2015 IEEE 54th Annual Conference on Decision and Control (CDC)

December 15-18, 2015. Osaka, Japan

Modeling and Control of a Multifingered Robot Hand for Object

Grasping and Manipulation Tasks
Matheus F. Reis, Antonio C. Leite and Fernando Lizarralde

Abstract— In this work, we address the problem of kinematic kinematic configuration of the fingers around the palm,
modeling and control design of a multifingered robot hand. anthropomorphism and dexterity levels, applicability, type
Each robot finger is modeled as a parallel manipulator and and placement of the actuators, among others [11].
its kinematic constraints are computed from empirical analysis
due to the inherent mechanical complexity of the mechanism. A multifingered robot hand can be modeled as a set of
The motion control problem for a grasped object is solved small robots fixed on a common base or palm. Indeed,
by using the kinematic control approach, which is able to the kinematic and dynamic models of a robot hand with
ensure the asymptotic stability of the output tracking error on-site actuation are quite similar to those of a typical
and the prehension of the object. The kinematics-based control industrial robot, which simplifies the modeling problem. On
scheme is designed to include the contact model in the hand
Jacobian matrix, allowing for the simultaneous control of the the contrary, remote actuation adds some challenges to the
object position as well as the relative position between the control design, such as, the compliance of the transmission
fingers. Experiments are carried out with a three-fingered robot system and non-negligible dynamic effects (e.g., friction and
hand executing grasping and manipulation tasks of soft objects. backlash). Meanwhile, most of control strategies developed
Practical results are shown to illustrate the performance and for coordinated manipulation between robot manipulators
effectiveness of the proposed methodology.
can be immediately extended and applied to a multifingered
I. INTRODUCTION robot hand grasping an object [12], [13]. However, the design
Most of robot manipulators used for industrial applications of high-level control schemes for a robot hand must take into
are, in general, composed of a large arm equipped with a account the interaction between the fingers and the object.
single gripper or tool for interacting with the environment. This mutual interaction depends on several aspects, such
These end effectors are custom designed for specific tasks as modeling of contact points (e.g., compliance, friction),
and are able to grasp only a particular class of objects and force/torque control at the contact points, the mobility of
workpieces, which reduces the versatility of the robotic sys- the fingers and the contacts (e.g., fixed, rolling), as well
tem [1]. Another major drawback usually found in industrial as planning algorithms for grasping/manipulating of objects
robots is the lack of dexterity of their large-scale manipu- [14], [15]. In this context, several control strategies for robot
lators. This behavior is more evident when the robot arm hands operating under a wide range of uncertainties have
has to accurately perform precise movements of the payload been developed using adaptive, robust, hybrid and artificial
in assembly tasks. Indeed, fine motions of a grasped object intelligence techniques [16]–[20].
require accurate motions of the robot joints that are difficult
to achieve owing to the size of the links. On the contrary,
a robot hand is able to reconfigure itself for grasping a
sort of objects and for achieving a dexterous manipulation
in several tasks [2]. Other capabilities added to robotic
systems endowed with robot hands are perception of physical
properties and active exploration of the environment, which
are usually unfeasible with simple grippers [3].
Following this trend, a number of robot hands have been
developed by universities and research centers around the
world for industrial, research and prosthetic applications, Fig. 1. Multifingered robot hand: (a) 3-Finger Adaptive Robot Gripper;
such as, the DLR Multisensory Articulated Hand [4], the (b) Articulated finger with elastic tendons. Courtesy from Robotiq.
NASA Robonaut Hand [5], the Barrett Hand [6], the Gifu
Hand [7], the Shadow Dexterous Hand [8], the Schunk In this work, we present a kinematic framework for
Anthropomorphic Hand [9], and the i-Limb Ultra Prosthetic modeling and control design of a multifingered robot hand.
Hand [10]. The main differences among these devices are Each robot finger is considered as a parallel manipulator and
related to key issues of design, such as, number of fingers, its kinematic constraints are computed from an empirical
modeling because of the inherent mechanical complexity of
This work was partially supported by the Brazilian funding agen- the mechanism. The kinematic control approach is used to
cies CAPES, CNPq and FAPERJ. The authors are with the Depart- tackle the position control problem for the mechanism during
ment of Electrical Engineering, COPPE/Federal University of Rio de
Janeiro, P. O. Box 68504, RJ 21941-972, Rio de Janeiro, Brazil. E-mail: the constrained motion, ensuring the asymptotic stability of
[email protected], {toni, fernando}@coep.ufrj.br the output tracking error and a firm grasping. The contact

978-1-4799-7886-1/15/$31.00 ©2015 IEEE 159

model and the relative position between the fingers are both and k = 1, 2, 3. Let the pair {pk , Rk } be the position vector
included in the kinematics-based control scheme to allow for and the rotation matrix, which represent respectively the
a successful motion control and prehension of the object. The location and the orientation of the k-th fingertip frame Fk
key idea behind our solution is based on the previous work with respect to the palm frame Fp . The pose of the k-th
on modeling and control design for general robotic systems finger can be obtained from the forward kinematics map as:
subject to kinematic constraints [21]. Experimental tests,
pk = pk (θk ) , Rk = Rk (θk ) , (1)
performed with the 3-Finger Adaptive Robot Gripper from
Robotiq (Fig. 1a), illustrate the applicability of the proposed where θk ∈ Rnk is the joint position vector for the k-th
methodology for object grasping and manipulation tasks by finger. The coordinate transformation between the frames Fk
general multifingered robot hands. and Fp for the k-th finger can be written in terms of the
II. MODELING AND CONTROL DESIGN homogeneous transformation matrix as [22]:
 
Consider the modeling and control design problem of a Rk (θk ) pk (θk )
Tk (θk ) = , (2)
multifingered robot hand subject to kinematic constraints. 0T 1
Here, we assume that the phalanges of the robot finger are
where the rotation matrices are given by R1 = Rx (−θk123 ),
interconnected by means of rigid links and springs (Fig. 1b).
R2 = Ry (θ2s ) Rx (θk123 ), R3 = Ry (−θ3s ) Rx (θk123 ), and the
Thus, each articulated finger can be considered as a closed-
position vectors are obtained by pk = (pkoffset +pklink ) with
loop kinematic chain, and we can model the robot fingers      
as micro parallel robots composed of actuated and passive 0 (lc +l0 ) ss −(lc +l0 ) ss
joints. The springs provide a net restitution force, which p1link = ls , p2link = −ls , p3link = −ls ,
keeps the fingers stretched as much as possible for a given lc (lc +l0 ) cs −(lc +l0 ) cs
configuration. The passive joints are indirectly controlled by
and pkoffset is the offset value between the palm frame Fp and
the motion of the actuated joints by means of a transmission
the base frame of the k-th finger Fbk . The elements of the
system. Then, for the purpose of modeling and control, we
vectors and matrices are denoted according to the following
also have to consider the problem of kinematic coupling
notation: s1 = sin(θk1 ), c1 = cos(θk1 ), ss = sin(θks ),
between the actuated joints and the passive ones.
cs = cos(θks ), s12 = sin(θk1 + θk2 ), c12 = cos(θk1 + θk2 ),
In this section, to develop the kinematic modeling for the
s123 = sin(θk1 + θk2 + θk3 ), c123 = cos(θk1 + θk2 + θk3 ),
robot hand, we follow an intuitive and simplified approach
ls = (l1 s1 + l2 s12 + l3 s123 ), lc = (l1 c1 + l2 c12 + l3 c123 ).
where we consider that each finger may move independently
Also, Rx (·), Ry (·), Rz (·) are elementary rotation matrices
of the others. Following this key idea, the robot hand can
around the x-, y- and z-axis respectively. Notice that, θkj for
be considered as a multi-robot system composed of a set of
j = 1, 2, 3 and θk123 = (θk1 +θk2 +θk3 ) are related to the k-th
robot fingers fixed on a common base or palm. Since each T T
finger with θ1 = [ θ11 θ12 θ13 ] , θ2 = [ θ21 θ22 θ23 θ2s ] ,
finger can be modeled as a single and independent parallel T
and θ3 = [ θ31 θ32 θ33 θ3s ] .
mechanism, it can be controlled separately. Considering the
Let vk ∈ R6 be the velocity vector of the k-th fingertip
robot hand adopted in this work, the convention for links
frame Fk with respect to the palm frame Fp . It can be
and joints motions is depicted in Fig. 2. The palm coordinate
system denoted by Fp is the inertial frame in which all related to the joint velocity vector θ̇k ∈ Rnk by the differential
vectors and matrices will be referred to hereafter. kinematics equation as:
 
ṗk
vk = = Jfk (θk ) θ̇k , (3)
ωk
where the pair {ṗk , ωk } denotes the linear and angular veloc-
ity components of vk and Jfk ∈ R6×nk denotes the geometric
Jacobian matrix of the k-th finger, whose expressions will
be omitted here for the sake of space saving. Notice that,
the angular velocity vector ωk can be related to the rotation
matrix Rk and its time-derivative Ṙk as S(ωk ) := Ṙk RkT ,
where S(·) is the cross product operator [22].
B. Empirical Model for the Mechanism Velocity
In this section, we consider that the motions of the robot
fingers are subject to kinematic constraints on the velocity.
As mentioned earlier, each finger can be considered as a
Fig. 2. Convention for the links and joints motions of the robot hand.
closed-chain mechanism and its passive joints are indirectly
controlled by the motion of the actuated joints using a sort
A. Forward and Differential Kinematics of transmission system. In this context, it is well-known that
Consider a robot hand composed of n articulated fingers the joint position vector θk and the corresponding geometric
coupled to the palm, where each one has nk revolute joints Jacobian matrix Jfk for the k-th finger could be partitioned

160
into actuated and passive parts [11]. Thus, it is possible to Similarly, taking the time-derivative of (9) we obtain:
rewrite the differential kinematics (3) only as a function of
θ̇2s − θ̇ms = 0 , θ̇3s + θ̇ms = 0 . (13)
the actuated joint velocities. To achieve this, it would be
necessary to find an explicit relationship for the passive joint Then, defining a modified Kronecker delta operator
variables in terms of the actuated ones, for example, by
θmk ≤ θj∗ ,

1,
using the direct integration of the velocity data. However, δkj = (14)
0, θmk > θj∗ ,
the successful finding of an analytical solution would also
depend on the mechanical complexity of the robot hand. it is possible to rewrite (10), (11) and (12) as:
Here, we propose an alternative and simplified solution, θ̇k1 = δk1 θ̇mk , (15)
where the key idea is to establish the kinematic relationship
between the joint position vector of k-th finger θk ∈ Rnk and θ̇k2 = κr (1−δk2 ) θ̇mk , (16)
the angular position vector of the motor shafts θm ∈ Rnm . It θ̇k3 = −δk1 δk3 θ̇mk , (17)
can be represented in the configuration space as:
where θj∗for j = 1, 2, 3 is a bound for the angular position
θk = gk (θm ) , (4) of the j-th joint for the k-th finger, such that, θj∗ = γj .
Now, consider that an actuator can be shared by two or
where gk (·) : Rnm 7→ Rnk is a set of algebraic functions. more robotic fingers. Thus, we can split the angular position
For the 3-Finger Adaptive Robot Gripper we consider that: vector of the motor shaft into unshared and shared terms
(i) θkj for j = 1, 2, 3 and θks correspond to the angular as θm := [ θuT θsT ]T , allowing us to rewrite the velocity
position of finger joints; (ii) θmk for k = 1, 2, 3 and θms kinematic relationship (5) according to:
correspond to the angular position of motor shafts.  
 θ̇u
If gk is differentiable with respect to time, the joint θ̇k = Juk (θm ) Jsk (θm )

, (18)
velocity vector of the k-th finger θ̇k can be written in terms θ̇s
of the corresponding angular velocity vector of the motor where Juk and Jsk are the unshared and shared actuator Jaco-
shafts θ̇m as: bian matrices, θ̇u and θ̇s are the unshared and shared angular
θ̇k = Jck (θm ) θ̇m , (5) velocity vectors of the motor shafts. For the 3-Finger Adap-
tive Robot Gripper, we have θm = [ θm1 θm2 θm3 θms ]T ,
where Jck (θk ) = (∂gk (θm )/∂θm ) ∈ Rnk ×nm is named as the
with θu = [ θm1 θm2 θm3 ]T and θs = θms . Therefore, using
actuator Jacobian matrix of the k-th finger. Because of the
(13), (15), (16) and (17), the Jacobian matrices in (18) take
mechanical complexity of the mechanism, we have recorded
the simple form:
the angular position data of the motor shafts and used these
T
data sets to empirically derive the algebraic functions. Our Ju1 = η1 0 0 , Js1 = 0 0 0 ,
  
experiments have demonstrated that the following relations T
Ju2 = 0 η2 0 , Js2 = 0 0 0 1 ,
  
hold for the joints of the 3-Finger Adaptive Robot Gripper: T
Ju3 = 0 0 η3 , Js3 = 0 0 0 −1 ,
  

θmk − 0.37 , θmk ≤ γ1 ,
θk1 = (6)
0.565 , θmk > γ1 , with ηj = [ δj1 κr (1−δj2 ) −δj1 δj3 0 ]T for j = 2, 3 and
η1 = [ δ11 κr (1−δ12 ) −δ11 δ13 ]T . Here, we have shown

0, θmk ≤ γ2 ,
θk2 = (7) that the actuator Jacobian matrices for each finger can be
κr (θmk − 0.959) , θmk > γ2 ,
 directly computed from the algebraic functions relating the
−θk1 , θmk ≤ γ3 ,
θk3 = (8) finger joint position to the angular position of the motor shaft.
−0.354 , θmk > γ3 ,
Finally, substituting (18) into (3), we obtain the following
where γ1 = 0.935 rad, γ2 = 0.959 rad, γ3 = 0.776 rad, and differential kinematics equation:
κr = 2.25 is a reduction factor due to the lever arm between
vk = Jfk Juk Jsk θ̇m = J˜k θ̇m ,
 
(19)
the second phalange and the rigid links behind the k-th finger.
In addition, because of the inherent coupling between the where J˜k ∈ R6×nm is the modified Jacobian matrix.
fingers 2 and 3 in the scissor mode, we have:
C. Multifinger Grasp
θ2s = θms , θ3s = −θms . (9) In this section, we present the kinematic modeling for
Now, we can take the time-derivative of (6)-(8) obtaining the a multifingered robot hand grasping and/or manipulating a
kinematic relationship in terms of angular velocities: given object. Here, we suppose that the object is held firmly
 between the palm and the fingertips, while it follows a de-
θ̇mk , θmk ≤ γ1 , sired reference trajectory. The velocity kinematic relationship
θ̇k1 = (10)
0, θmk > γ1 , between the grasped object and the motor shafts can be
obtained from the differential kinematics equation of the

0, θmk ≤ γ2 ,
θ̇k2 = (11) entire mechanism. According to (19), we can stack the linear
κr θ̇mk , θmk > γ2 ,
 and angular velocity vectors vk for all fingers to obtain:
−θ̇k1 , θmk ≤ γ3 ,
θ̇k3 = (12)
0, θmk > γ3 . vh = Jf Jc θ̇m = Jh θ̇m , (20)

161
where vh ∈ R18 is the velocity vector of the robot hand and force distribution [1]. This behavior can be avoided by using
Jh ∈ R18×nm is the well-known hand Jacobian matrix, writ- a hybrid position/force control scheme, which would require
ten in terms of the product of matrices Jf ∈ R18×(n1 +n2 +n3 ) that the robotic fingers were equipped with force/torques
and Jc ∈ R(n1 +n2 +n3 )×nm . Notice that, vh = [ v1T v2T v3T ]T , sensors [23]. Here, we present a kinematic formulation to
Jf = diag(Jf1 , Jf2 , Jf3 ) and Jc = [ (Jc1 )T (Jc2 )T (Jc3 )T ]T . deal with the object manipulation problem based on the
Now, let the pair {rko , Rko } be the position vector and relative position of the fingers. Let ph ∈ Rnk be the stacked
the rotation matrix of the object frame Fo with respect to position vector of the k-th fingertip frame Fk for k = 1, 2, 3
k-th contact point frame Fck . According to [21], [23], the with respect to the palm frame Fp . From the differential
object velocity vector vo can be related to the object velocity kinematics (19) we obtain:
vector at the k-th contact point vck by:
  ṗh = JfP Jc θ̇m , (28)
Rko S(rko )
vck = Ak (rko ) vo , Ak (rko ) = , (21) where JfP = diag(Jf1P , Jf2P , Jf3P )
is the position term of the
0 Rko
stacked geometric Jacobian matrix of the fingers. Now, let
where Ak ∈ R6×6 is the Adjoint matrix related to the k-th pr ∈ R6 be the relative position vector between the fingers
contact point. The velocity vector vck can be parameterized given by:
in terms of a generic velocity vector wk using:    
p −p I −I 0
pr := 1 2 = Ãp ph , Ãp := . (29)
vck = vk + HkT wk , (22) p2 −p3 0 I −I
nr ×6 From the time-derivative of (29) and using (28), we obtain
where the columns of the constraint matrix Hk ∈ R
represents the nr degrees of freedom for the k-th finger at the differential kinematics equation for the relative position:
the contact point [21]. We can stack (21) and (22) for all
ṗr = Ãp JfP Jc θ̇m = J˜r θ̇m . (30)
fingers to obtain:
where J˜r ∈ R6×nm is the constraint Jacobian matrix.
vh + H T w = vc , vc = A vo , (23)
E. Kinematic Control Approach
where w = [ w1T w2T w3T ]T , H T = diag (H1T , H2T , H3T ), and
A = [ AT T T T T T T T In this section, we consider the problem of controlling the
1 A2 A3 ] with vc = [ vc1 vc2 vc3 ] .
Then, combining (20) and (23) we obtain the following object position during the manipulation task, as well as the
differential kinematics equation: relative position between the fingers to ensure a firm grasp
of the object. The control design is based on the kinematic
Jh θ̇m + H T w = A vo , (24) control approach at the velocity level due to its simplicity
and pre-multiplying both sides of (24) by the pseudo-inverse of implementation and satisfactory performance when the
matrix A† and the annihilating matrix Ã, such that A† A = I task does not require too fast motion or rapid accelerations.
and ÃA = 0, we obtain the multifinger kinematic model as: Notice that, the accuracy of the results is improved if the
mechanism has high gear reduction ratios [22]. In this
vo = (A†Jh ) θ̇m + (A†H T ) w , (25) context, the movements of the fingers can be defined simply
T
0 = (Ã Jh ) θ̇m + (Ã H ) w , (26) by θ̇m := u, where u ∈ Rnm is the velocity control signal
applied to the motor drivers of the corresponding joints.
where A is of full column rank. Notice that, from (25) and Let po ∈ R3 be the position vector of the object frame
(26), the variable w can be interpreted as the velocity vector Fo with respect to the palm frame Fp . Here, we consider
of the contact point. Indeed, during the grasping situations, that the control goal for the robot hand is to follow a
the contact point between the fingertip and the object can time-varying reference trajectory pod (t) for the object, while
be modeled as a virtual rotational/prismatic joint performing keeping a constant relative position prd between the fingers.
rolling/sliding motion [21]. If we specify θ̇m , it is possible The control goals can be described as:
to solve (26) for w, provided that the matrix (ÃH T ) is non-          
singular. Thus, we can rewrite (25) as: po pod (t) eo pod − po 0
→ , := → , (31)
pr prd er prd − pr 0
vo = A† [ Jh − H T (Ã H T )† Ã Jh ] θ̇m = J˜h θ̇m , (27)
where eo ∈ R3 is the object position error and er ∈ R3 is the
where J˜h ∈ R6×nm is the modified hand Jacobian matrix. relative position error. Considering (27) and (30), we obtain
the following position control system:
D. Object Manipulation
J˜hP
   
Consider the object manipulation problem for a multifin- ṗo
= u, (32)
gered robot hand. In general, when the fingers hold a rigid ṗr J˜r
object their relative position and orientation have to be kept
where J˜hP is the position term of the modified hand Jacobian
constant during the constrained motion. Otherwise, if a rel-
matrix. From (31) and (32), the error equation takes the form:
ative motion between the fingers is allowed or commanded,
J˜hP
     
the object can be damaged due to the occurrence of unstable ėo ṗod
= − u. (33)
grasping situations, such as loss of contact or uncontrollable ėr 0 J˜r

162
 (a) Object pos, y-axis (b) Rel pos fingers 1-2, y-axis
A joint velocity control signal u ∈ Rnm , that linearizes and 40 −60
pod po
stabilizes the error system (33), can be given by
20 −70
†  † 
J˜hP J˜hP
   
νo ṗod + Λo eo
u= = , (34)

mm
mm
J˜r J˜r
0 −80
νr Λr er
−20 −90
where Λo = λo I and Λr = λr I are the position gain matrices,
prd pr
and the pair {νo , νr } denotes the Cartesian control signals. −40 −100
0 5 10 15 20 0 5 10 15 20
For any choice of λo and λr as positive scalar gains, we can
show that the error system dynamics is asymptotically stable (c) Rel pos fingers 2-3, x-axis (d) Rel pos fingers 2-3, y-axis
100 2
and, therefore, limt→∞ eo (t) = 0 and limt→∞ er (t) = 0 . prd pr
Notice that, the joint velocity control signal (34) holds 80 0
under the assumptions of full knowledge of the mechanism

mm

mm
kinematic parameters and provided that the mechanism is far 60 −2

from singular configurations. Thus, the kinematic singularity 40 −4

problem can be addressed from the full-rank analysis of the prd pr
Jacobian matrices, J˜hP and J˜r , as well as the stacked one. 20
0 5 10 15 20
−6
0 5 10 15 20
Time, s Time, s
Since some directions of movement are constrained by
the geometry of the tasks or kinematic limitations of the Fig. 3. Trajectory tracking and position regulation: (a) object position in
mechanism, we reduce the scope of the position control the y-axis; (b) relative position between the fingers 1 and 2, in the y-axis;
problem to command the motion of the object only in the (c)-(d) relative position between the fingers 2 and 3, in the x- and y-axis.
directions of interest. An intuitive and easy way to employ
(a) Object pos error (b) Rel pos error
the proposed control scheme is to left multiply the Jacobian 10 20
er,y 1−2
matrices by a selection matrix, composed of rows of ones er,x 2−3
and zeros, which selects only the elements corresponding to 5 10 er,y 2−3

the desired axes of motion.

mm

mm
0 0

III. EXPERIMENTS AND RESULTS −5 −10

In this section, we present experimental results in order to eo,y

−10 −20
illustrate the performance and effectiveness of the proposed 0 5 10 15 20 0 5 10 15 20
Time, s Time, s
methodology in terms of modeling and control design. The
3-Finger Adaptive Robot Gripper is a programmable and Fig. 4. Position errors: (a) object position error, in the y-axis; (b) relative
mechanically intelligent robot hand, which is able to ma- position error between the fingers 1 and 2, in the x-axis; relative position
nipulate a wide variety of objects. By means of the Robotiq error between the fingers 2 and 3, in the x- and y-axis.
Graphical User Interface it is possible to control force, speed
(a) Cartesian control signals (b) Joint control signals
and position parameters for each finger, thanks to many 1000 10
θ̇m1
options of communication protocols [24]. All experiments θ̇m2
500 5
were performed on a desktop PC with Intel Core i5 Processor θ̇m3
mm s−1

rad s−1

θ̇m4
(8M Cache, 2.8 GHz) 4GB RAM, running Windows OS. 0 0

The control algorithm was implemented in Matlab (The νo,y

−500 νr,y 1−2 −5
MathWorks Inc.) Release 2013a and a set of scripts were de- νr,x 2−3
νr,y 2−3
veloped to execute the control loop, send/receive data to/from −1000
0 5 10 15 20
−10
0 5 10 15 20
the Robotiq gripper and perform all necessary calculations. Time, s Time, s

An USB connection was created using the Matlab Serial

Fig. 5. Control signals: (a) position control signals in the Cartesian space;
Port class, and the communication link with the Robotiq (b) velocity control signals in the joint space.
gripper was established by the Modbus RTU communication
protocol. The data conversion between the angular position
and velocity of the motor shafts, θm and θ̇m , and the in the “pinch” mode and an ad-hoc Matlab routine closes the
corresponding sent/received one-byte values aj , bj ∈ [0, 255] fingers around the object until it is firmly grasped. From the
are given empirically for k = j = 1, 2, 3 by forward kinematics map, this routine calculates the desired
relative position between the fingers 1 and 2 in the y-axis,
θmk = 0.0071 aj −0.0419 , |θ̇mk | = 0.0290 bj +0.1884 ,
as well as between the fingers 2 and 3 in the x- and y-axis.
and for j = 4 by Finally, another routine releases the object by opening the
fingers, when the experiment ends. Thus, the desired relative
θms = −0.0023 a4 −0.3112 , |θ̇ms | = 0.0100 b4 +0.0650 .
position is given by prd ≈ [ −73.1 31.0 −1.3 ]T mm.
For the sake of simplicity, the object grasping and manip- The experimental test consists in tracking a sinusoidal
ulation task was performed using a plastic ball with a radius trajectory in the y-axis direction of the Cartesian space.
of r = 36.2 mm. In this case, the Robotiq gripper is set up The reference signal is defined by the analytical expression:

163
yod (t) = As sin(ωs t), where As and ωs are the amplitude and issues discussed here, some topics for future developments
angular frequency of the sinusoidal signal, and t = [ 0, Tmax ] are: (i) to consider the dynamic model of the robot hand
where Tmax = 60 s is the total runtime of the experiment. The in the control design; (ii) to implement hybrid position/force
lengths of the finger phalanges are l1 = 57 mm, l2 = 38 mm, control strategies for object grasping and manipulation tasks.
and l3 = 32 mm. The distance between the pitch and first R EFERENCES
joint axes is l0 = 22 mm. In the experiments, we set up
[1] D. Prattichizzo and J. C. Trinkle, “Grasping,” in Springer Handbook
As = 25 mm, ωs = π5 rad s−1 , λo = 40 s−1 and λr = 10 s−1 . of Robotics, B. Siciliano and O. Khatib, Eds. Springer Verlag, 2008,
The experimental results are shown in Figs. 3-5, where pp. 671–700.
it is possible to observe the time history of the trajectory [2] A. Bicchi, “Hands for Dexterous Manipulation and Robust Grasping:
a Difficult Road Toward Simplicity,” IEEE Trans. Rob. Aut., vol. 16,
tracking, the relative position regulation, the position errors, no. 6, pp. 652–662, Dec 2000.
and the control signals respectively. From the analysis of [3] J. Aleotti, D. Lodi Rizzini, and S. Caselli, “Perception and Grasping
the results, we conclude that the proposed control scheme of Object Parts from Active Robot Exploration,” J. Intell. Rob. Syst.,
vol. 76, no. 3-4, pp. 401–425, 2014.
was able to ensure the asymptotic convergence of the object [4] J. Butterfass, G. Hirzinger, S. Knoch, and H. Liu, “DLR’s Multisensory
position error to zero. Besides, it can be noticed that the Articulated Hand Part I: Hard- and Software Architecture,” in Proc.
relative position converges to the desired values, ensuring the of IEEE Int. Conf. Rob. Aut., May 1998, pp. 2081–2086.
[5] C. Lovchik and M. Diftler, “The Robonaut Hand: a Dexterous Robot
satisfactory prehension of the object (Fig. 4). Finally, from Hand for Space,” in Proc. of IEEE Int. Conf. Rob. Aut., May 1999,
the analysis of the analytical expression for the Jacobian pp. 907–912.
matrices, J˜hp and J˜r , we concluded that there are no occur- [6] W. T. Townsend, “The Barrett Hand Grasper - Programmably Flexible
Part Handling and Assembly,” Ind. Rob.: An. Int. J., vol. 27, no. 3,
rences of singularities in the task space. This analysis was pp. 181–188, 2000.
omitted here for the sake of space saving, but the expected [7] H. Kawasaki, T. Komatsu, and K. Uchiyama, “Dexterous Anthropo-
results was confirmed by means of exhaustive experiments. morphic Robot Hand with Distributed Tactile Sensor: Gifu hand II,”
IEEE/ASME Trans. on Mech., vol. 7, no. 3, pp. 296–303, Sep 2002.
[8] P. Tuffield and H. Elias, “The Shadow Robot Mimics Human Actions,”
IV. CONCLUSIONS AND FUTURE WORKS Ind. Rob.: An Int. J., vol. 30, no. 1, pp. 56–60, 2003.
In this work, we present a kinematic framework for mod- [9] H. Liu, P. Meusel, N. Seitz, B. Willberg, G. Hirzinger, M. Jin, Y. Liu,
R. Wei, and Z. Xie, “The Modular Multisensory DLR-HIT-Hand,”
eling and control design of a multifingered robot hand, which Mech. Mach. Theory, vol. 42, no. 5, pp. 612–625, May 2007.
is based on an empirical model and a control formalism [10] C. Connolly, “Prosthetic hands from Touch Bionics,” Ind. Rob.: An
developed for general robotic systems subject to kinematic Int. J., vol. 35, no. 4, pp. 290–293, 2008.
[11] C. Melchiorri and M. Kaneko, “Robot Hands,” in Springer Handbook
constraint conditions [21]. Experimental results, obtained of Robotics, B. Siciliano and O. Khatib, Eds. Springer Verlag, 2008,
with the 3-Finger Adaptive Robot Gripper from Robotiq, are pp. 345–360.
shown to illustrate the performance and effectiveness of the [12] R. Tinos, M. Terra, and J. Ishihara, “Motion and Force Control of
Cooperative Robotic Manipulators with Passive Joints,” IEEE Trans.
proposed methodology. Due to the presence of rigid links Contr. Syst. Tech., vol. 14, no. 4, pp. 725–734, Jul 2006.
and springs, each articulated finger can be modeled as a [13] G. M. Freitas, A. C. Leite, and F. Lizarralde, “Kinematic Control of
parallel mechanism and we empirically derive the kinematic Constrained Robotic Systems,” SBA Contr. Aut. Mag., vol. 22, pp. 559–
572, Dec 2011, http://dx.doi.org/10.1590/S0103-17592011000600002.
relationship between the velocities of the motor shafts and [14] I. Kao, K. Lynch, and J. W. Burdick, “Contact Modeling and Manipu-
the fingertips. We also model the robot hand during the lation,” in Springer Handbook of Robotics, B. Siciliano and O. Khatib,
constrained motion as a parallel mechanism and the contact Eds. Springer Verlag, 2008, pp. 647–669.
[15] T. Yoshikawa, “Multifingered robot hands: Control for grasping and
points between the fingers and the object as virtual revolute manipulation,” Ann. Rev. Contr., vol. 34, no. 2, pp. 199–208, 2010.
joints. Similarly to the case of free motion, we derive the [16] T. Takahashi, T. Tsuboi, T. Kishida, Y. Kawanami, S. Shimizu,
velocity relationship between the object and the motor shafts M. Iribe, T. Fukushima, and M. Fujita, “Adaptive Grasping by Multi
fingered Hand with Tactile Sensor based on Robust Force and Position
under the full knowledge of the object geometry. Control,” in Proc. IEEE Int. Conf. Rob. Aut., May 2008, pp. 264–271.
The position control problem is tackled by using a kine- [17] E. Engeberg and S. Meek, “Adaptive Sliding Mode Control for
matic control approach, which ensures the asymptotic stabil- Prosthetic Hands to Simultaneously Prevent Slip and Minimize De-
formation of Grasped Objects,” IEEE/ASME Trans. Mech., vol. 18,
ity of the output tracking error and a satisfactory performance no. 1, pp. 376–385, Feb 2013.
when the task motions are executed with low velocities and [18] Y. Yin, Z. Luo, M. Svinin, and S. Hosoe, “Hybrid Control of Multi-
slow accelerations. Since friction effects were not taken into fingered Robot Hand for Dexterous Manipulation,” in Proc. IEEE Int.
Conf. Syst., Man and Cyber., vol. 4, Oct 2003, pp. 3639–3644.
account in the proposed methodology, the kinematic control [19] S. Arimoto, “Intelligent Control of Multi-fingered Hands,” Ann. Rev.
approach without explicit closure of a force feedback loop is Contr., vol. 28, no. 1, pp. 75–85, 2004.
not able to guarantee the satisfactory prehension of the object [20] Y. Zhao and C. C. Cheah, “Neural Network Control of Multifingered
Robot Hands Using Visual Feedback,” IEEE Trans. Neur. Net.,, vol. 20,
during the manipulation task. To overcome this drawback, we no. 5, pp. 758–767, May 2009.
have combined the object position control with the relative [21] J.-Y. Wen and L. Wilfinger, “Kinematic Manipulability of Gen-
position control for the fingers, ensuring a firm grasp of the eral Constrained Rigid Multibody Systems,” IEEE Trans. Rob. Aut.,
vol. 15, no. 3, pp. 558–567, Jun 1999.
object. The position control scheme works better with soft- [22] B. Siciliano, L. Sciavicco, L. Villani, and G. Oriolo, Robotics: Mod-
finger contact models due to the frictional forces that arise elling, Planning and Control. Springer-Verlag London, 2009.
during the compression of the object. The theoretical results [23] F. Caccavale and M. Uchiyama, “Cooperative Manipulators,” in
Springer Handbook of Robotics, B. Siciliano and O. Khatib, Eds.
hold in the practical experiments under the assumptions of Springer Verlag, 2008, pp. 701–718.
full knowledge of the mechanism kinematic parameters and [24] Robotiq, Specification sheet for the 3-Finger Adaptive Robot Gripper,
no occurrences of singular configurations. Motivated by the Aug 2015, http://robotiq.com/products/industrial-robot-hand/.

164