Auton Robot (2014) 36:67–78

DOI 10.1007/s10514-013-9362-z

Design and control of a three-fingered tendon-driven robotic hand

with active and passive tendons
Ryuta Ozawa · Kazunori Hashirii · Yohtaro Yoshimura ·
Michinori Moriya · Hiroaki Kobayashi

Received: 27 February 2013 / Accepted: 3 August 2013 / Published online: 21 August 2013
© Springer Science+Business Media New York 2013

Abstract This paper presents a design of a three-fingered 1 Introduction

robotic hand driven by active and passive tendons and pro-
poses control methods for this hand. The tendon-driven Multi-fingered robotic hands have been developed for the last
robotic hand consists of the thumb, the index and the mid- three decades to realize dexterous grasp and manipulation
dle fingers. The robotic thumb can move all the joints inde- for applications of the outer space, works in daily life and
pendently. In contrast, the index and the middle robotic fin- prosthetic hands. These robotic hands are usually designed
gers are under-actuated using the combination of active and using small actuators, linkage mechanisms or tendon-driven
passive tendons, and move the terminal two joints synchro- mechanisms (TDMs).
nously, which is one of the important features of the human Small actuators can be mounted in fingers or the palm of
digits. We present passivity-based impedance and force con- robotic hands and makes the hands compact (Kawasaki et
trollers for tendon-driven robotic fingers and discuss how to al. 2002; Mouri et al. 2011). Therefore, it is easy to install
combine them for fast and secure grasps. We experimentally the hand on the robotic arm. However, these actuators limit
validate that the robotic hand moves fast and manipulates an the grasping force or speed while linkage mechanisms and
object and demonstrate that the robotic hand grasps objects TDMs can use large actuators to realize large grasping force
in diverse ways. and speed. Linkage mechanisms are useful in making hands
compact but usually limit their motions in a plane (Birglen
Keywords Robotic hand · Tendon-driven mechanisms · et al. 2008), except a finger using a special double plan-
Manipulation · Design etary gear system (Niikura et al. 2011). TDMs make the
transmission mechanism complex but can be applied to spa-
tial mechanisms, therefore TDMs have been used to design
R. Ozawa (B)
multi-fingered robotic hands (Jacobsen et al. 1984; Mason
Department of Robotics, Ritsumeikan University,
1-1-1 Noji-higashi, Kusatsu, Shiga 535-8577, Japan and Salisbury 1985; Grebenstein et al. 2010; Bridgwater et
e-mail: [email protected] al. 2012).
Multi-fingered robotic hands are often designed using
K. Hashirii
fewer actuators than the number of joints to reduce the weight
Research & Development Division, Nabel Co., Ltd., Kyoto, Japan
of the hand (Carrozza 2004; Fukaya et al. 2000), or to mimic
Y. Yoshimura the anatomical features, especially, kinematic connection
Mitsubishi Electric Corporation, Tokyo, Japan between the proximal and distal interphalangeal (PIP/DIP)
joints of the digits (Kaneko et al. 2007; Kawasaki et al. 2002).
M. Moriya
Construction Machinery Engineering Department, This kinematic connection increases agility of object manip-
Kubota Corporation, Osaka, Japan ulation when the controllers are appropriately selected (Bae
et al. 2005), and helps to passively respond the external forces
H. Kobayashi
(Ozawa et al. 2009). The kinematic connection has been real-
Department of Mechanical Engineering Informatics, Meiji University,
1-1-1 Higashi-Mita, Tama, Kawasaki, Kanagawa 214-8571, Japan ized using the stiff linkage mechanisms (Kaneko et al. 2007;
e-mail: [email protected] Kawasaki et al. 2002). However, these mechanisms do not

123
68 Auton Robot (2014) 36:67–78

Fig. 2 Active and passive tendons: (top) active tendon, (bottom) pas-
sive tendon

adopt a different approach: we appropriately combine sev-

eral force and impedance controllers to realize a variety of
grasps without any force, tactile and vision sensors. This
approach can regulate the internal grasping force and real-
ize fast and secure grasps. For force and impedance con-
Fig. 1 Three-fingered tendon-driven robotic hand. The right robotic
trollers of TDMs, several approaches have been proposed
finger is the thumb that has three joints driven by four active tendons.
The left two fingers are the index and the middle fingers that have four using tension sensors (Abdallah et al. 2013; Mason and Sal-
joints driven by four active and two passive tendons isbury 1985). These methods are very useful for TDMs with
sheaths in compensating joint torque errors caused by large
friction in sheaths. On the other hand, the developed hand
respond the external force passively. Hence, elastic mecha- does not use any sheath. Thus, we construct the controllers
nisms have been introduced to joints (Carrozza et al. 2006; using the passivity-based approach (Arimoto 1996) because
Liu et al. 1999). These mechanisms are mainly focused on this approach makes it possible to intuitively combine several
adaptive grasping but are not considered for object manipu- controllers and to apply the same controllers to both fully-
lation such as pushing or sliding. actuated robotic thumb and under-actuated robotic fingers.
In this paper, we develop a new robotic hand and its We present several passivity-based controllers for TDMs and
grasp controllers. First, we present design of a tendon-driven explain how to combine them to realize five different grasps.
robotic hand, as shown in Fig. 1. This robotic hand con- We experimentally validate the performances of the hand and
sists of modulated motor boxes, full-actuated robotic thumb these controllers.
and two under-actuated fingers. The two under-actuated fin-
gers has two proximal fully actuated degrees of freedom
(DOF) and two distal coupled DOFs using the combination 2 Tendon kinematics and definition of TDMs
of active and passive tendons connected to actuators and elas-
tic elements, respectively. This special coupling is different First, the basic tendon kinematics is described. Two kinds of
from the decoupled combination of classical controllability tendons are used in the mechanisms; one is an active tendon
of the fingers with the passive adaptation (Carrozza et al. and the other is a passive tendon. An active tendon is con-
2006; Kaneko et al. 2007; Liu et al. 1999, 2007; Lotti et al. nected to an actuator, as shown in Fig. 2 (top). The tensile
2005), and enables us to design the biomimetic feature of the force of an active tendon is determined by the output of the
PIP/DIP joints that move synchronously in free space while actuator connected to the tendon. It makes no difference if
they conversely move to absorb the external force in contact the tendon is rigid or elastic. It is assumed that an active ten-
tasks. We demonstrate a method to design this special feature don can generate any positive tensile force. A passive tendon
based on the tendon kinematic analysis. is not connected to an actuator, but rather to an elastic ele-
Next, we present several grasp controllers for the hand. ment, as shown in Fig. 2 (bottom). The tensile force depends
Recently, postural hand synergies were introduced to ana- on its deflection. A passive tendon can adjust the pretension,
lyze a variety of hand motions using the principal compo- and we assume that the pretension is large enough to pre-
nent analysis of hand posture (Santello et al. 1998), and have vent the tendons from loosening during operation. Thus, the
been used to control multiple grasps to combine several prin- tensile force is uniquely determined according to the joint
cipal components of hand posture (Brown and Asada 2007; configuration.
Catalano et al. 2012). However, it is well known that hand Let L, M and N be the number of tendons, actuators and
posture is separate from the regulation of the contact force joints, respectively, l ∈ R L be the deflection of the tendons,
(Santello et al. 1998) and this regulation is important to real- q ∈ R N be the joint angle vector and θ ∈ R M be the motor
ize grasps (Johansson and Westling 1984). Therefore, we angle vector. Then, the following relation is established:

123
Auton Robot (2014) 36:67–78 69

 
la (1998) treated only CF-TDMs because the TDMs have only
l= = J j q + Jθ θ + l0
le active tendons. Definition 2 can be considered as the exten-
     
J R 0 sion of this concept to TDMs with active and passive tendons.
= a q+ θ θ + , (1) Both CS-TDMs and U-TDMs are used for the design of
Je 0 l e0
under-actuated robotic hand to adaptively grasp an object
where J j ∈ R L×N and J θ ∈ R L×M are Jacobian matri- with palm (Hirose and Umetani 1978; Dechev et al. 2001)
ces, l 0 ∈ R L is an initial tendon expansion, l a ∈ R M and and are frequently treated without distinction (Birglen et al.
l e ∈ R L−M are the expansion of active and passive tendons, 2008). However, the difference between these TDMs is crit-
J a ∈ R M×N and J e ∈ R (L−M)×N are Jacobian matrices with ical to design a robotic hand, e.g., CS-TDMs can grasp an
respect to the active and the passive tendons. Rθ ∈ R M×M is object at the fingertips while U-TDMs cannot. Therefore, we
a diagonal matrix and each diagonal element is the radius of design the thumb robot using a CF-TDM and the index and
a pulley mounted on a motor axis. The initial expansion of the middle finger robots using CS-TDMs.
an active tendon is adjustable, thus we assume that it equals
zero without loss of generality. On the other hand, an initial
expansion of a passive tendon l e0 ∈ R L−M must be positive 3 Design of tendon-driven robotic fingers
to prevent the tendon from loosening.
Let tensile force be f t . Using the virtual work principle, Figure 1 shows the three-fingered tendon-driven robotic hand
the relationship between f t ∈ R L and the joint torque τ ∈ that realizes some functions of the thumb, the index and the
R N is given by: middle fingers. The index and the middle fingers have a sym-
τ = −J Tj f t . (2) metric structure and are aligned side-by-side. The thumb is
located in front of the index finger with the incline to oppose
Then, the tensile force vector is: the middle finger and leans back 45◦ to grasp an object at an
f t = A1τ + f b = A1τ + A2ξ , (3) appropriate position. The lengths of the thumb and the digits
between the first joint and the fingertip are 93 and 133 mm,
where respectively, which are almost the same sizes of the human’s
 +
A1 = − J Tj and A2 = I L + A1 J Tj . (4) hand. Each finger uses four actuators that are Maxon DC
motors with the power of 11 W for generating more than
 + 50 N tensile forces. As a result, each joint can move about
f b is a bias force vector that does not affect τ , and J Tj is
3 Hz with the peak-to-peak amplitude of 30◦ , and the finger
the Moore–Penrose pseudo-inverse of J j . generates more than 10 N fingertip force.
Definition 1 If there exists ξ such that In the following subsection, we describe the details of the
thumb robot and the index and the middle finger robots.
f b = A2ξ > 0, (5)
then a mechanism driven by tendons is called a TDM. 3.1 The thumb robot

TDMs can be classified into the following mechanisms in Figure 3 shows the photo, kinematics, tendon route and
terms of the ranks of the matrices J j and J a (Ozawa et al. schematic graphs of the thumb robot. The thumb has three
2009): joints (N = 3). The first two and the third joints correspond
Definition 2 1. A TDM is called a controllable full-TDM to middle phalanx (MP) joints and the interphalangeal (IP)
(CF-TDM) if both the ranks of the matrices J j and J a joint of the human’s thumb, respectively. The thumb robot
equal the number of the joint N . must be designed as a CF-TDM that requires at least N + 1
2. A TDM is called a controllable semi-TDM (CS-TDM) if actuators to move all the joints independently. Therefore, we
the rank of the matrix J j equals N and that of the matrix used four actuators that are located in a modulated box at the
J a is less than N . basement. The tendon Jacobian matrix of the thumb robot is
3. A TDM is called an uncontrollable TDM (U-TDM) if the designed as follows:
⎡ ⎤
rank of the matrix J j is less than N . r r r
⎢ r −r −r ⎥
CF-TDMs are conventional robotic mechanisms that can J j = Ja = ⎢ ⎣−R
⎥, (6)
R 0⎦
drive all the joints independently and CS-TDMs have some
−R −R 0
joints that move synchronously. U-TDMs are kinds of non-
holonomic manipulators that cannot determine the joint con- where r = 4.5 mm and R = 7.0 mm. The rank of J a as well
figuration uniquely (Suzuki et al. 1996). The controllability as that of J j is N in this case, thus this tendon Jacobian matrix
of TDMs that was originally proposed in Kobayashi et al. J j satisfies the requirement of a CF-TDM in Definition 2.

123
70 Auton Robot (2014) 36:67–78

Fig. 4 The index and middle finger robots. (Top left) photo of the
fingers. The right and the left are the index and the middle fingers,
respectively. (Top right) tendon wiring of the index finger. The wiring
Fig. 3 Thumb robot. (Top left) photo, and (top right) tendon wiring,
of the middle finger is symmetric to this. (Bottom left) kinematics of the
(bottom left) kinematics and (bottom right) the sketches of the thumb
index and the middle fingers. (Bottom right) the sketches of the index
robot with the dimensions in millimeter
and the middle fingers with the dimensions in millimeter

3.2 The index and the middle finger robots

The index and the middle fingers are usually modeled

as the four joint mechanisms, as shown in Fig. 4 (bot-
tom left). The first two joints are the MP joints to realize
the abduction/adduction and curling/extension, respectively.
The terminal two joints are DIP/PIP joints to realize curl-
ing/extension. When the finger moves in free space, these IP
joints move synchronously, as shown in Fig. 5b. This means
that the finger uses only one driving DOF for the two joints,
i.e., the finger is under-actuated. On the other hand, the finger
Fig. 5 Connected motion of the DIP and PIP joints a normal, b free
resists the external force at the fingertip to rotate the DIP joint
motion, c contact motion
conversely to the PIP joint when the fingertip contacts with
the environment, as shown in Fig. 5c. For the sake of mod-
ularity, it is preferable for all fingers to have identical kine- (iii) The remaining three DOF, i.e., two MP joints and one
matic structure. Therefore, we need to consider the following of the IP joints of the index and the middle fingers, must
points to design the index and the middle finger robots: be controlled independently.

(i) The finger robots must be CS-TDMs because they are This feature is realized effectively if control of the remaining
under-actuated and their configuration must be deter- DOF does not affect the synchronization of the DIP and the
mined uniquely for the external force. PIP joints, and vice versa.
(ii) The DIP and the PIP joints must rotate synchronously To make this point clear, we assume that the finger robot
if no external force affects. is in the equilibrium, the actuators are fixed and the active

123
Auton Robot (2014) 36:67–78 71

and the passive tendons generate tensile forces proportional Next, to drive the mechanism using the tendons, there exists
to the tendon deflection, as follows: f b > 0 such that
 
f J Tj f b = J aT f ba + J Te f be = 0.
f t = ta = KJ j (q − q0 ) + f b (15)
f tb
  As the row vectors of J a and J e are linearly independent, Eq.
K a J a (q − q0 ) f
= + ba , (7) (15) can be rewritten as follows:
K e J e (q − q0 ) f be

where K, K a and K e are the diagonal stiffness matrices of all J aT f ba = 0 and J Te f be = 0. (16)
the tendons, the active ones and the passive ones, respectively,
Definition 1 and Eq. (16) indicated that the mechanism driven
q0 is the joint angle at the equilibrium, and f ba and f be are
by only the active tendons or only the passive tendons is an
the bias force vectors of the active and the passive tendons,
independent TDM. Thus, we can design the finger robots in
respectively. Then, from Eq. (2), the joint torque τ in the
the following procedure:
vicinity of the equilibrium point q0 is given as follows:
Step (1) Design J e satisfying the following conditions:
τ = −J Tj f t = −J Tj KJ j (q − q0 )
= −(J aT K a J a + J Te K e J e )(q − q0 ). (8) (a) J e q = 0 if and only if cT q = 0.
To guarantee that q0 is the unique equilibrium point, we need (b) rank J e = 1 and there exists f be > 0 such that
that J Te f be = 0.

rank J j = N = 4. (9)
Step (2) Design J a satisfying the following conditions:
The relationship between the DIP and PIP joint move-
ments is nearly linear in unconstrained motion (Darling et (c) rank J a = 3 and there exists f ba > 0 such that
al. 1994). Therefore, in this design, we assumed that the J aT f ba = 0.
relationship is linear and the ratio is 1:1. Then, the con- (d) rank J j = rank J a + rank J e = 4.
straint condition of the DIP and the PIP joints is expressed as
follows:
First, we consider Step (1) for the design of the index
 
cT q = 0 0 1 −1 q = q3 − q4 = 0. (10) finger. We combined Eq. (10) multiplied by the pulley radius
r with the symmetric equation to guarantee the existence of
Equation (8) indicated that the equilibrium point of TDMs f be satisfying Eq. (10) as follows:
generally depends on both active and passive tendons. This
   
dependency complicates the derivation of the equilibrium cT 0 0 r −r
Jeq = r q= q = 0. (17)
point. In addition, the active tendons must generate larger −c T 0 0 −r r
tension to keep the constraint condition (10) than to increase
the gripping force. These drawbacks are removed by decou- J e satisfies (a) and (b) in Step (1) when we choose f be =
pling the passive tendons from the active tendons. (1, 1)T > 0.
Therefore, the following relationship must be established Next, we consider Step (2). The active tendon Jacobian
among the ranks of these tendon Jacobians: is designed to drive the remaining DOF independently. The
joint motion q = (q1 , q2 , q3 , q4 )T can be decomposed as
rank J j = rank J a + rank J e = 4. (11) follows:
As the passive tendons remain unchanged under the con- ⎡ ⎤ ⎡ ⎤
q1 0
straint, the following equation must be satisfied: ⎢ q2 ⎥ ⎢ 0 ⎥
⎢ ⎥ ⎢ ⎥
J e (q − q0 ) = J e q − l e0 = 0. (12) q = a1 + a2 = ⎢ ⎥ ⎢ ⎥
⎢ q3 +q4 ⎥ + ⎢ q3 −q4 ⎥ . (18)
⎣ 2 ⎦ ⎣ 2 ⎦
q3 +q4
Note that l e0 can be neglected in the design process because 2 − q3 −q
2
4

it is adjustable. The IP joint connection is only the constraint

in this system. Thus, where a1 and a2 are vectors in the nullspace of J e and in the
range space of J Te , respectively. To guarantee the condition
rank J e = 1. (13) (d) in Step (2), the equation J a a2 = 0 must be satisfied for
Equations (13) and (11) indicate that any a2 . This can be done by choosing the row vectors of J a
from the nullspace of J e . Then, we can use the following J a
rank J a = 3. (14) to control the remaining DOF:

123
72 Auton Robot (2014) 36:67–78

⎡ ⎤
r r r r (a) (b)
⎢ r −r −r −r ⎥
Ja = ⎢
⎣−R
⎥. (19) Initial Position Bending IP joints Bending only
R 0 0⎦ connectedly DIP joint

−R −R 0 0
J a satisfies the condition (c) when we choose f ba =
(R, R, r, r )T > 0. Finally, J j is given as follows:
⎡ ⎤
r r r r
⎢ ⎥
  ⎢ r −r −r −r ⎥
Ja ⎢ −R R 0 0 ⎥
Jj = =⎢ ⎢ −R −R
⎥.
⎥ (20)
Je ⎢ 0 0 ⎥
⎣ 0 0 r −r ⎦
0 0 −r r Springs absorb
Springs constrain IP joint motion external force
The index and the middle finger robots are the same struc-
Fig. 7 The spring mechanisms and its effect in the digits. a The wire
ture, but the actuation and the tendon-wiring of the robots are
routes of the passive tendons. b The effect of the spring mechanism.
symmetric. These finger robots as well as the thumb robot When the terminal two joints (IP joints) move in free space, the springs
use four actuators and the first columns of J a in Eqs. (6) keep the lengths and allow the IP joints to move synchronously. In
and (19) have the same structure. Therefore, the same actu- contrast, the springs are shrunken and stretched when the external force
is acted on the fingertip
ator boxes can be used as shown in Figs. 3 and 4. Figure 6
focuses on the tendon routes of the IP joints. The two passive
tendons cross between the IP joints and two active tendons contact softer, but during the contact, all the joints move in
are aligned parallel. The IP joints synchronously move with- the same direction (or only the DIP joint moves). If the fin-
out changing the length of the passive tendons in free motion ger pushes the environment stronger, then the finger becomes
(Fig. 6b). On the other hand, when an external force is added straighter and the contact point moves in the tangential direc-
to the fingertip, the lengths of passive tendons change and tion. This makes the contact point slippery, and this is unde-
the DIP and PIP joints move in the anti-directions ( i.e., the sired motion during pushing operation.
PIP joint rotates in clockwise and the DIP joint rotates in The spring mechanism was designed, as shown in Fig. 7a.
counterclockwise. See Fig. 6c). The passive tendon depicted as the lines connected from the
Note that many conventional index finger robots have been left bottom of the fingertip to the upper end of the rods where
designed to realize connected IP joint motion, and are clas- the springs are attached. As shown in Fig. 7b, it was experi-
sified into two classes; (i) DIP and PIP joints are rigidly mentally demonstrated that the length of the springs was not
connected (Liu et al. 2007; Kaneko et al. 2007) or (ii) joint- changed from the initial position (left) when the IP joints
independent elastic components are inserted into the joints moved synchronously (middle). In contrast, when an exter-
(Carrozza et al. 2006; Liu et al. 1999; Lotti et al. 2005). The nal force is applied to the fingertip, the length was changed
former, the robot with the rigid joints, realizes connected to absorb the force.
motion in free space, but it is difficult to realize soft contact
motion. The latter, the robot with elastic components, can
4 Passivity-based control of TDM
(a) (b) (c)
A tendon driven robotic finger consists of rigid-body linkages
and actuators connected by tendons. Therefore, the dynamics
of the i-th finger is given as follows:

H i q¨i + d iq + giq + (J ij )T f it = 0 , (21)

i i
H iθ θ̈ + Biθ θ̇ + (J iθ )T f it = τ ia . (22)

The superscript i is used to distinguish the fingers and we

assign i = 1, 2, 3 for the thumb, index and middle fingers,
respectively. H i and H iθ are the inertia matrices of the robot
and the motors, respectively, d iq and giq are the Coriolis and
Fig. 6 A passive tendon mechanism to realize the special movements
between the DIP and PIP joints a normal, b free motion, c contact centrifugal force and the gravitational force of the links, Biθ
motion is the damping matrices of the motor and τ ia is the motor

123
Auton Robot (2014) 36:67–78 73

torque vector. The first and second equations are the dynam- 4.2 Taskspace control of TDMs
ics of the finger linkages and the motors, respectively. The
motor dynamics is connected to the link dynamics through The taskspace controller is used to control the position and
the active tendons. The elastic force generated by the passive force described in the task space. In the joint space control,
tendons affects only on the linkages. The control problem we needed to take the joint constraint into account for the
is how to design the control input τ a = (τ a1 , τ a2 , τ a3 ) for joint motion planning. In contrast, in the force control, we
the full-actuated thumb and the under-actuated fingers. PI do not need to consider about the joint constraint, because
controllers were frequently used for tendon-driven robotic each finger have three actuated DOF that are equivalent to
fingers with sheaths to cope with large friction (Mason and the dimensions of the force. The desired force depends on
Salisbury 1985; Abdallah et al. 2011). However, the designed tasks. In this paper, we design three taskspace controllers for
hand does not use any sheath in the transmission. Therefore, positioning, pushing and pinching. To realize a desired force,
we use passivity-based impedance and force controllers for we control the desired joint torque τ itaskk (k = 1, 2, 3) using
fingers and combine them to realize selected grasps. the following motor torque controller:
i
   
4.1 Joint space impedance control of TDMs τ iak = −K iv θ̇ + (J iθ )T Ai1 giq + τ itaskk + f ib , (26)

where K iv is the feedback gain. The first term of Eq. (26) gives
This subsection presents the joint space impedance controller
the motor damping, and the second term is for the gravity
(Ozawa et al. 2009). Let qid be the desired joint angle of the i-
compensation, one of the following task controllers and the
th finger. Then we can define the corresponding motor angle
bias force. In the following subsection, we design the desired
θ id as follows:
torque τtaskk
i .
θ id = −(Riθ )−1 J ia qid (23)
4.2.1 Impedance controller
For the convergence of qi on qid , we design a controller as
follows: A taskspace impedance controller that regulates a position
 
i
τ ia0 = −K ip θ i − K iv θ̇ + (J iθ )T Ai1 giq + f ib , (24) si ∈ R 3 of the fingertips in the taskspace to the desired one
sid is given as follows:
where K ip and K iv are the positive definite matrices, θ i = τ itask1 = −(J is )T K is si , (27)
θ i − θ id and f ib is the bias force. In general, the bias force
∂s i
is determined by using a constant value which is adequately where J is = ∂q i and s
i = si − si . Unlike the joint
d
large, by solving optimization problems (Mason and Sal- impedance controller, the dimension of the fingertip posi-
isbury 1985; Ma et al. 1993; Abdallah et al. 2013) or by tion si is three, which is equivalent to the driving DOF of the
determining the passive joint compliance (Kobayashi et al. fingers. Therefore, the joint configuration qi corresponding
1998). We adopt the constant bias force for the hand thanks to the position si is uniquely determined.
to the large actuator power. However, we can use any of them
except (Kobayashi et al. 1998) because this controller does 4.2.2 Force controller
not depend on bias force algorithms.
Note that the thumb robot can freely choose the desired When we would like to have the finger scratch or slide an
joint angle qid , while the desired angle of the index and object, we must control the endpoint force in the normal
the middle finger robots is restricted because of the under- direction to the environment. Then, we use the following
actuation. Let q̂id be the nominal desired joint position, which force controller (Ozawa and Moriya 2010):
is designed by an operator. If the passive tendons play role of  
joint constraints, a desired joint angle qid should be calculated τ itask2 = −(J is )T K is Sip si + Sif f id , (28)
from a nominal desired joint angle q̂id = (q̂d1 i , q̂ i , q̂ i , q̂ i )
d2 d3 d4
as follows: where Sip and Sif are the selection matrices that determine the
⎧ position-controlled and force-controlled directions, respec-
⎪
⎨q̂d
i
for thumb, tively, f id ∈ R 3 are the desired contact force. This position
qd =
i
 T controller is feedback but the force controller is feedfoward.
⎪
⎩Pi q̂i = q̂ i q̂ i q̂d3 i +q̂ i
d4 q̂d3 +q̂d4
i i
e d d1 d2
otherwise, Therefore, this controller does not suitable for accurately
2 2
(25) track a desired contact force. However, many operations used
in hand such as sliding and pushing an object usually do not
where Pie = I−(J ie )+ J ie , where (J ie )+ is a the Moore-Penrose require precise tracking of the contact force. Thus, we do not
pseudo-inverse of J ie . adopt a force feedback controller in this paper. If the force

123
74 Auton Robot (2014) 36:67–78

feedback is needed for application, we can modified it by middle finger, thus τ a = (τ a3 1 , τ 2 , τ 3 ). In contrast, three-
a3 a0
adding the force sensor at the fingertip and the integral term fingered grasp is realized only the blind grasping controller
of the force error to Eq. (28). τ a = (τ a3
1 , τ 2 , τ 3 ).
a3 a3
Adducted thumb is similar to the envelope grasp. How-
4.2.3 Blind grasping controller ever, the motion of the thumb is different from that in the
case of envelope grasp. In this case we consider a virtual
The blind grasping controller is designed for pinching an stick that the fingers roll up and that the thumb pushes.
object and is defined as follows (Ozawa et al. 2005): To roll up the stick, the fingers are controlled by the joint
impedance controller. The thumb is controlled by the task-
τ itask3 = −k ip (J is )T (xi − x j ), (29) space impedance controller to press the fingertip on a point
in the longitude axis of the virtual stick. Thus, the controller
where k ip is the feedback gain, xi is the position of the i-th
is given as τ a = (τ a1
1 , τ 2 , τ 3 ).
a0 a0
fingertip and x j is that of an opposed finger or the opposed Lateral grasp consists of two processes: preshaping and
thumb. This controller is similar to the taskspace impedance pinching. First, all the joints in the digits bend adequately and
controller (27) but the error vector is different. The taskspace the thumb move to the side of the index finger for preshaping.
in the impedance and force controllers may depend on the Then, the fingertip of the thumb puts on the side of the index
environment, but the fingertip position does not depend on finger to pinch an object. Therefore, in the first process, the
the environment and can be described in the local hand coor- joint-space impedance controller τ a = (τ a0 1 , τ 2 , τ 3 ) is
a0 a0
dinates. The error vector xi −x j expresses the opposed direc- used for all the fingers and the thumb. Then, the controller
tion between the thumb and one of the remaining digits. k ip is switched to τ a = (τ a2 1 , τ 2 , τ 3 ), where the joint-space
a0 a0
corresponds to one-dimensional spring between the finger- impedance controllers are used for the digits while the task-
tips, thus k ip is scalar. This controller drastically improves space force controller is used for the thumb.
the stability of the pinching motion.

4.3 Grasping controllers 5 Experiment

We can realize several hand motions by combining the above First, we executed two basic experiments to validate the joint
joint and task-space controllers. From Cutokosky’s taxon- impedance and the force controllers. Next, we combined the
omy (Cutkosky 1989) we selected five different grasps to task-space and the joint impedance controller to implement
manipulate a variety of objects as follows: several grasps.
Figure 8 shows the first and the second joint motion of the
1. Envelope grasp, index finger robot in the experiment, where the index finger
2. two-fingered precision grasp, kept the angle of the first joint unchanged while it moved the
3. three-fingered precision grasp, rest of the joints to track the sinusoidal function. The joint
4. adducted thumb, and peak-to-peak amplitude gradually varied between 50 and 30◦
5. lateral grasp. and the period gradually changed between 2.0 and 0.3 s. The
first joint almost kept the angle although the active tendons
Envelope grasp is used to grasp a heavy object with palm. are kinematically coupled among all the joints. The residual
Two-fingered and three-fingered precision grasps are used to error caused by the friction between the tendons and pul-
pinch a small object with the fingertips and to manipulate leys and by the unmodeled tendon stretch. The second joint
it. Lateral grasp is used to pinch a thin object such as paper tracked the desired motion until 6.5 s where the time period
and a key with the side of the index finger and the fingertip. was about 0.6 s. After 6.5 s, the joint still moved quickly
Adducted thumb is used to grasp a stick-shaped object such although the overshoot appeared because of the inertial force
as a screw driver and an umbrella. of the finger.
Envelope grasp requires all the joints in the fingers to close Figure 9 shows the photo of this experimental setting
gradually until the palm and the fingers are firmly contact for the pushing and pulling task and the fingertip posi-
with a grasped object. Therefore, we use only the joint space tion in the horizontal(x) direction. In this experiment, the
impedance controller τ a = (τ a0 1 , τ 2 , τ 3 ).
a0 a0 index finger was controlled by the controller (28) to track
Two-fingered precision grasp requires the thumb and the x(t) = 0.015 sin(4π t) + 0.075 m and to push the object
index finger to oppose each other and the middle finger to with f d = 4 N. The fingertip tracked the desired posi-
keep an appropriate configuration. Thus, this motion is real- tion with the response delayed by about 0.1 s, as shown in
ized using the blind grasping controller for the thumb and Fig. 9b. When the finger pulled the object, the passive ten-
the index finger and the joint impedance controller for the don on the near side was compressed to absorb the external

123
Auton Robot (2014) 36:67–78 75

Fig. 8 The trajectory of the 60 60

first and second joints of the real
desired
index finger. The desired angle
of the first joint is constant, as
shown in the left. In contrast, the 40 40

Angle [deg]

Angle [deg]
second joint motion in the right
is given as a sinusoidal function,
whose peak-to-peak amplitude
20 20
gradually varied between 50 and
30◦ and whose period gradually
changed between 2.0 and 0.3 s
real
0 0 desired

0 2 4 6 8 10 0 2 4 6 8 10
Time [sec] Time [sec]

Fig. 9 Pushing and pulling an

(a) Springs are in equilibrium Spring are absorbing the force
object using the index finger
robot. a The photo during the
experiment. b The endpoint
trajectory in the horizontal
direction

0.09
(b)
Fingertop position[m]

0.08

0.07

0.06
target fingertop path
fingertop path
0.05
0 1 2 3 4 5
Time[sec]

force, as shown in Fig. 9a. More detail results about slid- 6 Conclusion
ing manipulation have been reported in Ozawa and Moriya
(2010). In this paper, we designed a robotic hand and its controllers.
Figure 10 shows the snapshots of the five grasps. Two- We presented the basic kinematics of the TDMs with active
fingered and three-fingered precision grasps, envelope grasp and passive tendons. Based on the kinematics we developed
and adducted thumb were captured in 10 Hz. In contrast, lat- the three-fingered robotic hand which consists of the thumb,
eral grasp was captured in 5 Hz because this grasp consists the index and the middle fingers. The terminal two joints of
of the two actions: preshaping and pinching. In each grasp, a the index and the middle fingers were designed using the
helper manually placed an object between the fingertips and cross-coupled passive tendons to realize the connected IP
then the robotic hand moved to grasp it. As shown in Fig. 10, joint motion. We confirmed that the passive tendon mecha-
the hand reached to the final shapes within four flames after nisms in the index and the middle fingers worked as a con-
moving, and kept grasping the object after a helper released straint in the free motion and absorbed the external force in
it. Therefore, the hand realized fast and secure grasps with a constraint motion. In the robotic fingers, we used the linear
the combination of the joint and task-space impedance con- spring for these passive tendon mechanisms. In contrast, we
trollers. adjust the stiffness of the corresponding tendons to suit for

123
76 Auton Robot (2014) 36:67–78

(a) improve the size and weight of the hand systems to mount
on robotic arms or to use it as a prosthetic hand. We also pre-
sented passivity-based controllers for the fingers and com-
bined them to realize five grasps. Experiment validated that
the developed robotic hand moved about 3 Hz with the peak-
to-peak amplitude of 30◦ in free space using the joint space
impedance controller. The index finger robot slid an object in
2 Hz using the force controller. The passive stiffness of the
finger smoothly absorbed the disturbances during the slid-
(b)
ing. We also demonstrated that the robotic hand fast and
securely grasped a variety of objects using five different
grasps without any external sensors. These grasps are use-
ful in some tasks in our daily life. However, these controllers
cannot adaptively control the grasping force, which is impor-
tant, e.g., to grasp a slippery object (Johansson and Westling
1984). In addition, we still need more controllers for execut-
ing tasks in our daily life. Therefore, we will develop more
(c) controllers and increase the grasping ability.

References
Abdallah, M. E., Platt R. Jr., Hargrave, B., & Permenter, F. (2011). Posi-
tion control of tendon-driven fingers with position controlled actua-
tors. In IEEE international conference on robotics and automation,
Saint paul, MN (pp. 2859–2864).
Abdallah, M. E., Platt, R, Jr., & Wampler, C. W. (2013). Decoupled
(d) torque control of tendon-driven fingers with tension management.
The International Journal of Robotics Research, 32(2), 247–258.
Arimoto, S. (1996). Control theory of non-linear mechanical systems:
A passivity-based and circuit-theoretic approach. Oxford: Oxford
University Press.
Bae, J. H., Arimoto, S., & Yoshida, M. (2005). Control and dexterity in
robotic pinching under linking of movements of joints like human
fingers. In Proceedings of the first international conference on com-
plex medical engineering, Takamatsu, Japan (pp. 899–903).
Birglen, L., Laliberte, T., & Gosselin, C. (2008). Underactuated
robotic hands, Springer tracts in advance robotics (Vol. 40). Berlin:
(e) Springer.
Bridgwater, L., Ihrke, C., Abdallah, M., Radford, N., Rogers, J., Yayathi,
S., Askew, R., & Linn, D. (2012). The robonaut 2 hand. In IEEE
international conference on robotics and automation, Saint Paul,
MN (pp. 3425–3430).
Brown, C. Y. & Asada, H. H. (2007). Inter-finger coordination and
postural synergies in robot hands via mechanical implementation of
principal components analysis. In Proceedings of IEEE/RSJ interna-
tional conference on intelligent robots and systems, San Diego, CA
(pp. 2877–2882).
Carrozza, M. C. (2004). The SPRING hand: Development of a self-
Fig. 10 Object manipulation. a Envelope grasp, b two-fingered preci- adaptive prosthesis for restoring natural grasping. Autonomous
sion grasp, c three-fingered precision grasp, d adducted thumb and e Robots, 16, 125–141.
lateral grasp. Each snapshot flows from the top left to the bottom right. Carrozza, M. C., Cappiello, G., Micera, S., Edin, B. B., Beccai, L., &
a–d are captured in 10 Hz, and d is captured in 5 Hz Cipriani, C. (2006). Design of a cybernetic hand for perception and
action. Biological Cybernetics, 95, 629–644.
Catalano, M., Grioli, G., Serio, A., Farnioli, E., Pazza, C., & Bicchi,
A. (2012). Adaptive synergies for a humanoid robot hand. In IEEE-
operations in our daily life. We will investigate the non-linear RAS international conference on humanoid robots, Osaka, Japan
(pp. 7–14).
passive tendon mechanisms for fingers and design it. Cutkosky, M. R. (1989). On grasp choise, grasp models, and the design
The developed robotic fingers are almost same size of of hands for manufacturing tasks. IEEE Transactions on Robotics
our fingers, but the actuation boxes are still large. We will and Automation, 5(3), 269–279.

123
Auton Robot (2014) 36:67–78 77

Darling, W. G., Cole, K. J., & Miller, G. F. (1994). Coordination of Ozawa, R. & Moriya, M. (2010). Effects of elasticity on an under-
index finger movements. Journal of Biomechanics, 27(4), 479–491. actuated tendon-driven robotic finger. In Proceedings of IEEE inter-
Dechev, N., Cleghorn, W., & Naumann, S. (2001). Multiple finger national conference on robotics and biomimmetics, Tianjin, China
passive adaptive grasp prosthetic hand. Mechanism and Machine (pp. 891–896).
Theory, 36, 1157–1173. Santello, M., Flanders, M., & Soechting, J. F. (1998). Postural hand
Fukaya, N., Toyama, S., Asfour, T., & Dillmann, R. (2000). Design of synergies for tool use. The Journal of Neuroscience, 18(23), 10105–
the tuat/karlsruhe humanoid hand. In Proceedings 2000 IEEE/RSJ 10115.
international conference on intelligent robots and systems (pp. 1754– Suzuki, T., Koinuma, M., & Nakamura, Y. (1996). Chaos and nonlin-
1759). ear control of nonholonmic free-joint manipulator. In Proceedings
Grebenstein, M., Chalon, M., Hirzinger, G., & Siegwart, R. (2010). of the IEEE international conference on robotics and automation
Antagonistically driven finger design for the anthropomorphic DLR (pp. 2668–2675).
hand arm system. In Proceedings of IEEE international conference
on intelligent robots and systems, Taipei, Taiwan (pp. 609–616).
Hirose, S. & Umetani, Y. (1978). The development of soft gripper for the
versatile robothand. Mechanism and Machine Theory, 13, 351–359. Ryuta Ozawa received his Ph.D
Jacobsen, S. C., Wood, J. E., Knutti, D. F., & Biggers, K. B. (1984). The degrees in Mechanical Engineer-
UTAH/M.I.T. dextrous hand: Work in progress. The International ing from Meiji University in
Journal of Robotics Research, 3(4), 21–50. 2001. From 2001 to 2002 he was
Johansson, R. & Westling, G. (1984). Roles of glabrous skin receptors a research assistant at Meiji Uni-
and sensorimotor memory in automatic control of precision grip versity. From 2002 to 2003 he
when lifting rougher or more slippery objects. Experimental Brain was a post-doctoral research fel-
Research, 56, 550–564. low of the Japan Society for the
Kaneko, K., Harada, K., & Kanehiro, F. (2007). Development of multi- Promotion of Science. He joined
fingered hand for life-size humanoid robots. In Proceedings of IEEE the department of robotics, Rit-
international conference on robotics and automation, Roma, Italy sumeikan University in 2003.
(pp. 913–920). From 2008 to 2009 he was a
Kawasaki, H., Komatsu, T., & Uchiyama, K. (2002). Dexterous anthro- visiting scholar at IRISA/INRIA,
pomorphic robot hand with distributed tactile sensor: Gifu hand II. Rennes, France. He is currently
IEEE/ASME Transactions on Mechatronics, 7(3), 296–303. an associate professor at the
Kobayashi, H., Hyodo, K., & Ogane, D. (1998). On tendon-driven department of robotics, Ritsumeikan University. He was awarded the
robotic mechanisms with redundant tendons. The International Jour- encouraging prize for young researchers from the department of System
nal of Robotics Research, 17(5), 561–571. Integration in the Society of Instrument and Control Engineers (SICE) of
Liu, H., Butterfass, J., Knoch, S., Meusel, P., & Hirzinger, G. (1999). A Japan in 2006, the best paper award from the Robotics Society of Japan
new control strategy for DLR’s multisensory articulated hand. IEEE in 2007, the outstanding paper prize from the Society of Instrument
Control Systems Magazine, 15, 105–110. and Control Engineers of Japan in 2009. He is a member of the IEEE,
Liu, H., Meusel, P., Seitz, N., Willberg, B., Hirzinger, G., Jin, M. H., et SICE, ISCIE and the Robotics Society of Japan. His research interests
al. (2007). The modular multisensory DLR-HIT-hand. Mechanism are manipulation of robotic hands and robotic arm, teleoperation of
and Machine Theory, 42, 612–625. robotic hands, full-actuated/under-actuated tendon-driven robotic arm,
Lotti, F., Tiezzi, P., Vassura, G., Biagiotti, L., Palli, G., & Melchiorri, mass estimation with a variable stiffness device in micro gravity, bal-
C. (2005). Development of ub hand 3: Early results. In Proceed- ance control of a biped and so on.
ings of IEEE international conference on robotics and automation,
Barcelona, Spain (pp. 4488–4493).
Ma, S., Hirose, S., & Yashinada, H. (1993). Design and experiments for
a coupled tendon-driven manipulator. IEEE Control Systems Maga-
Kazunori Hashirii was born in
zine, 13, 30–36.
1984. He received his BS and ME
Mason, M. T. & Salisbury, J. K. (1985). Robot hands and the mechanics
degrees in robotics and computer
of manipulation. Cambridge: The MIT Press.
science from Ritsumeikan Uni-
Mouri, T., Endo, T., & Kawasaki, H. (2011). Review of gifu hand and
versity, in 2007 and 2009, respec-
its application. Journal of Mechanics, 39, 210–228.
tively. He has worked for Nabel
Niikura, R., Kunugi, N., & Koganezawa, K. (2011). Developement of
Co., Ltd, Japan, since 2009 and
artificial finger using the double planetary gear system. In Proceed-
he is currently a leader of the
ings of IEEE/ASME international conference on advanced intelligent
R&D division of the mechanical
mechatronics, Budapest, Hungary (pp. 481–486).
design group.
Ozawa, R., Arimoto, S., Nakamura, S., & Bae, J. H. (2005). Control
of an object with parallel surfaces by a pair of finger robots without
object sensing. IEEE Transactions on Robotics, 21(5), 965–976.
Ozawa, R., Hashirii, K., & Kobayashi, H. (2009). Design and control of
underactuated tendon-driven mechanisms. In: Proceedings of IEEE
international conference on robotics and automation, Kobe, Japan
(pp. 1522–1527).

123
78 Auton Robot (2014) 36:67–78

Yohtaro Yoshimura received Hiroaki Kobayashi received the

his BE and ME degree in robot- BE, ME, and Ph.D. degrees in
ics and mechanical engineering precision engineering in 1971,
from Ritsumeikan University, 1973, and 1977, respectively, all
Japan, in 2010 and 2012, respec- from Kyoto University. Since
tively. He is currently working as 1976 he has been with the School
a software developer of the head of Science and Technology at
unit in Mitsubishi Electric Cor- Meiji University, Japan, where
poration, Japan. he is currently a Professor. From
1988 to 1989, he was a postdoc-
toral fellow in Computer Infor-
mation Science at the School
of Engineering and Applied Sci-
ence, University of Pennsylva-
nia. His research interests are in
the area of intelligent control of robotic systems including machine
learning. He is a member of IEEE, JSME, CISE, ISCIE, and RSJ.
Michinori Moriya received his
BE and ME degree in robot-
ics and mechanical engineer-
ing from Ritsumeikan Univer-
sity, in 2007 and 2009, respec-
tively. He is currently working as
an engineer in the construction
machinery engineering depart-
ment, Kubota Corporation. His
interest is the design of backhoes.

123