then we measured the torque applied to
However, the compliant mechanism also has a disadvantage. the two motors on the robotic finger. Fig. 11 shows the result
Because of the low mechanical impedance, the response of the normal model. The motor that actuating the PIP joint
speed of the system was adversely affected. For all the endured approximately 700 N-mm torque, and the motor
reasons stated above, some criteria were addressed. actuating the MCP joint and DIP joint received about 350 N-
mm torque.
1) RESPONSE TIME

FIGURE 11. Impact on motors of normal model.

FIGURE 8. Open-loop positioning response of the normal model.

6 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

Finally, the torques are listed in Table III. Based on the

properties, this model was considered to be best.

E. ANALYSIS OF ROBOTIC HAND GRASPING AND

CONTROL
1) GRASP STABILIZER

The main purpose of grasping is to explore, restrain, or

control an object [36]. Here, grasp refers to how the hand
contacts the object along with the dynamic properties of both
the hand and the object. Given a grasp, we can study the
FIGURE 12. Impact on motors of models with springs.
force distribution and stability of a configuration. To
manipulate and explore the object, the grasp may be changed
The torque applied to the motors of the models with by, for instance, finger gaiting or rolling to achieve the goal,
springs is shown in Fig. 12. Although the finger oscillated for and therefore, the dynamics and stability in such evolution
a longer time, some concerns arose with the torque, which should be considered. Another problem is the motion from
was the lowest. With mechanical compliance, the robot hand the initial configuration of the hand to the grasped object,
would obtain high stability against the impact of collision. which includes the positioning of the contacts and the force
control.
3) CORRELATION BETWEEN OUTPUT TORQUE AND In grasping, the total contact wrench λ can be split into
SPRING DEFORMATION two parts: the manipulation force λ manipulation that realizes the
desired object motion and the internal force λ stabilization that
TABLE III
MODEL CHARACTERISTICS does not affect the object’s motion but maintains a stable
grasp on the object, specifically,
Model
Settling time 0.8 (sec) λ  λ manipulation  λ stabilization (8)
.
Motor
Endured 200 (N-mm) Because λ stabilization is the force that guarantees the stability of
1
torque on a grasp, it should be always considered as the primary
Motor
motors 100 (N-mm) condition for grasping.
2
That means when an SEA-driven We summarized the required condition of λ stabilization as
Correlation robot smacks into something, it follows. First, λ stabilization should be in the internal friction.
between output just bumps off without causing Second, as λ stabilization should not influence the object’s
torque and harm—and that is the affordable motion, λ stabilization should be in the null space of grasping
deformation. way to implement safety factor matrix G . Third, λ stabilization should be in the space of
into the design. wrenches that the fingers can control, meaning in the range
of the hand’s Jacobian matrix J . Combining the above three
conditions, we obtained (9) as the conclusion.
λ stabilization  int( FC) (G) (J ) (9)

2) GRASP STRATEGY

However, (9) may be too complicated for practical

applications. To simplify it for online planning, we make the
following assumptions to facilitate quick calculations.
 ASSUMPTIONS:
1) The grasp is a planar.
FIGURE 13. Deformation and torque of model.
2) The grasped object is a rigid body.
To measure the torque by the deflection of the spring, the 3) The contacts between the grasped object and fingertips
deflection must have a similar trend as the actual torque. Fig. are point contacts with friction.
13 shows the spring deformation and the output torque model. 4) The center of mass is on the grasp plane.
The red curve is the given output torque; the other colored 5) The kinematic model of the multi-fingered hand is
curve is the spring deformation. precise.

VOLUME XX, 2019 7

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

The first assumption, which confines all contact points to

the grasp plane, simplifies the grasp matrix. On the other
hand, since the MCP joint of each finger of the NTU five-
finger robotic hand is equipped with an SEA, the small joint
angle measurement error can be tolerated.
In force and moment balance, there are at least one force
opposed with the other forces; for example, in the NTU five-
finger robotic hand, the thumb is always in a position
opposite to the other fingers. Since the index finger, ring
(a) (b)
finger, and little finger are coupled, we assume abduction and
adduction angles are zero; that is, all fingers are straight. As a FIGURE 14. Two conditions for three-fingered force closure.

result, we can simplify the hand Jacobian matrix as well.

The three-finger case: Because a robotic hand with three
fingers satisfies the necessary condition for force and
moment balance, we proposed two conditions to verify
whether the forces were in the friction cones in planar
grasping. Let f1 , f 2 , f 3 be the three forces exerted by the
three fingers, respectively. The grasp forms force closure if
and only if one of the following conditions is satisfied:
1) f1 , f 2 , f3 are coplanar and intersect at one point.
(a) (b)
2) f1 , f 2 , f3 are coplanar, parallel, and the two forces are
FIGURE 15. Two conditions for cases with more than three fingers.
opposite in directions.
That is, if fi  int( Fci ), for i  1, 2, 3 ( f i is strictly in the
3) CONTACT FORCE GENERATOR
internal friction cone Fci ), the contact points C1 , C2 , C3 can
construct a force-closure grasp.
Deciding the contact force is an important issue for grasping.
In Fig. 14, the yellow part is the object, and the red parts
In recent works, the multi-fingered force planning included
are the friction cones Fci . With the NTU five-finger robotic
nonlinear planning, linear planning, and intelligent planning
hand, this three-fingered force-closure condition is used for [37], [38], [39]. However, in actual applications, it is difficult
the contact points of thumb, index finger, and middle finger. to use those algorithms because of costs, time, and
Cases with more than three fingers: For situations with
insufficient information. As a result, we are only concerned
more than three contact points Ci (i  1, 2,3,..., n) , there exist with the contact force and contact position in order to
at least three points, and it is force closure if and only if the calculate the contact force compensation.
following conditions are satisfied: Fig. 16 shows the relationship of contact points. Cg is the
1) Ci , C j , i, j {1, 2,3,..., n} and i  j , along with a geometric centroid point of all contact points, which can be
line L, satisfying the normal vector of contact points calculated as
ni , n j , have different projection vectors on L and
1 n
Cg   Ci , (10)
Lp  Fc  Fc . n i 1
i j

2) k {1,..., n}\{i , j} ,  k   , where  is the friction where n is the number of contact points, Ci for i=1,2,3, … ,n
coefficient and  k is the angle between normal vector are the contact points, li for i=1,2,3, … ,n, are the vectors
from the contact points to Cg which is based on the unit
nk and the plane defined by the composition of
vector
Ci , C j , Ck .
Cg  Ci
3) l {1,..., n}\{i , j} , L  Fcl   . lˆi  (11)
|| Cg  Ci ||
For the NTU five-finger robotic hand, we chose the
contact points of thumb and middle finger to be the two and satisfies the magnitude relation
points to fulfill the condition of line L. Then we checked n
whether the index finger, ring finger, and little finger  ai lˆi  0, i  1, 2,3,..., n . (12)
i 1
satisfied line L in the friction cone of each finger. If it did, we
could assure that at least three fingers were in the force When grasping, we will obtain the desired joint torque as
closure. If not, it would become unstable, as shown in Fig. 15.
 d i  ai J F fˆi , i  1, 2,3,..., n , (13)
i

8 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

and its magnitude depends on the property of the grasped motor (FAULHABER 1512 012 SR with IE2-8 encoder and
object. Since the kinematics of robotic hands do not always 324:1 gearbox) actuated the coupled DIP/PIP joints of the
have the DOFs to give the force f i , whose direction is the finger. Thus, the DIP and PIP joints of the robotic hand were
same as li , some directions will be neglected. Note that ai is designed to be under-actuated, and their rotating motions
the gain constant, and J F is the Jacobian matrix. were coupled with each other.
i
When a disturbance is exerted on the grasped object or the The rotating ranges of these two joints were determined, as
robotic hand, the geometric centroid point will be changed to shown in Fig. 18, where the distal phalange was expected to
Cg , as shown in Fig. 17. In order to maintain a stable grasp, be opposite to the proximal phalange and it satisfied the
 ci will be changed in accordance with the new geometric constraint 3  0.822 . The lengths of the phalanges were
centroid point. the same as the optimized values, and the parameters of the
phalanges and joints were according to the result from
Section III C. Furthermore, a cable-driven mechanism was
used for the transmission, because unlike gear-driven
transmissions, a cable-driven transmission does not have any
backlash problem [32]. The cable paths are shown as the red
line in Fig. 18(a) and Fig. 18(b).

(a) Cable path.

FIGURE 16. The relationship of contact points.

(b) Top view of the DIP and PIP joints.

81

99

(c) Maximum rotating angles of the DIP and PIP joints.

FIGURE 17. Contact point change.

IV. MECHANISM

A. UNDER-ACTUATED FINGER
The term “under-actuated” means that the mechanism has (d) Maximum rotating angles of the MCP Joint.

fewer actuators than the DOF [40], [41]. Because this type of FIGURE 18. The under-actuated robotic finger.
mechanism reduces the number of actuators and the
complexity, this simplified mechanism can be easier to Given that the robotic hand is the most important part of
control and is much lighter and more affordable than its fully any interactions with the physical environment, it should be
actuated counterparts. compliant with safety requirements. Hence, an SEA with a
Since the human PIP and DIP joints work together in steel cable and springs for the robotic fingers was designed
many cases, we designed PIP and DIP joints to be for this study. Further, because dexterity is one of the key
mechanically coupled and driven by one actuator. A DC features of robotic hands, achieving the dexterity of a human

VOLUME XX, 2019 9

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

hand was an important goal. The design and rotation range of finger bases: thumb, middle finger, and others. If one of the
the fingers shown in Fig. 18(c) and Fig. 18(d) are similar to fingers is broken, it can be easily repaired using modular
those of human fingers, and the robotic fingers can curl spares, thereby reducing inventory.
completely up to 270 degrees. Each MCP joint is actuated by The CAD view of the whole hand model that made use of
one DC motor (FAULHABER 1516 012 SR with IE2-512 the size of a human hand for design is shown in Fig. 20. The
encoder and 546:1 gearbox) located in the palm, which pulls frame of the robotic hand was constructed of aluminum alloy
a cable along a linear slider connection to the motor through to minimize the weight, which is about 1,065 g for each hand.
a pulley. Moreover, for achieving the abduction and adduction motion,
the linkage mechanism was devised, as shown in Fig. 20(b).
Since the index finger, ring finger, and little finger are
coupled, the spread range of motion for the joints is given as
much as 15o for the joint (abduction/adduction of base
joint). The DC motor (FAULHABER 1512 012 SR with IE2-
8 encoder and 324:1 gearbox) in the metacarpal phalange of
the finger actuates the vertical DOF of the MCP joint.
To achieve stable grasping, precise manipulation, and
modularization, a thumb was designed by introducing an
additional joint with 1 DOF into the finger module and a
little modification of the finger module. Fig. 21 displays the
basilar joint of the thumb structure. The thumb can abduct
and adduct by moving from the palmar position to the lateral
position, and vice versa, as depicted in Fig. 21(a). The DC
motor (FAULHABER 1512 012 SR with IE2-8 encoder and
324:1 gearbox) at the base of the thumb actuates the
FIGURE 19. Three different finger designs. carpometacarpal joint (CMC joint, or CM joint) to obtain an
approximation of the abduction/adduction motion. The
anteposition (opposition) motion was accomplished by the
basilar joint mechanism, as shown in Fig. 21(b). Thus, one of
the design concepts was realized.

B. SERIES ELASTIC ACTUATOR (SEA) MECHANISM

According to the results of the analysis, the MCP joint
implants had to be compliant. Therefore, a compliant
mechanism based on the SEA [33] was designed, as
displayed in Fig. 18. The cable-driven transmission was
employed again; the blue arrows in Fig. 22(a) and the red line
(a) (b)
in Fig. 22(b) are the cable paths. The cable terminals were
FIGURE 20. CAD model of the hand. fixed at the proximal phalange of the robotic hand, as shown
in Fig. 22(b). From Fig. 22(b), the relationship between the
spring deformation x and the T predicted via Hooke’s law
can be derived as
2T  k x , (14)
where k is the stiffness of the spring. From Fig. 22(a) and Fig.
22(b), the schematic drawing was sketched and is shown in
Fig. 22(c). The output torque at the joint is
(b) Anteposition (opposition)
(a) Basilar joint of the thumb
structure.
motion of the humanoid robotic
hand.
  r (T1  T2 ) , (15)

FIGURE 21. Basilar joint of thumb. where T1 and T2 are the forces transmitted to the cable by
the pulley and r is the radius of the pulley. Finally, by
To reduce the cost and maintenance of the robotic hand, combining (14) and (15), the relationship between torque 
the robotic fingers were designed using a modular concept. and the spring deformation x can be derived as
In other words, the main parts of the five fingers are identical
except for the bases, which are connected to the hand’s palm.   rk x . (16)
As shown in Fig. 19, there are three different designs for the

10 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

As a result, the output torque can be measured with a D. WORKSPACE

potentiometer equipped on the joint. The MCP joint is driven The dexterous workspace for each fingertip was derived from
by an actuator that can change the related positions of the the forward kinematics of the robotic hand with the link
two links connected by that joint. The proprioceptors are parameters and the rotating ranges of the joints. The
encoder and potentiometer sensors, which measure both dexterous workspace for each fingertip was obtained and is
kinematic and dynamic parameters of the robot. Based on shown in Fig. 24, where the density of the mark shows the
those measurements, the control system activates the dexterity of each fingertip, and the red markers indicate the
actuators to generate torques so that the articulated origin of the world coordinate frame. If the contact points
mechanical structure performs the desired motion. between the fingers and an object are located in the
workspace, the robotic hand can perform dexterous
manipulations as required.

(a) Cable path of MCP joint.

T T

x

Potentiometer
(b) Top view of SEA.

T1
FIGURE 24. The workspace of the robotic hand.

E. SPECIFICATIONS OF THE NTU FIVE-FINGER

T2 ROBOTIC HAND REAL-TIME CONTROL SYSTEM
(c) Schematic drawing. New circuits were designed for the multiprocessor control
FIGURE 22. SEA mechanism. structure to improve flexibility for advanced developments.
All related assembly codes were also embedded into
microprocessors to handle the coordination of the robot
C. WHOLE HAND hand’s controller modules, and a control was used for
After the design parameters and kinematic configuration tolerant force disturbance.
were determined, a detailed mechanical design of the Microchip Technology, Inc., an American manufacturer of
manipulator was also done. Based on the results from the microcontrollers, memory, and analog semiconductors, was
previously designed hand anatomy and finger mechanism, used for the digital signal processing (DSP) in the controller
the whole robotic hand was devised. Its dimensions are area network (CAN) bus. In the application area, we enabled
shown in Fig. 23, and the size of this robotic hand is similar an ADC channel to measure the motor current. The QEI
to that of a human hand. module (encoder module) provided information on the angle
of the rotor so that the joint torque could be measured
through the modified joint torque observer. We can update
the properties of the dynamics on a PC, which communicates
with the microcontroller unit using USB-to-CAN bus
adapters. The distributed 16-bit dsPIC30F4011 controllers
and their 32-bit master PIC32MX795F512H were all state
machines that constantly checked their states and executed
commands. The dsPIC30F4011 was the low-level control,
and PIC32MX795F512H was the USB-to-CAN bus adapter.
The CAN communication system was integrated into the
FIGURE 23. Overall profile of the NTU five-finger dual robotic hand. palm. The state machine architecture of the C30 controllers

VOLUME XX, 2019 11

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

and the C32 controllers are illustrated in Fig. 25 and Fig. 26,
respectively.
The multiprocessor control system consists of two parts:
the central communication module and the joint control
module. Hence, this control structure can distribute the entire
computation load across several control modules. The
architecture of the joint control module was designed to
perform a faster computation. A joint control module was
able to control two joints, as shown in Fig. 27.
In the real-time control system, the clock time influences
the response time of the control system. In C++ on a laptop
computer, multi-thread programming was used to send
control commands and read feedback signals through the
USB ports in parallel.

FIGURE 28. Real-time control system.

With respect to the hardware system and the real-time

control system, the important issue was to upgrade the
bandwidth of the upper-level controller (PC-based) [42], [43],
[44].
Fig. 28 shows the real-time control system. To achieve the
overall control framework, a high-level controller was
programmed on a PC using C++. This included computation
FIGURE 25. C30 – NJM2670 motor control module.
of the dynamics of the compensator about the feedback states
of the manipulator’s system, detection of the entry/leave
point, computation of the admittance compensator, and
compensation for the impedance effect on each joint torque
command. The high-level controller was programmed using
C++ with a sampling frequency of 200 Hz, while the
sampling frequency for the low-level controller was 5 kHz.
In such a large system, reducing the number of wires is
vital for effective setup and maintenance of the robot. To
achieve an optimal solution for real-time control and the
FIGURE 26. The USB-to-CAN bus adapter module. wiring system, we connected the robot’s limbs and sensors to
the central control laptop computer using USB-to-CAN bus
adapters. The limbs and sensors were connected in several
CAN buses, and all nodes were connected with four wires
Sensor
PC Joint Control Module Signal (VCC, GND, CANH, and CANL) in each CAN bus. Two
wires were digital power lines, and two wires were for CAN
Trajectory Planning Joint Control Module
Sensor
Signal
bus communication. This reduced the number of required
wires as compared to connecting the devices directly using
UART Sensor
RS-232, I 2 C bus (inter-integrated circuit bus), USB, or other
Joint Control Module Signal methods. The laptop computer can process higher-level
Central algorithms such as image processing, path planning, and
Communication Sensor
Module
Joint Control Module Signal decision making; thus, it was defined as the higher-level
controller. The microcontrollers on the driver boards were
defined as lower-level controllers, which were used mainly to
FIGURE 27. The multiprocessor control structure of the robotic hand. control the motors and read signals from the sensors.

12 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

F. NTU FIVE-FINGER ROBOTIC HAND robotic hand was made of aluminum alloy and nylon, and the
SPECIFICATIONS weight of the hand can be further reduced if any other
material is available.
TABLE IV
SPECIFICATIONS OF THE ROBOTIC HAND
Max. Max.
Max.
continuous continuous
torque
torque angular
output
output velocity
(mNm)
(mNm) (rpm)
Proximal
interphalangeal joint 51.52 79 15
of four fingers
Metacarpophalangeal
300 830 22
joint of four fingers
Basilar joint of the
51.52 79 15
thumb FIGURE 29. End effector frames of thumb, index, and middle fingers
Max. output force at (grasping the chess piece).
6N
fingertip
Curl range of a
270 degrees
finger
1,065 g (include electronic
Weight
printed circuit boards)
Number of actuators 12
Number of joints 19
Nominal voltage 12 Volts
Current consumption 0~3A

FIGURE 30. Thumb fingertip stress.

For the quasi-static situation, the force at the end effector is
balanced by joint torques. The force Jacobian matrix J F is
the transpose of velocity Jacobian matrix, as
  J T F  JF F , (17)

where F  61 denotes the force (Cartesian force-moment

vector) acting on the end effector, and   n1 denotes
torques at n joints.
In order to calculate the feasible force that is allowed to act
on the end effector, the specifications used for the hand
motors were shown in Table IV, where based on force FIGURE 31. Index fingertip stress.
Jacobians, the stresses of the fingertips can be calculated.
According to the joint angles, the Jacobian matrix can be
constructed. From (17), the force on the end effector frame
can be calculated according to the possible torques at joints
J1,1 , J1,2 , J 2,1 , J 2,2 , J 3,1 , and J 3,2 .
In terms of the joint torques, the stress range that each
fingertip was able to exert on an object was analyzed. From
Fig. 29 through Fig. 32, the stresses along xi and yi of the
i th fingertip grew greater, while the joint torques increased
and reached their maximum values as the torques reached
FIGURE 32. Middle fingertip stress.
their limitations shown in Table IV.
The fundamental finger of the robotic hand has two
G. CONTROL OF THE HAND–ARM SYSTEM
independent DOFs: the PIP joint implants and the MCP joint
This section discusses the control of the hand–arm system.
implants. The qualifications of all PIP joint implants are the
Since the robot arm in our application was designed to move
same as those of all MCP joints implants. The specifications
to the desired position along a scheduled trajectory, we used
of the robotic hand are listed in Table IV. The prototype
position control for the robot arm. By contrast, the robotic

VOLUME XX, 2019 13

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

hand, used for grasping, requires exerting force on an object, For automatic interaction with the human player, these
and thus, force control is necessary. The whole hand–arm moves were recorded by webcam (Logitech C930e, Logitech,
system can be separated into four parts: planning, high-level Newark, CA, USA) and automatically analyzed. No manual
control, local-level control, and hardware. (keyboard) input was necessary, and the chess-playing
Before the robot could begin to execute the operation, we robotic system was provided. The learning algorithm was
needed to give it information obtained from vision-based employed in the system. This research presents a simplified
planning, such as the position of the grasped object, the grasp synthesis [47] and control approach for practical
orientation for the hand to grasp, the required force for applications [20], [48], [49], [50], [51]. The main idea was to
stabilization, and the manipulation trajectory. In the simplify grasping conditions and achieve compliant behavior
computer, hand–arm planning generated reference in grasping so that the hand–arm system can use frictional
trajectories for both the arm and the hand according to the forces to grasp an uncertain object efficiently and without
position and orientation of the grasped object. In turn, the breaking the object.
arm and hand delivered information on the actual angles for
each joint to the hand–arm planning system using forward
kinematics to determine the Cartesian position of the arm and
hand.
The high-level controller dealt with the robot kinematics
by using online inverse kinematics to determine the joint
angle with inverse dynamics [45], generating joint torque for
the hand and dealing with hand–arm coordination. The local
controller, which was embedded in the robot, provided
reliable and stable control for the single joint angles in order
to track the command.
The control operations for the hand–arm system are shown
in Fig. 33. X Ar is the reference trajectory of the arm, X Hr is
the reference trajectory of the hand, q Am is the joint angles of
the arm, τ Hm is the joint torque of the hand, q Arm is the real
joint angles of the arm, q Hand is the real joint angles of the
hand, and f ext is the external force sensed by the tactile
sensors.

FIGURE 34. Standard HIWIN Technologies Corporation articulated six

DOFs industrial dual robot arms and NTU five-finger dual robotic hands.

FIGURE 33. The control system of the hand–arm system.

V. EXPERIMENTAL RESULTS
 METHODS FOR GRASPING VARIOUS OBJECTS
Computer Vision-Based Chess Gaming Capabilities of
Intelligent Dual Robotic Hand–Arm System FIGURE 35. Examples of object grasping with the chess piece. This
indicates that our robotic hand reproduced human grasping postures
faithfully.
This section describes how a standard HI-tech WINner
(HIWIN) Technologies Corporation articulated six DOFs To simplify the grasping condition, the contacts were
industrial robot arm (RT605) [46] and the NTU five-finger restricted to lying on the same plane, and the contact model
robotic hand were equipped with additional hardware and was assumed to be a point of contact with friction. Under this
software to enable the resulting machine to play chess with a assumption, when the sum of forces and moments balance,
human. We combined the robot arm and robot hand to there is at least one force opposed to the other forces. For
facilitate planning and real-time control, as shown in Fig. 34. instance, the thumb of the NTU five-finger robotic hand

14 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

should always be positioned against the other fingers, and a The robot versus robot experiment is illustrated in Fig. 36.
normal plane should be chosen to align with the direction of First, the robot picked up the white bishop (Fig. 36[a]) and
the desired manipulation. The initial planning first considered released it at the end point (Fig. 36[b]). When the hand had
the contact positions for the thumb, index, and middle fingers fully released the chess piece, the robot moved to the clock
so that these three fingers touched the object (e.g., grasping and pressed it (Fig. 36[c]). When the other robot received the
the chess piece), as shown in Fig. 35. The line segments of signal from the clock, it immediately tracked the moving
the three contact points were included in the friction cone white bishop and found the best move, which it executed (Fig.
[52]. Using this relationship iteratively, each finger’s desired 36[d] and Fig. 36[e]) before pressing the clock at its side.
contact position on the plane was then chosen. For the finger The robot versus human match is shown in Fig. 37. The
where the workspace was outside the selected plane, a robot was the first player, and the tactics was decided by the
desired position was chosen to minimize the closest distance chess engine. As can be seen from Fig. 37(a), Fig. 37(b), and
to the convex formed by the other fingers. Based on this Fig. 37(c), the robot moved the white pawn forward two
grasping model, the desired force was chosen such that the grids. When the robot pressed the clock (Fig. 37[d]), the
sum of forces exerted zero wrench on the plane formed by human opponent decided on a move and executed it. The
the convex hull of the fingertips. Thus, while retaining game continued until one of the players won.
compliance with the plane, the robotic hand can grasp an The chess-playing robot discussed in this paper can detect
object by friction forces in the normal direction. the chessboard, track moves, and find the best next move.
Next, we carried out grasping experiments to verify that The robot plays chess with a human opponent or another
the developed device satisfied the functions of robotic hands. robot by using visual feedback, and for the tracking step, this
To grasp an object, the thumb was first rotated so that it was study proposed a method that provides a more robust result
opposite to other four fingers, which were initially fully by combining the chessboard grid state and image-
extended, where only the distal and intermediate phalanges subtraction information to eliminate the weakness of using
moved, but the proximal phalanx was fixed. At last, our only one method. The robot was tested in different
robotic hand can be able to grasp objects just like a human environments, and the tracking method was provided to
hand because it has similar mechanisms. overcome noise in the environment. Finally, the chess-
The technique described above was then extended to a full playing robot was debuted at the 2015 International Robot
hand–arm system. The robot arm was controlled with Exhibition (iRex), Tokyo Big Sight, Tokyo, Japan, where it
position control, and the robotic hand was controlled with demonstrated the ability to play chess with humans (Fig. 36,
simplified grasping with the chess strategy. Before the Fig. 37; see video for supplementary materials,
operation was executed, the hand–arm planning system [Link]
generated a reference trajectory for the arm and hand
according to the position and orientation of the chess piece to
be grasped.

(a) (b)

(a) (b)

(c) (d)
(c) (d)

(e) (f)
(e) (f)
FIGURE 36. Robot vs. robot.
FIGURE 37. Robot vs. human.

VOLUME XX, 2019 15

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

an adult human hand, so it has a wide range of applications.

In addition, its flexibility and computational power makes it
meet future requirements for autonomous manipulation.
Therefore, it has been confirmed that our robotic hand is
widely versatile. In the future, we will validate this hand as a
prosthetic hand by controlling it with myoelectric signals, as
(a) Three-jaw chuck or palmar
(b) Lateral pinch. shown in Fig. 38. Afterward, it can be applied to humanoid
tripod pinch.
robots and master-slave robot systems, and with our
developed humanoid robot, NINO [50], which learned from
human demonstration to introduce itself using Taiwanese
Sign Language [49], [54].

REFERENCES
[1] Y. Matsuoka, “The mechanisms in a humanoid robot
(c) Hook grasp. (d) Power grasp. hand,” Int. J. Auton. Robots, vol. 4, no. 2, pp. 199–209,
1997.
[2] A. Saudabayev and H. A. Varol, “Sensors for robotic
hands: A survey of state of the art,” IEEE Access, vol. 3,
pp. 1765–1782, Sep. 2015.
[3] N. E. N. Rodríguez, G. Carbone, and M. Ceccarelli,
“Optimal design of driving mechanism in a 1-DOF
anthropomorphic finger,” Int. J. Mechanism & Machine
(e) Centralized grip or Theory, vol. 41, no. 8, pp. 897–911, 2006.
(f) Free hand (rest). [4] M. J. Farrugia and M. A. Saliba, “Optimisation of
indexing.
anthropomorphic robot hand design through human
FIGURE 38. Real-time control of a multi-fingered robotic hand using manual dexterity testing,” Int. J. Verein Deutscher
electromyography signals (a wearable MYO™ armband [53]). Ingenieure-Berichte (VDI-Berichte), vol. 1956, pp. 147–
161, 2006.
VI. CONCLUSIONS [5] J. H. Lee, B.-J. Yi, S.-R. Oh, and I. H. Suh, “Optimal
design of a five-bar finger with redundant actuation,” in
This article gave an overview of the design solutions and
Proc. IEEE Int. Conf. Robot. Automat. (ICRA), Leuven,
implementation of an autonomous chess-playing robot, Belgium, 1998, vol. 3, pp. 2068–2074.
focusing on the dexterity and compliance of grasping control [6] Edmonds Orthopedic Center. Edmonds, WA, USA,
by implementing a robot hand–arm system that is composed (2003, Apr. 14). [Online]. Available:
[Link]
of a HIWIN articulated six DOFs industrial robot arm and =6bb97628783df0ff0eb7a551c12fa2c2&disp_type=topi
the NTU five-finger robotic hand. We paid particular c_detail&area=11&topic_id=699047371a7cfea5c5064b
attention to mechanical and electrical robustness throughout 8bdb7d68b5
[7] A. M. R. Agur and M. J. Lee, “Upper limb,” in Grant’s
the design, assembly, and testing of the robotic hand and Atlas of Anatomy, 10th ed., Baltimore, Maryland:
proposed a simplification of the conditions for the chess Williams & Wilkins, pp. 480–513, 1999.
pieces to facilitate efficient grasp planning. The goal of the [8] H. Liu, K. Wu, P. Meusel, N. Seitz, G. Hirzinger, M. H.
Jin, Y. W. Liu, S. W. Fan, T. Lan, and Z. P. Chen,
project was to show that the possibilities of robot intelligence, “Multisensory five-finger dexterous hand: The
and due to the above-mentioned characteristics, our robotic DLR/HIT Hand II,” in Proc. IEEE/RSJ Int. Conf. Intel.
hand is effective for use as an end effector of robots carrying Robots Syst. (IROS), Nice, France, 2008, pp. 3692–3697.
out skillful grasping and manipulation, substituting for [9] T. Mouri, H. Kawasaki, and S. Ito, “Unknown object
grasping strategy imitating human grasping reflex for
humans. As for the mechanical design, the structure with anthropomorphic robot hand,” J. Advanced Mech. Des.,
spring-based compliant hinges is proved to be quite Syst., & Manufacturing (JAMDSM), vol. 1, no. 1, pp. 1–
satisfactory in terms of structural simplification, ease of 11, 2007.
[10] Shadow Robot Company Ltd., Highgate, London,
manufacturing, and convenient assembly. Construction of a England, 1997. [Online]. Available:
five-fingered master-slave robot system is one possible future [Link]
work. [11] Learning Dexterity. (2018, Jul. 30). [Online]. Available:
[Link]
In the work detailed in this paper, an innovative humanoid [12] L.-R. Lin and H.-P. Huang, “NTU hand: A new design
robotic hand was successfully designed. The presentation of dexterous hands,” J. Mech. Des. (Trans. of the Amer.
was focused on an efficient control architecture, which, Society of Mech. Engs. (ASME)), vol. 120, no. 2, pp.
282–292, Jun. 1998.
through analysis of hand anatomy and the use of optimization
[13] S. C. Jacobsen, E. K. Iversen, D. F. Knutti, R. T.
processes, allowed the robotic hand, using a simplified Johnson, and K. B. Biggers, “Design of the Utah/M.I.T.
mechanism, to accomplish functions of human hands, dextrous hand,” in Proc. IEEE Int. Conf. Robot. Automat.
including fine manipulation. Moreover, this humanoid (ICRA), San Francisco, CA, USA, 1986, pp. 1520–1532.
[14] W. Townsend, “The BarrettHand grasper –
robotic hand has been equipped with compliant components programmably flexible part handling and assembly,”
to enhance its functionality. The robotic hand, including the Industrial Robot: An International Journal, vol. 27, no.
complete controller system, is almost the same size as that of 3, pp. 181–188, 2000.

16 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

[15] Barrett Tech., Inc., Newton, MA, USA [Online]. Humanoid Robots (Humanoids), Tsukuba, Japan, 2005,
Available: [Link] pp. 327–332.
[16] K. Akazawa, T. E. Milner, and R. B. Stein, “Modulation [33] E. Torres-Jara, “Sensitive manipulation,” Ph.D.
of reflex EMG and stiffness in response to stretch of dissertation, Dept. Elec. Eng. & Com. Sci.,
human finger muscle,” J. Neurophysiology, vol. 49, no. Massachusetts Inst. Tech., Cambridge, MA, USA, 2007.
1, pp. 16–27, Jan. 1983. [34] E. Torres-Jara and J. Banks, “A simple and scalable
[17] Q. Meng, H. Wang, P. Li, L. Wang, and Z. He, force actuator,” in Proc. 35th Int. Symp. on Robot., 2004.
“Dexterous underwater robot hand: HEU hand II,” in [35] A. Edsinger-Gonzales and J. Weber, “Domo: A force
Proc. of IEEE Int. Conf. Mechatronics Automat. (ICMA), sensing humanoid robot for manipulation research,” in
Luoyang, China, 2006, pp. 1477–1482. Proc. 4th IEEE/RAS Int. Conf. Humanoid Robots
[18] P. Beckerle, M. Bianchi, C. Castellini, and G. Salvietti, (Humanoids), Santa Monica, CA, USA, vol. 1, 2004, pp.
“Mechatronic designs for a robotic hand to explore 273–291.
human body experience and sensory-motor skills: a [36] A. Bicchi and V. Kumar, “Robotic grasping and contact:
Delphi study,” Advanced Robot. (AR), vol. 32, no. 12, A review,” in Proc. IEEE Int. Conf. Robot. Automat.
pp. 670–680, Jun. 2018. (ICRA), San Francisco, CA, USA, vol. 1, 2000, pp. 348–
[19] L. Sciavicco and B. Siciliano, “Differential kinematics 353.
and statics,” in Modelling and Control of Robot [37] M. Mukadam, C.-A. Cheng, X. Yan, and B. Boots,
Manipulators, 2nd ed., London, England: Springer- “Approximately optimal continuous-time motion
Verlag, 2000, pp. 87–96. planning and control via probabilistic inference,” Robot.,
[20] W.-Y. Lee, M.-B. Huang, and H.-P. Huang, “Learning Com. Sci., Cornell Univ. Library, Ithaca, NY, arXiv
robot tactile sensing of object for shape recognition preprint arXiv:1702.07335, Feb. 27, 2017 & in Proc.
using multi-fingered robot hands,” in Proc. 26th IEEE IEEE Int. Conf. Robot. Automat. (ICRA), Singapore,
Int. Symp. Robot Human Interactive Comm. (RO-MAN), 2017, pp. 664–671.
Lisbon, Portugal, 2017, pp. 1311–1316. [38] J. Dong, M. Mukadam, F. Dellaert, and B. Boots,
[21] M. Malvezzi, G. Gioioso, G. Salvietti, and D. “Motion planning as probabilistic inference using
Prattichizzo, “SynGrasp: A MATLAB toolbox for Gaussian processes and factor graphs,” in Proc. Robot.:
underactuated and compliant hands,” IEEE Robot. & Sci. Syst. Conf. XII (RSS-2016), Ann Arbor, Michigan,
Automat. Magazine, vol. 22, no. 4, pp. 52–68, Dec. 2015. USA, 2016.
[22] K. Lakshminarayana, “Mechanics of form closure,” [39] R. M. Murray, Z. Li, and S. S. Sastry, “Multifingered
ASME Paper, New York, USA, Tech. Rep. vol. 78- hand kinematics,” in A Mathematical Introduction to
DET-32, pp. 2–8, 1978. Robotic Manipulation, 1st ed., Boca Raton, FL, USA:
[23] C. Ott, A. Albu-Schäffer, A. Kugi, S. Stramigioli, and G. CRC Press (Inc.), Mar. 1994, pp. 211–240.
Hirzinger, “A passivity based Cartesian impedance [40] T. Laliberté and C. M. Gosselin, “Simulation and design
controller for flexible joint robots—Part I: Torque of underactuated mechanical hands,” Mech. & Mach.
feedback and gravity compensation,” in Proc. IEEE Int. Theory, vol. 33, no. 1/2, pp. 39–57, Jan.–Feb. 1998.
Conf. Robot. Automat. (ICRA), New Orleans, LA, USA, [41] A. M. Dollar and R. D. Howe, “Joint coupling design of
vol. 3, 2004, pp. 2659–2665. underactuated grippers,” in Proc. 30th Annual ASME
[24] A. Albu-Schäffer, C. Ott, and G. Hirzinger, “A passivity Mechanisms & Robot. Conf., Parts A & B & Int. Des.
based Cartesian impedance controller for flexible joint Eng. Tech. Conf. (IDETC) & Computers & Inf. in Eng.
robots—Part II: Full state feedback, impedance design Conf. (CIE), Philadelphia, PA, vol. 2, 2006, pp. 903–911.
and experiments,” in Proc. IEEE Int. Conf. Robot. [42] H.-P. Huang, J.-L. Yan, T.-H. Huang, and M.-B. Huang,
Automat. (ICRA), New Orleans, LA, USA, vol. 3, 2004, “IoT-based networking for humanoid robots,” J.
pp. 2666–2672. Chinese Inst. Engs. (JCIE), vol. 40, no. 7, pp. 603–613,
[25] S. L. Chiu, “Control of redundant manipulators for task Oct. 2017.
compatibility,” in Proc. of IEEE Int. Conf. Robot. [43] T. Choi, H. Do, D. Park, C. Park, and J. Kyung “Real-
Automat. (ICRA), Raleigh, NC, USA, vol. 4, 1987, pp. time synchronization method in multi-robot system,” Int.
1718–1724. J. Elec. Letters, vol. 50, no. 24, pp. 1824–1826, 2014.
[26] J.-O. Kim and P. K. Khosla, “Dexterity measures for [44] J.-Y. Kim, I.-W. Park, J. Lee, M.-S. Kim, B.-K. Cho,
design and control of manipulators,” in Proc. IEEE/RSJ and J.-H. Oh, “System design and dynamic walking of
Int. Workshop Intel. Robots Syst. (IROS) & Intelligence humanoid robot KHR-2,” in Proc. IEEE Int. Conf.
for Mech. Syst., Osaka, Japan, vol. 2, 1991, pp. 758–763. Robot. Automat. (ICRA), Barcelona, Spain, 2005, pp.
[27] T. Yoshikawa, “Analysis and control of robot 1431–1436.
manipulators with redundancy,” in Robot. Research: 1st [45] C.-A. Cheng, H.-P. Huang, H.-K. Hsu, W.-Z. Lai, and
Int. Symp., Cambridge, MA, 1984, pp. 735–747. C.-C. Cheng, “Learning the inverse dynamics of robotic
[28] T. Yoshikawa, “Manipulability of robotic mechanisms,” manipulators in structured reproducing kernel Hilbert
Int. J. Robot. Research, vol. 4, no. 2, pp. 3–9, 1985. space,” IEEE Trans. Cybernetics, vol. 46, no. 7, pp.
[29] T. Yoshikawa, “Manipulability and redundancy control 1691–1703, Jul. 2016.
of robotic mechanisms,” in Proc. IEEE Int. Conf. Robot. [46] HIWIN Technologies Corp., Taichung City, Taiwan
Automat. (ICRA), St. Louis, MO, vol. 2, 1985, pp. 1004– (R.O.C.), Oct. 1989. [Online]. Available:
1009. [Link]
[30] I. Yamano, K. Takemura, and T. Maeno, “Development 5_710_gb.aspx
of a robot finger for five-fingered hand using ultrasonic [47] T. Yoshikawa, “Multifingered robot hands: control for
motors,” in Proc. IEEE/RSJ Int. Conf. Intel. Robots Syst. grasping and manipulation,” Int. J. Annual Rev. Control,
(IROS), Las Vegas, Nevada, USA, vol. 3, 2003, pp. vol. 34, no. 2, pp. 199–208, 2010.
2648–2653. [48] M.-B. Huang, H.-P. Huang, C.-C. Cheng, and C.-A.
[31] I. Yamano and T. Maeno, “Five-fingered robot hand Cheng, “Efficient grasp synthesis and control strategy
using ultrasonic motors and elastic elements,” in Proc. for robot hand-arm system,” in Proc. 11th IEEE Int. Conf.
IEEE Int. Conf. Robot. Automat. (ICRA), Barcelona, Automat. Sci. Eng. (CASE), Gothenburg, Sweden, 2015,
Spain, 2005, pp. 2673–2678. pp. 1256–1257.
[32] E. Torres-Jara, “Obrero: A platform for sensitive [49] S.-Y. Lo and H.-P. Huang, “Realization of sign language
manipulation,” in Proc. 5th IEEE-RAS Int. Conf. motion using a dual-arm/hand humanoid robot,” J. Intel.
Serv. Robot., vol. 9, no. 4, pp. 333–345, Aug. 2016.

VOLUME XX, 2019 17

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]
 This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2891007, IEEE Access
M.-B. Huang and H.-P. Huang: Innovative Human-Like Dual Robotic Hand Mechatronic Des. & Its Chess-Playing Expt. (Jan. 2019)

[50] H.-K. Hsu, H.-P. Huang, and S.-Y. Lo, “A humanoid 2002, the Research Fellow Award twice from 2002 to 2008, the
robotics simulation and control platform for NINO,” Int. Distinguished Research Fellow Award from the National Science Council,
J. Digital Human, vol. 1, no. 2, pp. 169–194, 2016. Taiwan, in 2009, the Distinguished Education Award on Radio Frequency
[51] Y.-R. Liu, M.-B. Huang, and H.-P. Huang, “Automated Identification from EPCglobal in 2010, the TECO Outstanding Science
grasp planning and path planning for a robot hand-arm and Technology Research Achievement Award in 2012, the HIWIN
system,” presented at the 11th IEEE/SICE Int. Symp. on Technologies Corp. Distinguished Industry Creation Award in 2014, the
Syst. Integration (SII), Sorbonne Université, Paris, Automation Engineering Medal, Chinese Institute of Automation
France, Jan. 14-16, 2019. Engineers in 2018. He has been editor-in-chief of the Journal of Chinese
[52] J.-H. Bae, S.-W. Park, D. Kim, M.-H. Baeg, and S.-R. Fuzzy System Association from 1997 to 1999 and the International Journal
Oh, “A grasp strategy with the geometric centroid of a of Fuzzy System from 1999 to 2002; associate editor of IEEE
groped object shape derived from contact spots,” in Proc. TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING from 2003 to
IEEE Int. Conf. Robot. Automat. (ICRA), St. Paul, MN, 2005, the International Journal of Advanced Robotics Systems,
USA, 2012, pp. 3798–3804. IEEE/ASME TRANSACTIONS ON MECHATRONICS; member of the Publication
[53] MYO™, Thalmic Labs Inc., Kitchener, Canada [Online]. Management Committee of IEEE/ASME journals; and editorial board
Available: [Link] member of the International Journal of Advanced Robotics from 2004 to
[54] J. Falconer, “Humanoid robot demonstrates sign 2008 and the International Journal of Service Robots. He is currently on
language,” IEEE SPECTRUM. New York, NY, USA. the editorial board of the International Journal of Electronic Business
Oct. 31, 2013. [Online]. Available: Management, Intl. J. Fuzzy Systems, an editor of the International Journal
[Link] of Automation and Smart Technology, and editor of iRobotics. He was
/ntu-taiwan-humanoid-sign-language named in Who’s Who in the World 2001 and 2002 and in Who’s Who in
the R.O.C. in 2002. He has served as the committee chair/co-chair and
member of several national, international, and IEEE conferences. He was
also a fellow of the Chinese Institute of Automation Engineers in 2010, the
Chinese Society of Mechanical Engineers in 2011, and Robotics Society of
Taiwan in 2018.

MING-BAO HUANG received a B.S. degree in

mechanical engineering from Chinese Culture
University, Taipei, Taiwan, in 2009, and an M.S.
degree in engineering science from National
Cheng Kung University, Tainan, Taiwan, in 2011.
He is currently pursuing a Ph.D. degree in
mechanical engineering with the NTU Robotics
Laboratory, Department of Mechanical
Engineering, National Taiwan University, Taipei,
Taiwan.
His current research interests include human and robotic grasping and
dexterous manipulation, mechanisms and machine design,
anthropomorphic robotic hand design, sensors for artificial hands, human-
robot interaction, and machine learning and pattern recognition.

HAN-PANG HUANG (S’83–M’86) received

M.S. and Ph.D. degrees in electrical engineering
from the University of Michigan, Ann Arbor, MI,
USA, in 1982 and 1986, respectively.
Since 1986, he has been with National Taiwan
University, Taipei, Taiwan, where he is currently
a professor with the Department of Mechanical
Engineering and the Graduate Institute of
Industrial Engineering and serves as the Director
of the NTU Robotics Laboratory. He has been
distinguished professor of the National Taiwan
University since 2006, and held the position of Zhong Zhuo-Zhang Chair
Professor since 2009. He served as the Director of the Manufacturing
Automation Technology Research Center, Associate Dean at the College
of Engineering, Director of the Graduate Institute of Industrial Engineering,
and Chairperson of Mechanical Engineering with National Taiwan
University. His current research interests include machine intelligence,
humanoid robots, intelligent robotic systems, prosthetic hands,
manufacturing automation, nanomanipulation, and nonlinear systems. He
has authored more than 380 papers on these topics that have been
published in refereed technical journals and conference proceedings.
Prof. Huang was a recipient of the Ford University Research Award
from 1996 to 1998, the Distinguished Research Award thrice from 1996 to

18 VOLUME XX, 2019

This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see [Link]