582

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 32, NO. 5, SEPTEMBER 2002

Motion in Human and Machine:

A Virtual Fatigue Approach
Veljko Potkonjak, Dragan Kostic, Student Member, IEEE, Milan Rasic, Student Member, IEEE, and
Goran S. ordaevic, Member, IEEE

AbstractAchieving human-like behavior of a robot is a

key issue of the paper. Redundancy in the inverse kinematics
problem is resolved using a biological analogue. It is shown that
by means of virtual fatigue functions, it is possible to generate
robot movements similar to movements of a human arm subject
to muscle fatigue. Analytic method enabling control of robot
motions in a human-like fashion is described. An example of an
anthropomorphic robot arm performing a screw-driving task
illustrates the method.

robots supports the expectations that future personal robots

will be of anthropomorphic structure. These robots will live
among humans, hence, they should feature human-like characteristics in their behavior (regarding motion, communication
and intelligence).
Objectives of this paper are twofold: 1) to investigate the
problem of human-like behavior in the sense of motion and
2) to advise how to achieve human-like motions in robots.

Index TermsHumanmachine analogy, inverse kinematics,

redundancy.

II. HUMAN-LIKE MOTION: BACKGROUND AND NEW APPROACH

I. INTRODUCTION

NDUSTRIAL robots perform various tasks improving

the quality and efficiency of manufacturing. They also
substitute human workers in tasks that may impact humans
health and safety. Despite a long history of industrial robots,
their usage still requires a well-structured working environment where physical presence of humans is not allowed [1].
Considering the fast-paced development of computer hardware
and sensor techniques, as well as the design of advanced
nonlinear control strategies that substantially improve the
intelligence of robots, it seems reasonable to expect a direct
cooperation between human workers and robots in the near
future. Their mutual safety will be a basic condition of direct
cooperation inside a common working space. Apart from
safety requirements, it is also important to provide comfortable
working conditions, since they may considerably influence
the workers productivity. During future cooperation with a
robot, the humans feeling of comfort will not be dependent
solely on the specific nature of a working task, which means
specific operation procedure and working dynamics, but also
on psychological impressions with respect to the robotized
companion. This is why robots should adapt to ordinary human
feelings and behave in a human-like manner.
A new and growing field of robot application is service
robotics. Within this class, the most promising perhaps, are
home or personal robots. The fast growing interest in humanoid
Manuscript received April 16, 2000; revised February 28, 2002. This paper
was recommended by Associate Editor R. A. Hess.
V. Potkonjak is with the Faculty of Electrical Engineering, University of Belgrade, 11000 Belgrade, Yugoslavia (e-mail: potkonjak@www.mail.yu).
D. Kostic is with the Department of Mechanical Engineering, Eindhoven University of Technology, Eindhoven, The Netherlands (e-mail: D.Kostic@tue.nl).
M. Rasic is with the Faculty of Electronic Enginering, University of Nis,
18000 Nis, Yugoslavia (e-mail: mrasic@elfak.ni.ac.yu).
G. S. ordevic is with the Faculty of Electronic Enginering, University of
Nis, 18000 Nis, Yugoslavia (e-mail: gorandj@elfak.ni.ac.yu).
Digital Object Identifier 10.1109/TSMCA.2002.804792

The stated objectives can be naturally achieved by designing

the robots that imitate humans (or animals) both morphologically, and as to the way they execute motions. Research in this
field is related to a so-called bio-robotic approach. In this approach, solutions to various technical problems resemble the
patterns recognized in nature. These patterns may concern kinematic configuration, shape, mechanical construction, etc. Here
we consider biological patterns providing human-like behavior
of a robot.
A. Kinematic Redundancy
From a mechanical point of view, a robot resembling a
human should be kinematically redundant, i.e., its mechanism
should feature a higher degree of mobility than required for
a given motion in operational space. Kinematic redundancy
contributes to robot dexterity and facilitates coping with unpredictable changes within its environment. It enables avoidance
of mechanical limits of robot joints [2], obstacle avoidance
[3], singularity avoidance [4], fault-tolerant operation [5],
optimization of robot dynamics [6], [7], human-like distribution of joint motions [8][14], etc. Unfortunately, redundancy
usually increases the mathematical complexity of the robot
control problem. Its implications are particularly emphasised
in the well-known inverse kinematics (IK) problem. This
is a problem of searching for joint motions that provide a
desired trajectory of a robot end-effector in operational space.
The presence of kinematic redundancy means that the same
end-effectors trajectory can be executed with different joint
motions. Hence, the problem of the particular choice between
available joint motions arises. Conventional methods to handle
this problem are divided into two groupsglobal and local,
upon the criterion of what one needs to know in advance about
the operational space trajectory in order to find the appropriate
IK solution [1]. Both groups of methods adopt some additional
performance criterion (e.g., energy consumption) and search
for joint motions that execute the desired end-effectors trajectory along with optimzing the criterion. However, global

1083-4427/02$17.00 2002 IEEE

POTKONJAK: et al. MOTION IN HUMAN AND MACHINE: A VIRTUAL FATIGUE APPROACH

methods require knowledge about the complete end-effectors

trajectory in advance, since they look for joint motions that
provide a global optimum for the adopted criterion. On the
other hand, local methods perform only a local optimization
of the criterion and just call for information about the actual
pose of the end-effector. The choice between global and local
methods is task-dependent. If a robot is going to operate in
a well-structured working environment, performing some
predefined operational space trajectory, then global methods
are preferable [1], [15]. When environmental conditions are not
controllable and feature unpredictable changes, local methods
are inevitable. The latter case requires maximal autonomy from
the robot in both planning and executing its motions. Such a
case also corresponds to future mutual cooperation between
humans and robots, which presumably, will have plenty of
possible variations still yet unknown. Therefore, for on-line
robot operation, and especially in the physical presence of
humans, local methods for redundancy resolution are required.
Some characteristic local methods are the following: gradient
projection methods [1], [2], extended Jacobian method [16],
configuration control [17], Lyapunov function based methods
[18], distributed positioning (DP) concept [8][14] and the inverse kinematic function approach [19], [20].
B. Human Arm as a Redundant Manipulator
This paper suggests one local method for redundancy resolution resulting in human-like motions of the robot joints. The
subject of the theory can be illustrated considering a human arm
as an example. It is an ultimate model of highly accurate and
dextrous motions, due to the presence of kinematic redundancy.
Imitating movements of a human arm, robot performance can
be improved substantially. Benefits of human-like motion, however, cannot be gained without previous knowledge of the principles that the central nervous system (CNS) employs during
the generation of arm movements. It is particularly important to
understand how the CNS resolves the redundancy problem in
the arm, by distinguishing one solution for joint motions from
an infinite set of possible motions. Unfortunately, a thorough
explanation of such principles is not yet known [21]. In this
paper, we rely on a hypothesis already presented in [22] and
suggest its application in robotics. According to that hypothesis, it is possible to associate some cost function to each joint
of a human arm. The arm performs movements that optimize
these functions. Joint movement time is an example of a kinematic cost function. Examples of dynamic cost functions are
the following: quadratic norm of joint control torques [7], kinetic energy [23], jerks in joints (third derivative of joint position) [24]. Several neuro-physiological and psychophysical cost
functions were also suggested [25]: input energy was defined
as a quadratic norm of input neural signals of motor units (muscles), while input fatigue denote the magnitude of such neural
signals. The authors of [25] suggest some proper combination
of these functions, rather than their separate application. In this
study, special attention is paid to functions of joint discomfort,
which were experimentally derived to identify arm postures of
maximum comfort [22]. They were determined upon analysis of
recorded electromiographic (EMG) signals taken from subjects

583

engaged in experiments, as well as by using their subjective psychophysical evaluations of maximum comfort postures.
The fact that a variety of cost functions has already been used
to explain principles of human arm motor control indicates that
the CNS does not obey any one particular cost function, but also
does not violate general physical and technical principles of optimality, from which particular cost functions come about [26].
Hence, additional efforts in searching for new appropriate and
effective cost functions are justified. They contribute to a better
understanding of biological principles of motor control. The majority of the above mentioned cost functions are of integral type,
meaning that they are supposed to be used with global optimization methods. This paper suggests a mathematical function appropriate for on-line use in robotics. It is of local nature, suitable for generation of human-like motions in robot joints and
especially in those tasks where substantial robot autonomy is
expected.
C. Virtual Fatigue Idea
The underlying idea of the paper has both theoretical and
experimental foundations. Practical experience shows that the
human arm commonly takes those postures and executes those
movements that are the most comfortable. The term comfortable relates to joint positions and engagement of motor units
and may also be described by the term pleasant (a more precise definition will be given later). On the other hand, endurance
contractions of motor units cause muscle fatigue, thus introducing an unpleasant feeling, that is, a sense of discomfort. In
everyday life it is easy to observe that after a sensation of discomfort caused by muscle fatigue, the human arm normally reduces engagement of the fatigued motor units, by taking postures that require lower participation of these units. This means
that while performing repetitive movements requiring continual
repetition of motions in operational space (like in screw-driving
tasks), the human arm occasionally reconfigures itself by taking
a more comfortable posture, rather than proceeding with some
particular pose. The ability to rearrange its motion is enabled by
the presence of both actuator and kinematic redundancy in the
human arm [27]. Actuator redundancy comes from the possibility to use several motor units for the same motion of any arm
joint. Kinematic redundancy results from the existence of seven
degrees of freedom (DOFs) in the arm (from shoulder to wrist),
which is more than six independent movements required for an
arbitrary positioning and orientation of an object in operational
space [14], [18]. Actuator redundancy and its implementation
in robotics, are challenging problems that deserve particular attention. They are not considered in this paper, although their
role in performing movements in biological mechanisms must
be pointed out. Instead, this paper focuses on kinematic redundancy and investigates possibilities to distribute the engagement
of robot joints in a human-like fashion, imitating the arms inherent property to execute comfortable motions. The main objective is to achieve a human-like motion. This can be done if
an adequate mechanism is established that simulates biological
processes of comfort and discomfort in the arm. It would be
useful to rely on relevant findings from already published results
of theoretical and experimental investigations. A result which

584

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 32, NO. 5, SEPTEMBER 2002

is strongly correlated with our work deals with psychophysical

cost functions of joint comfort/discomfort and is presented in
[22]. Their validity was practically justified for arm reach posture prediction in [28]. A psychophysical cost function describes
an immediate deviation of joint position from the location of
maximum comfort. According to experimental findings given
in [22], the CNS controls arm motion by minimizing the efforts
(from a psychophysical point of view) invested during the movements. Physiological and psychophysical investigations indicated that, in the absence of muscle fatigue, a more comfortable
joint pose is closer to the middle of the physiological motion
range in that joint. Locally minimizing the function describing
a deviation from the position of maximum joint comfort, it is
possible to determine comfortable motions of a kinematically
redundant mechanism.
Mathematical functions representing current distances from
middle positions of joints are used in robotics for joint limits
avoidance [2], [29]. The applied IK method takes care of these
distances and forces joint motions to the direction opposite to
mechanical boundaries. In this paper, such functions are chosen
as starting points in the formulation of an analytic procedure for
generating joint movements, that are equivalent to the movements of a human arm after appearance of discomfort due to
muscle fatigue.
Physiological sources of fatigue, although extensively
studied, are still not thoroughly known. Basically, fatigue appears after long-standing and powerful contractions of muscle
motor units. Increase of lactic acid concentration accompanies
the progress of fatigue sensation (pH value decrease in muscle
tissue). Simultaneously, the oxygen distribution is reduced,
while concentration of some substances particularly influencing the mechanism of muscle contractions and dilatations
decreases, e.g., Adenosine Triphosphate (ATP). As a result,
muscular activity declines. The progress in a humans feeling
of discomfort due to fatigue grows simultaneously with the
progress of fatigue itself. The beginning manifestation is like a
slight sense of discomfort in a certain part of the arm, then the
discomfort transforms into an unpleasant squib which, finally,
results in obtuse pain [30]. Additional engagement of other
motor units is then required to sustain the necessary muscle
force.
Muscle fatigue can be quantified by means of objective and
subjective methods. Objective methods include mechanical,
electromagnetic (EMG), metabolic and physiological measurements [31]. Another group of methods is based on the
subjective evaluation of the sensed fatigue level, given by the
subjects participating in experiments [30]. Because there is a
variety of factors indicating the current level of fatigue, it is not
possible to distinguish an ultimate method for fatigue quantification. The same statement holds for the models of fatigue,
available in literature (see, for example, [32]). However, no
matter which method is applied for fatigue quantification, it
seems reasonable to consider fatigue as an increasing function.
That function is often assumed to be exponential [33][35].
The slope of the function depends on the actual engagement of
motor units and the current level of fatigue. After some time,
saturation appears, as a result of reduced activity of exhausted
motor units. An example of a diagram with such characteristics

is available in [36] and corresponds to the increase of lactic acid

concentration in a muscle engaged in demanding movements.
Keeping in mind the described principal characteristic of the
fatigue function, we shall suggest for a mathematical nondimensional variable that can be used as an analogue of biological fatigue. That variable should describe a virtual fatigue of robot
actuators, while its temporal characteristic must be equivalent to
the functional characteristic of biological fatigue. The temporal
characteristics of the virtual fatigue could then be used for the
generation of human-like motions of the robot joints. The aim is
to force a redundant anthropomorphic robot arm to track a given
end-effector operational space trajectory, along with producing
the most comfortable configurations in the sense of the above
mentioned psychophysical cost function. The time history of
robot virtual fatigue should also resemble the biological one.
The model of virtual fatigue will be derived based on a model
of thermal dynamics of dc motors. The dynamic behavior of the
motor thermal model will be deliberately accelerated and the
faster coordinate of the resulting model will take a role of virtual
fatigue. Functionally, virtual fatigue will have a response equivalent to the biological muscle fatigue, that is, similar dynamic
behavior. Results presented in the rest of the paper will justify
this approach. An anthropomorphic seven-DOF robot arm performing the screw-driving task is simulated. It will be shown
that the robot arm attains postures and executes motions similarly to that of the human arm performing the screw-driving
task.
The rest of the paper is organized as follows. Section III gives
the mathematical description of the method for the generation
of human-like motions of robots. Particularly, an appropriate
solution of the IK problem is suggested. Then, candidates
for functions of psychophysical discomfort are presented.
The derivation of the virtual fatigue models follows. In the
last part of Section III, penalty functions are utilized for the
required human-like control of robot arm motion. The quality
of the suggested approach is verified by means of computer
simulations and results are given in Section IV. Final remarks
are given in the Conclusion. Numerical data of the models is
given in Appendixes A and B.
III. MATHEMATICAL FORMULATION
A. Solution of IK Problem
The robot arm kinematics will be defined in terms of velocities [1], [18]. The relation between vectors of configuration
(joint) velocities and operational (end-effector) velocities ,
is given by the Jacobian form
(1)
It is assumed that redundancy exists, i.e., the number of operational velocities, denoted by , is strictly less than the number
of configuration velocities, denoted by . Dimension of the nonis then
. The redundancy
square Jacobian matrix
implies a nonunique IK solution, since a given motion task,
, can be accomdefined in terms of operational velocities
plished with an infinite number of configuration velocities .
We are interested in those joint velocities that would be executed by a human arm in the same task. Adequate velocities can

POTKONJAK: et al. MOTION IN HUMAN AND MACHINE: A VIRTUAL FATIGUE APPROACH

be found by local minimization of the cost function, formed by

two quadratic terms [18]
(2)
and
denote
positive definite symmetric weighting
represents an -component column vector.
matrices, while
The first term enables us to penalize motions of some joints relative to others. In this paper it should provide a distribution of
joint motions in accordance with the the concept of distributed
positioning (DP) [8][14], which means to stimulate motions of
the joints with low inertia and penalise motions of joints with
high inertia. It should also enable a proper reconfiguration of
the robot arm, in accordance with the progress of virtual fatigue. The second term aims at the utilization of kinematic redundancy in the sense of a secondary criterion. Minimization of
objective (2) will be performed using the method of Lagrange
multipliers [37] to incorporate constraint (1). The Lagrangian
corresponding to objective (2) and kinematic constraint (1) is
given by

(3)

585

Vector
enables the local optimization of some secondary ob, used for the proper utilization of kinejective function
is defined as the gradient
matic redundancy. Following [2],
:
of
(12)
is a real scalar coefficient. The final form of the IK
where
solution is obtained by substituting (12) into (10):
(13)
B. Choice of the Secondary Objective Function
is discussed
Choosing the secondary objective function
in this subsection. The distances of current joint positions
from the mechanical joint limits
and
will
be the basis for definition of the secondary objective function.
In the previous section it was already pointed out that the middle
values of human arm joint ranges coincide with positions of the
as a
maximum comfort. This fact justifies the choice of
function penalizing deviation from the middle values [2], [29]

1 vector of Lagrange multipliers. The

where denotes an
required IK solution has to satisfy the necessary conditions:
(14)
and

(4)

or

The first conditions results in:

(15)
(5)

which can be rewritten as follows:

C. Complete Model of Robot Arm Dynamics

(6)
Substitution of (6) into (1), gives:

(16)
(7)

Solving for

The dynamics of the robot arm plays an important role in

establishing a procedure that provides human-like motions of
robot joints. Consider the standard representation of robot dynamics [39], [40]

yields
(8)

Using (8) in (6) to eliminate , one obtains

1 vectors of joints positions, vewhere , and denote

inertia matrix, is an
locities and accelerations, is an
1 vector of centripetal, Coriolis, friction and gravitational
1 vector of driving torques in
torques, while denotes the
the joints. Driving torques are produced by actuators. We are interested in dc drives as joint actuators. The model of an armature
controlled dc motor, acting at the th joint, is given by

(9)
(17)
which can be rewritten in a more compact form:
(10)
denotes the weighted pseudoinverse of the Jacobian
where
matrix [38]
(11)

,
,
and
denote armature voltage, current,
where
and
denote
resistance, and inductance, respectively,
constants of back electromotive force and torque, respectively,
is the central moment of inertia of the rotor, and
is
the gear-box ratio. Inclusion of the actuator model (17) into the
model of the robot mechanical structure (16), provides a complete model of robot arm dynamics.

586

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 32, NO. 5, SEPTEMBER 2002

D. Virtual Fatigue: Second-Order Model

This subsection suggests a method that should be used in
robots to simulate the effects of biological fatigue of human
muscles. The expectation is that such a method might provide
human-like movements of robot joints. By means of the
, it is possible to establish a functional
weighting matrix
relation between joint motions and time history of the virtual
fatigue. At this point a new question arises: how to formulate
virtual fatigue? Practically speaking, the question is which
physical variable related to a robot could be considered as an
equivalent to biological fatigue. Intuitively, the temperature of
dc motor armature winding seems an appropriate candidate.
The thermal dynamics of a dc motor can be described by a
second-order model [40], [41]

(18)
and
denote temperatures of the armature windwhere
is the ambient temperings and the housing, respectively,
and
are thermal capacities of the windings and
ature;
and
are windings-to-housing
the housing, respectively;
and housing-to-ambient thermal conductances; the term
indicates power loss in armature windings. Temperatures of the
armature windings and the housing are the state coordinates in
the given model. They are governed by the magnitude of power
loss, which depends on the instantaneous value of armature current. Current is a state coordinate of the complete model of the
robot dynamics (16), (17).
With respect to the model of thermal dynamics (18), it should
be pointed out that motor catalogues usually do not offer values
and
, nor of thermal conneither of thermal capacities
and
, directly. These parameters are calcuductances
lated using the catalogue values of rotor and housing thermal
and
), as well as thermal time constants
resistances (
and
). The following
for the rotor and the housing (
equations are applied [41], [42]:

ture heating with constant power loss, taken from catalogue

[42], is as follows:
(20)
denotes the equivalent thermal time constant of the
where
dc motor. The above expression shows the exponential nature of
the armature temperature step response, which is analogous to
the exponential nature of fatigue models given in, e.g., [34], [35].
According to the above analysis one may draw the conclusion that a time history of armature temperature is equivalent
to the functional characteristic of biological fatigue, which can
recommend it for the role of virtual fatigue. However, such a
choice requires a more comprehensive analysis. Namely, a critical criterion in servo sizing (choosing appropriate dc motors) is
whether the selected actuators can produce the torques needed
in the most demanding tasks [40], [41]. This means that in practice robot actuators should be properly dimensioned, preventing
the possibility of failures due to overheating. Hence, critical armature temperatures are not reached in practice. In addition to
this, the temperature dynamics of a dc motor operating within
the recommended operating range is still slower than the rate
of biological fatigue increase. These facts show that the desirable functional characteristic of armature temperature is compromised by unsatisfactory slowness of its dynamics. This in
turn disputes the appropriateness of armature temperature as an
equivalent to biological fatigue.
The above problem can be resolved by an artificial acceleration of the dc motor thermal dynamics, deliberately using lower
values of thermal time constants and higher values of thermal
resistances, instead of the original ones. Reduced thermal time
constants give a faster change of the states in the model (18).
On the other hand, thermal resistances influence the steady
state values of the temperatures. Higher thermal resistances
provide higher steady state temperatures. Hence, the level of
the steady-state temperatures can be adjusted by means of the
thermal resistances.
In accordance with these remarks for speeding up the DC
motor thermal dynamics, as well as for scaling the steady-state
magnitudes of motor temperatures, it is possible to formulate a
model of virtual fatigue in the th robot joint:

(19)
The temperature of armature windings is a critical value in
the thermal model (18), since it has faster dynamics than the
temperature of the housing. This is due to the considerably
lower value of the armature thermal capacity with respect to
the thermal capacity of the housing [42]. For constant angular
velocity of the motor shaft and assuming a constant load torque,
the temperature of the armature windings rises until a balance
between power loss, on one hand and dissipation of heat from
the armature through the housing into the air, on the other, is
established. With a balance between power loss and dissipation,
armature temperature reaches a steady state. Qualitatively,
such temperature dependence resembles a time history of the
increase of lactic acid concentration due to fatigue progress in
muscles. Furthermore, the analytical expression for the arma-

(21)
where variable denotes the virtual fatigue, while is an auxiliary variable. Parameter enables the set up of a reference level
and
are defined
for the virtual fatigue. Time constants
and
by the desired rates of and , while parameters
are chosen in accordance with preferable magnitudes of and
. For some particular motor, as starting values for tuning these
parameters, one may take motor thermal time constants and
thermal resistances from the catalogue. It means that appropriate
and
are obtained by an adequate decrease of
values of
and
, respectively, while
motor time constants
and
should attain higher values than motor thermal resisand
.
tances

POTKONJAK: et al. MOTION IN HUMAN AND MACHINE: A VIRTUAL FATIGUE APPROACH

E. Virtual Fatigue: First-Order Model

The variable , describing virtual fatigue in the -the robot
joint, is obtained by solving (21). Now, another question arises.
Is it necessary for the model of virtual fatigue to be of the second
order? In other words, is it possible to reduce the model order,
or, should the order be increased? The latter certainly does not
seem to be attractive since it contributes to the mathematical
complexity of the problem. A reduction of model dimension is
preferred, since it enables easier and more efficient calculations,
as well as better control over virtual fatigue time history. A reduced-order model of virtual fatigue in th robot joint is given
by
(22)
denotes the state. The reference value of the virtual
where
fatigue can be set by . Model parametersscalar gain
and time constant
, should be determined such that the dynamic behavior of virtual fatigue from the reduced order model
(22) is close to the dynamic behavior of the equivalent variable from the original model (21). Naturally, we can expect
nothing more than simple similarity between these two coordinates, since identical behavior is certainly not possible. For
simulation purposes (results will be presented in the next secand
were identified using the graphotion) parameters
analytical method of multiple integrations [43]. This method
provides the following expressions:

(23)
denotes the steady-state value of the second-order
where
model virtual fatigue response, assuming constant power loss
(24)
Suggesting proper means for establishing the functional dependence between virtual fatigue time history and robot arm
motions should conclude this section. For this purpose, we
should reinstate the expected effects of virtual fatigue usage.
During manipulation, a human arm performs movements
adequate to the desired manipulation task, permanently accommodating its configurations to the actual level of muscle
fatigue. Present kinematic and actuator redundancy allows
execution of the manipulation task in a comfortable way, by
appropriate distribution of joint motions and participation of
different motor units. In such ways exhausted muscles may
recover and other muscles increase their activity. A similar
strategy could be applied for a robotic arm. However, capabilities of the kinematic redundancy will be utilized only. Actuator
redundancy is out of scope of this paper.
F. Application of Penalty Functions
For each robot joint it is necessary to specify an appropriate
of the virtual fatigue. If the value of is less
critical level
, then an IK solution should provide merely the most
than
comfortable pose, from a psychophysical point of view. When
virtual fatigue reaches, or exceeds a critical level, appropriate

587

depression of joint motion is required. Such control of joint motions can be achieved by means of the weighting matrix . Reis formed by simple summing of two matrices,
ferring to (6),
and
. Matrix
should provide human-like distrii.e.,
bution of joints motions (DP concept) and an adequate reconfiguration of robot arm with respect to actual levels of joint virtual
one can specify higher engagement of
fatigue. By means of
some joints in realization of the secondary objective, i.e., local
given by (14) and (15). In this paper, it
minimization of
is assumed that all joints have equal priority in realization of
( dethe secondary objective, which in turn gives
determines
notes identity matrix). In such a way, the role of
, in the sense that
should provide
a particular choice of
. This can be
a behavior in accordance with the objective of
achieved with the following choice of :
diag

diag

(25)

are positive constants, while

where weighting coefficients
are constant until reaches
and monofunctions
. Having in mind the role of
tone nondecreasing above
in the basic cost functional (2), it follows that 1) weighting cocan be used for initial distribution of joint involveefficients
contribute to
ment according to DP concept; 2) functions
reduced movement of each joint where virtual fatigue exceeds
an assigned limit.
are positive and monotone nondecreasing.
Functions
They are required to reduce motions in joints after magnitudes
of related virtual fatigues reach critical values. Control of joint
motions can be achieved using the technique of penalty functions [37]. Choice of particular penalty function is task dependent. For the simulation studies considered in the next section,
a quadratic form of the penalty function is used
(26)
determines the desired slope of the
where scalar
penalty function.
Finally, we should mention one potentially delicate point of
the suggested control algorithm. Speeding up the motor thermal
dynamics by deliberate modification of model parameters
may cause inconvenient oscillatory behavior in virtual fatigue
time histories. Oscillations are unavoidable characteristic of
temperature diagrams of dc motors operating in dynamic mode
[41]. Nonuniform motor speeds and changeable load torques
cause them. Suppose that cyclic end-effectors trajectory results
in cyclic joint motions and consider engagement of one joint
motor. In some phases of joint motion, a higher output torque
is required from the motor. This in turn requires injection
of higher armature current, which results in higher power
loss. This increases temperature of armature windings for
some amount. In other phases of joint motion, engagement of
motor is reduced, as well as requirements on armature current
magnitude. Consequently, power loss is decreased causing a
short-term drop of armature temperature. Short-term temperature drop is not enough to redirect the slope of temperature
diagram. Average value of armature temperature proceeds with
monotone climbing until a steady state, while its actual value
feature high-frequency oscillations around it.

588

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 32, NO. 5, SEPTEMBER 2002

Fig. 1. Anthropomorphic robot arm in screw-driving task.

Such oscillations are emphasized in virtual fatigue models

(21) and (22). They act as a kind of disturbance in the control algorithm presented above. Especially, they may negatively affect
the penalty function (26). That is why their influence should be
suppressed. A viable solution is filtering using a lowpass filter
(27)
and
denote actual and filtered virtual fatigue
where
can be
values in Laplace domain, respectively. The filtered
.
employed in (26) instead of
This concludes the mathematical procedure for redundancy
resolution that should result in human-like motions of robot
joints. In the next section, quality of the suggested procedure
is evaluated using simulations.

IV. EXAMPLE: SCREW-DRIVING TASK

Effects of the suggested method are analyzed for the screwdriving task. This task is typical for industry and is an unavoidable part in assembling, where almost 25% of all operations are
screw insertions [44]. It should be noted that industrial robots
perform the screw-driving task in a way conceptually different
from that applied in our simulations. The main role of industrial robot is to place a screw-driving tool in an appropriate position with respect to the target screw and not to provide further driving movements. The screw-driving tool itself generates
the additional revolute movements. This means that the robot
is merely used as a space positioner. Having in mind our main
idea, achieving human-like motion, we require that the robot
performs screw driving without any automated tool, thus putting
it in the position of a manual worker.

An anthropomorphic robot arm is considered in the screwdriving task (Fig. 1). It has seven DOFs, meaning that it is redundant for the given task. Its kinematic scheme is a simplified model of a human arm excluding the hand [12][14]. Such
a kinematic scheme is deliberately chosen. An anthropomorphic model is the most appropriate for testing and evaluation
of human likeness. In the previous paragraph, it was pointed
out that the robot arm is not going to perform screw driving
in the manner of industrial robots. It will perform the task in
a human-like fashion, which means that apart from enabling the
required position of the screwdriver in the space, the robot arm
should also provide angular screw-driving movements about the
longitudinal screwdrivers axis (Fig. 1). Human-like conception
of screw driving enables a direct comparison of motions in humans and robots.
Robot arm kinematics can be described using the notation of
Denavit and Hartenberg (DH) [18]. Joint coordinate frames, assigned according to DH convention, are indicated in Fig. 1. The
values of DH parameters are given in Table I, in Appendix A. In
the same Appendix A one can find Table II, containing: kinematic and dynamic parameters of the considered robot arm,
weighting coefficients that provide DP concept in (25), parameters of the penalty functions (26) and parameters of the low-pass
filters (27). Finally, parameters of the dc actuators and virtual fatigue models (21) and (22), are presented in Table III given in
Appendix A.
are indicated.
In Fig. 1, independent joint coordinates
These coordinates do not coincide with rotational angles
according to DH notation. The coordinates in Fig. 1. are more
explicit and thus easier to observe. They are related to DH angles
by

(28)

POTKONJAK: et al. MOTION IN HUMAN AND MACHINE: A VIRTUAL FATIGUE APPROACH

589

Initial position of the robot arm is defined with respect to

(29)
, in its initial posture the robot arm has a
According to
stretched forearm (wrist) including the screwdriver.
A screw-driving task requires rotational motion of a screw
about its longitudinal axes. It is assumed that screwing consists
of a series of single revolute movements (forward and back[rad]. A forward rotation (screwing
ward), each of exactly
in) is indicated in Fig. 1. Backward rotation does not drive a
screw, but only brings screwdriver in the initial position. Each
s. Different studies investigating
movement takes
motor control of human movements have shown that human
arm performs smooth voluntary movements with bell-shaped
velocity profiles (for illustration, see [24]). In our simulations
the bell-shaped velocity profile was approximated with a cosine
velocity profile [6] and applied to rotational motions of the
screwdriver. This defines the motion task of the end-effector.
In the first simulation, the force and the torque applied to
the screw are not taken into account. The IK solution is calculated according to the procedure described in the previous section. Underlying idea is to achieve comfortable joint motions
(closer to the middle locations of joints ranges), while taking
care of the actual levels of joints virtual fatigue. When virtual
fatigue in some joint exceeds a critical limit, the joint engagement should be reduced. It is followed by higher involvement of
other joints. Distribution of joints involvement should be similar with a human arm.
Different diagrams are obtained by simulating the anthropomorphic robot arm in the screw-driving task. Simulation started
with a virtual fatigue model of second-order (21). Joint motions
,
calculated by suggested redundancy
resolution procedure, are presented in Fig. 2 (the order 1, 3,
2, instead of regular 1, 2, 3, uses the space in the figure
more economically). Note that total time of monitoring the event
was 500 s. Since a cycle due to screwing-in (forward-plus-backward rotation) lasts 1 s, it is not possible to observe a particular oscillation in joint motions. One can only see the envelope,
but it is enough to understand the results. The only exception
is the zoomed diagram for joint 2. Bearing in mind the initial
robot arm position (29), it is clear why initially screw driving is
performed by joint 5 alone. It is the only joint able to provide
the rotational motions about the longitudinal axis of the screwdriver. Time histories of joints virtual fatigues are presented
(in
next (Fig. 3). The figure shows armature temperatures
-th joint, regular order
), as well as time histories of virtual fatigues. When talking about virtual fatigue, one
vershould distinguish between the original and filtered
sion. Original versions , shown in Fig. 3, are subject to high
frequency oscillations (period equals 1 s). The filtered versions
cannot be observed in the figure since they are covered by oscillatory . However, this does not compromise the possibility
to understand the character and the role of virtual fatigue. Initial
level of all curves is identical and equal to the adopted ambient
C. It means that the reference level of
temperature
. Such a choice enables direct comvirtual fatigues is also

Fig. 2. Joint motions: q , i = 1; 3; 2; 4; . . . ; 7. Note that q is shown in two

versions: normal and zoomed view. Zoomed view shows the motion drift in
shoulder: elbow starts to move toward the trunk.

parison of time histories of all three variables (

, , and ).
Noisy shape of values justifies their filtering.
and
(or
)
Relations between time histories of
should be considered next. This will be done by comparing diagrams from Figs. 2 and 3. Let us start from joint 5. Diagram
(Fig. 3) features a slope drop, after reaching the assigned
. The drop is not sharp, but comes some
critical level
. This resembles the situation after aptime after reaching
pearance of the feeling of discomfort in a human forearm performing a similar screw-driving task. Humans certainly would
not try to immediately and completely relax a group of muscles
sensing the discomfort, but rather make some slow attempts to
rearrange involvement of available motor units and partially reduce participation of those muscles. Hence, discomfort caused
by fatigue would keep increasing for some time, but with a lower
reaches the critical value, penalty function
slope. When
starts to work, causing decreased amplitude of motion
(Fig. 2). This means that with appearance of
in diagram
fatigue, engagement of that joint decreases. This is enabled by
proper participation of other joints, which starts exactly at the
moment when joint 5 reaches its critical value. This effect is ap.
parent from all diagrams
Diagrams for motions in joints 2, 3, and 7 deserve special
attention, because of their prevalent participation in generating
the required motions of the screwdriver, compensating the reduced involvement of joint 5. This is also similar to the natural
behavior of human arm, which after sensing the discomfort (fa-

590

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 32, NO. 5, SEPTEMBER 2002

pearance of biological fatigue. Both robot arm and a human put

their elbows closer to the trunk, after subject to fatigue caused
by endurance movements during screw driving. This posture of
a human arm is more comfortable to work. It is important to note
that the robot arm proceeds its normal operation, just taking the
new posture. Redistribution of motion, depression of some joint
motions and stronger engagement of others, does not compromise the end-effector motion task. The complete redistribution
is shown in diagrams
(Fig. 2).
In addition, one may notice that armature temperatures are not
critical during the robot operation (Fig. 3), since each of them is
under the allowed limit taken from the catalogue [42]. That dc
actuators are still in their recommended operating ranges means
that requirements on motor thermal dynamics are not so high in
the suggested method.
The effects of different robot arm postures can be analyzed
using time responses of mechanical and electrical work in joints.
Mechanical work in th joint is calculated using
(30)
where denotes the control torque, is the joint velocity, and
is the total duration of motion. Electrical work is computed
as follows
(31)

Fig. 3. Time histories of armature temperatures

; ; ;
; .
filtered virtual fatigues z , i

= 1 2 3 ... 7

, virtual fatigues z , and

tigue) in the forearm, engages exactly the same joint rotations to

relax exhausted muscles. Increased engagement of robot joints
results in equivalent increase of their fatigue. Two of the three
most actively participating joints (2 and 3) are in the shoulder
of the robot arm. These joints provide the required motions,
along with compensating the gravity load of the complete robot
mechanism. The third active rotation is in the wrist (joint 7) and
coincides with the rotation of a human wrist that is able to participate in endurance movements. It can stand significant dynamical demands and slowly fatigues. Its engagement in providing
speeded-up motions in the writing task has been investigated in
[14]. These facts imply that fatigue effects should appear in the
shoulder first, rather than in the wrist. This is equivalent to the
natural behavior of a human arm. To achieve these effects, we
size the dc servos in accordance with the power capabilities of
the joints in human arms. As a consequence, virtual fatigue
is the next (after ) to reach adopted threshold
. After
reaching the threshold, the penalty function will keep the fatigue
in the vicinity of that level. The motion in joint 2 (shoulder)
drifts toward lower values of . This is shown in Fig. 2 and even
zoomed for better observation. This is apparently equivalent to
an elbow shift toward the trunk, occurring after the shoulder is
subject to fatigue. The suggested method for redundancy resolution based on virtual fatigue, provides reconfiguration of a
robot mechanism in a way identical to a human arm after ap-

and
denote armature voltage and current of the
where
drive. Time responses of mechanical and electrical work are
monotone increasing, with different slopes for different robot
arm postures (Fig. 4).
By inspection of the presented diagrams, one can notice that
time histories of virtual fatigue, as well as electrical work, start
to grow from the very beginning of robot arm motion, even before the majority of joints start their movements. This is a consequence of the robot arm dynamical model (16), consisting of a
system of coupled nonlinear differential equations. The torque
in a joint depends on its own motion, on motions of all other
joints and finally on gravitational load. So, it could be different
from zero even when the joint is not moving. Control torques are
generated by dc actuators, which feature certain power losses
in motor armatures. Power loss increases state coordinates in
motor thermal model (18) and causes rise of the virtual fatigue
in (21) and (22). This is in accordance with a human who may
feel fatigue in a joint that was not engaged in motion.
and
, on one hand and diInspection of diagrams for
(i.e.,
), on the other, reveals the influence
agrams for
of critical values of virtual fatigues on slopes of time histories
of mechanical and electrical work. For example, slopes of
and
are lower until virtual fatigue in the joint 5 reaches its
critical level. Then, joint 2 starts to move, requesting a higher
mechanical and electrical power, which in turn increases slopes
and
. When the virtual fatigue in joint 2 reaches its
of
limit, control torque and voltage are reduced. Torque is reduced
since: 1) motion in joint 2 is depressed thanks to the redundancy
resolution method and 2) center of mass of the mechanism is
shifted down and toward the trunk, decreasing the gravitational

POTKONJAK: et al. MOTION IN HUMAN AND MACHINE: A VIRTUAL FATIGUE APPROACH

591

Fig. 5.

Principal time histories of force F and torque M acting on the screw.

Fig. 6. Motion in joint 2 (force and torque applied to the screw).

Fig. 4. Time histories of mechanical A

and electrical A

work.

Fig. 7. Time histories of virtual fatigue z , filtered virtual fatigue z and

armature temperature
for joint 2 (force and torque applied to the screw).

load. Lowering the center of mass is a result of reconfiguration

of the robot arm, accomplished in a human-like fashion (elbow
drifts toward the trunk). Reduced motion and load decrease the
mechanical and electrical power, which is confirmed by lower
and
.
slopes in the time responses of
We now recall that the results presented above did not include the force and the torque applied to the screw. The next
simulation is performed taking into account these effects. A
N
second-order virtual fatigue model is still used. Force
Nm are applied to solve the reaction of the
and torque
object. The force and the torque are applied while screwing in,
that is, while performing rotation of the screwdriver in the forward direction. During backward rotation there is no need for
are presented in Fig. 5. Zero levels of
force or torque. and
force and torque correspond to free backward turns of the screwincrease the dynamic load that should
driver. Nonzero and
be compensated by the joint actuators. For the simulations accounting for and , another servo sizing was required. A new
selection of dc actuators was accomplished, resulting in the list
of motors given in Table IV presented in Appendix B.
We present simulation results for joint 2 only (Figs. 68).
, fatigue
(note that
is
Given diagrams for motion
,
, show qualitacovered by oscillations of ) and
tively similar behavior of the robot arm compared with the case
were zero (compare with related diagrams from
when and
(Fig. 6), the pheFigs. 24). According to the diagram for

Fig. 8. Time histories of mechanical A

2 (force and torque applied to the screw).

and electrical A

work for joint

nomenon of human-like reconfiguration of robot arm posture,

by moving the elbow closer to the trunk, is appearing again.
The final simulation was accomplished using the first order
model of virtual fatigue (22). Simulations were carried out for
both cases, omitting and considering the effects of and .
Since qualitatively similar results were obtained, there is no
need to present both of them. We present only the results with
and
omitted and give only diagrams for joint 2 (Figs. 911).
The parameters in (22) are indicated by in Appendix A.
, fatigue
(
),
Observing the diagrams for motion
and
, one can notice similar behavior
and work
like in the previously considered cases. Moreover, presented

592

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 32, NO. 5, SEPTEMBER 2002

TABLE I
DH PARAMETERS OF ANTHROPOMORPHIC ROBOT ARM

Fig. 9.

Motion in joint 2 (first-order model of virtual fatigue).

The simulation example of a seven-DOF anthropomorphic

robot arm performing a screw-driving task, shows that the
virtual fatigue method provides human-like motions of robot
joints. The nature of the approach and the task considered
implies a possibility of practical application of the proposed
method.
Fig. 10. Time histories of virtual fatigue z , filtered virtual fatigue z and
armature temperature
for joint 2 (first-order model of virtual fatigue).

Fig. 11. Time histories of mechanical A

2 (first-order model of virtual fatigue).

and electrical A

work for joint

diagrams are behaving more regularly than related results

from previous simulations (Figs. 24 and Figs. 68). This statement can be supported by the following observation related
. According to the diagram, after each reconfiguration
to
of the robot arm, caused by actual progress in virtual fatigue,
joint 2 comes into a different stationary phase. Each phase is
characterized by oscillations with bounded amplitude range.
In previous simulations, the amplitudes were not stable but
increased. It is more natural for movements of a human arm to
perform cyclic motions of stable amplitude during each level
of perceived fatigue. It seems that the reduced model of virtual
fatigue better imitates human-like behavior than the original
second order model. Benefits of the reduced model are also
obvious from the lower involvement of noise in the diagram
of actual virtual fatigue, as well as from apparently noticeable
differences in slope values in the diagrams of mechanical and
electrical work at the moments of robot arm reconfiguration.
It can be concluded that a first order model of virtual fatigue
gives better results consistent with the objectiverealization
of human-like motions in a robot.

V. CONCLUSION
The paper suggests generation of robot motions using a biological analogue. An artificial value called virtual fatigue is
suggested. It is used as an equivalent of biological fatigue in
human muscles. Its purpose is to provide robot motions in accordance with motions in a human arm subject to biological fatigue. Mathematical models of virtual fatigue are derived based
on existing models of dc motors thermal dynamics. Reduction
of the model order is suggested. A method for redundancy resolution using temporal values of joint virtual fatigue is presented. The method works as follows. During motions, the robot
takes comfortable configurations in a psychophysical sense. It
means that joint positions should be as close as possible to the
middle values of joint motion ranges. Time history of virtual
fatigue in each joint depends on its involvement in the overall
motion of the robot, as well as on the static and dynamic load
that should be compensated in that joint. After virtual fatigue
in a joint reaches some pre-assigned critical value, motion in
that joint is depressed, with a proper reengagement of other
joints. It enables an exhausted joint to take some rest, since
the other joints may compensate its reduced involvement. As
a consequence, the robot posture is changed. This method is
consistent with the natural behavior of a human arm. The suggested method is verified by means of computer simulation.
A seven degrees-of-freedom anthropomorphic robot arm performing a screw-driving task is simulated. The mechanical construction of the robot imitates the kinematics of a human arm.
Simulation results show human-like effects in robot arm postures after applying the suggested method. The robot arm performs movements like a human arm in a similar manipulation
task. A human-like phenomenon of bringing the elbow closer to
the trunk after the shoulder is subject to virtual fatigue, appears
and strongly supports validity of the method. Practical effects
of both second and first order models of virtual fatigue are analyzed. It appears that a first-order model provides better results,
recommending it for further implementation in real robots.

POTKONJAK: et al. MOTION IN HUMAN AND MACHINE: A VIRTUAL FATIGUE APPROACH

TABLE II
KINEMATIC AND DYNAMIC PARAMETERS OF ANTHROPOMORPHIC
ROBOT ARM AND PARAMETERS FOR REDUNDANCY RESOLUTION
USING THE VIRTUAL FATIGUE APPROACH

593

TABLE IV
PARAMETERS OF ASSUMED DC DRIVES
(CONSIDERING CONTACT WITH THE OBJECT)

TABLE III
OF ASSUMED DC DRIVES
(DISREGARDING CONTACT WITH THE OBJECT)
PARAMETERS

order virtual fatigue models, disregarding the contact with an

object. Notation is in accordance with that applied in the text
above.
APPENDIX B
Models of assumed dc drives [42], in the case when contact
with an object is taken into account:
14. RE075-201-31GAA200A (250 W);
5. RE036-072-01EAB200A (70 W);
6. RE035-071-34EAG201A (40 W);
7. RE025-055-35EBD202A (20 W).
Corresponding parameters are given in Table IV. Notation is
in accordance with that applied in the body of the paper.
REFERENCES
APPENDIX A
DH parameters of the anthropomorphic robot arm are given
in Table I.
Kinematic and dynamic parameters of the anthropomorphic
robot arm, as well as parameters for redundancy resolution using
the virtual fatigue approach, are presented in Table II. The parameters correspond to second and first (indicated by ) order
virtual fatigue models. Notation is in accordance with that applied in the text of the paper.
Model numbers of assumed dc drives [42]
14. 2260.885-51.216-200 (80 W);
5. 2332.909-12.256-200 (11 W);
6. RE025-055-38EAA200A (10 W);
7. RE026-059-02EAL200A (18 W).
Corresponding parameters are presented in Table III. Given
parameters correspond to second and first (indicated by )

[1] Y. Nakamura, Advanced Robotics: Redundancy and Optimization. Amsterdam, The Netherlands: Addison-Wesley, 1991.
[2] A. Ligeois, Automatic supervisory control of the configuration and
behavior of multibody mechanisms, IEEE Trans. Syst., Man, Cybern.
SMC-, vol. 7, pp. 868871, 1977.
[3] A. A. Maciejewski and C. A. Klein, Obstacle avoidance for kinematically redundant manipulators in dynamically varying environments,
Int. J. Robot. Res., vol. 4, no. 33, pp. 109117, 1985.
[4] T. Yoshikawa, Manipulability of robotic mechanisms, Int. J. Robot.
Res., vol. 4, no. 2, pp. 39, 1985.
[5] K. N. Groom, A. A. Maciejewski, and V. Balakrishnan, Real-time
failure-tolerant control of kinematically redundant manipulators, in
Proc. IEEE Int. Conf. Robot. Automat., Albuquerque, NM, Apr. 1997,
pp. 25952600.
[6] M. Vukobratovic and M. Kircanski, Kinematics and Trajectory Synthesis
of Manipulation Robots. Berlin, Germany: Springer-Verlag, 1986.
[7] J. M. Hollerbach and K. C. Suh, Redundancy resolution of manipulators through torque optimization, IEEE J. Robot. Automat., vol. RA-3,
pp. 308316, 1987.
[8] V. Potkonjak, Distributed positionong for redundant robotic systems,
Robotica, vol. 8, no. 1, pp. 6176, 1990.
[9]
, New approach to the application of redundant robots, Robot.
Comput-Integr. Manuf., vol. 8, no. 3, pp. 181185, 1991.

594

IEEE TRANSACTIONS ON SYSTEMS, MAN, AND CYBERNETICSPART A: SYSTEMS AND HUMANS, VOL. 32, NO. 5, SEPTEMBER 2002

[10] V. Potkonjak and A. Krstulovic, Contribution to the kinematics and dynamics of redundant robots via distributed positioning, J. Intell. Robot.
Syst., vol. 5, pp. 229239, 1992.
Milosavljevic, D. Antic, and D.
[11] V. Potkonjak, G. S. ordevic, C.
Popovic, Variable structure system for control of redundant robots,
Robot. Autom. Syst., vol. 13, no. 1, pp. 1324, 1994.
[12] V. Potkonjak and A. Krstulovic, Mathematical modeling of a redundant anthropomorphic arm (part I), Robot. Autonom. Syst., vol. 9, pp.
165170, 1992.
[13]
, Simulation of a redundant anthropomorphic arm (part II),
Robot. Autonom. Syst., vol. 9, pp. 171179, 1992.
[14] V. Potkonjak, M. Popovic, M. Lazarevic, and J. Sinanovic, Redundancy problem in writing: From human to anthropomorphic robot
arm, IEEE Trans. Syst., Man, Cybern. B, vol. 28, pp. 790805, Dec.
1998.
[15] D. P. Martin, J. Baillieul, and J. M. Hollerbach, Resolution of kinematic redundancy using optimization techniques, IEEE Trans. Robot.
Automat., vol. 5, pp. 529533, Aug. 1989.
[16] J. Baillieul, Kinematic programming alternatives for redundant manipulators, in Proc. IEEE Conf. Robot. Automat., St. Louis, MO., 1985,
pp. 722728.
[17] H. Seraji, Configuration control of redundant manipulators: Theory and
implementation, IEEE Trans. Robot. Automat., vol. 5, pp. 472490,
Aug. 1989.
[18] L. Sciavicco and B. Siciliano, Modeling and Control of Robot Manipulators. London: McGraw-Hill, 1996.
[19] D. R. Baker and C. W. Wampler II, On the inverse kinematics of redundant manipulators, Int. J. Robot. Research, vol. 7, no. 2, 1988.
[20] V. Potkonjak, G. S. ordevic, M. Rasic, and D. Kostic, Representation
of robot motion control skill, IEEE Trans. Syst., Man, Cybern. C, vol.
30, pp. 219238, May 2000.
[21] D. Michel, Learning from humans strategies of motor control: A challenge for robotic systems design, in Proc. IEEE Int. Workshop Robot
Human Commun., 1996.
[22] H. Cruse, E. Wischmeyer, M. Bruwer, P. Brockfield, and A. Dress, On
the cost functions for the control of the human arm movement, Biol.
Cybern., no. 62, pp. 519528, 1990.
[23] O. Khatib, Dynamic control of manipulators in operational space, in
Proc. 6th IFToMM Congr. Theory Machines Mechan., New Delhi, India,
1983, pp. 11231131.
[24] N. Hogan, An organizing principle for a class of voluntary movements, J. Neuroscience, no. 4, pp. 27452754, 1984.
[25] A. H. Sief-Naraghi and J. M. Winters, Changes in musculosceletal control strategies with loading: Inertial, isotonic, random, in Proc. ASME
Biomech. Symp., vol. AMD-98, 1989, pp. 355358.
[26] M. L. Lathash, Control of human movement. Champaign, IL: Human
Kinetics , 1994.
[27] O. Fuentes and R. C. Nelson, Morphing Hands and Virtual Tools (Or
What Good is an Extra Degree of Freedom), Univ. Rochester, Comput.
Sci. Dept., Rochester, NY, Tech. Rep. 551, 1994.
[28] E. S. Jung, J. Choe, and S. H. Kim, Psychophysical cost function of joint
movement for arm reach posture prediction, in Proc. Human Factors
Ergonomics Soc. 38th Annu. Mtg., 1994.
[29] T. F. Chan and R. V. Dubey, A weighted least-norm solution based
scheme for avoiding joint limits for redundant joint manipulators, IEEE
Trans.Robot. Autom., vol. 11, pp. 286292, April 1995.
[30] T. berg, Subjective and objective evaluation of shoulder muscle fatigue, Ergonomics, vol. 37, no. 8, pp. 13231333, 1994.
[31] J. Mizrahi, Fatigue in muscles activated by functional electrical stimulation, Crit. Rev. Physi. Rehab. Med., vol. 9, no. 2, pp. 93129,
1997.
[32] T. Kiryu, M. Morishiata, H. Yamada, and M. Okada, A muscular fatigue
index based on the relationships between superimposed M wave and
preceding background activity, IEEE Trans. Bio-Med. Eng., vol. 45,
pp. 11941204, Oct. 1998.
[33] P. H. P., Electrical Excitation of Skeletal Muscle: Alterations in Force,
Fatigue and Metabolic Properties, Case Western Reserve University,
Cleveland, OH, Report, no. EDC 4-72-32, 1972.
[34] L. Vodovnik and S. Reberzek, Improvements in voluntary control of
paretic muscles due to electrical stimulation, in Electrical Stimulation
as a Rehabilitation Method to Improve Abnormal Locomotion and Manipulation Ljubljana, Slovenia, 1975, pp. 6478.

[35] Y. Giat, J. Mizrahi, and M. Levy, A model of fatigue and recovery

in paraplegics quadriceps muscle subjected to intermittent FES, J.
Biomech. Eng. , vol. 118, no. 3, pp. 357366, 1996.
[36] D. G. Jenkins and B. M. Quigley, Endurance training enhances critical
power, Med. Sci. Sports Exercise, vol. 24, pp. 12831289, Nov. 1992.
[37] B. S. Gottfried and J. Weisman, Introduction to Optimization
Theory. Englewood Cliffs, NJ: Prentice-Hall, 1973.
[38] D. E. Whitney, The mathematics of coordinated control of prosthetic
arms and manipulators, J. Dyn. Syst., Measur. Cont., vol. 94, no. 4,
Dec. 1972.
[39] M. Vukobratovic and V. Potkonjak, Dynamics of Manipulation Robots:
Theory and Application. Berlin, Germany: Springer-Verlag, 1982.
[40]
, Applied Dynamics and CAD of Manipulation Robots. Berlin,
Germany: Springer-Verlag, 1985.
[41] V. Potkonjak, Thermal analysis and dynamic capabilities of DC motors
in industrial robotic systems, Robot. Comput.-Integr. Manuf., vol. 5, no.
2/3, pp. 137143, 1989.
[42] Maxon Motor Catalogue, 97/98 ed..
[43] P. Eikof, Fundamentals of Control Systems Identification (in Russian). Moscow, USSR: Mir, 1975.
[44] B. Lara, K. Althoefer, and L. D. Seneviratne, Automated robot-based
screw insertion system, in Proc. IEEE IECON, Aachen, Germany, Sept.
1998, pp. 24402445.

Veljko Potkonjak was born in Belgrade, Yugoslavia

in 1951. He studied at the Faculty of Electrical
Engineering, University of Belgrade, Yugoslavia and
graduated in 1974. The same year he was accepted
in postgraduate studies at the same faculty to finally
finish in 1977. In 1981, he defended his doctoral
thesis. He received the M.Sc. and Ph.D. degees in
robotics.
After graduation, he started his work at the Faculty
of Electrical Engineering, University of Belgrade, as
a Teaching Assistant in the field of mechanics. He
became an Assistant Professor in 1985, and in 1990, he was promoted to the
rank of associate professor and, finally in 1995, to the rank of full professor.
During his educational career, he has been teaching mechanics, robotics and
biomechanics. He was also a teacher or a visiting researcher at the Faculty of
Electronics, University of Nis, the Technical Faculty in Cacak, National Technical University of Athens, and American University of Athens. His research interests primarily concern robotics. His attention has been oriented to problems
concerning the modeling of robotic systems and the implementation of these
models to design and control. He is the a author/coauthor of two international
research monographs (in English), two chapters in the Handbook of Mechanical Systems, several textbooks for universities and secondary schools, about
46 international journal papers, 25 papers on international conferences and several papers in Yugoslav journals and domestic conferences. He has made a large
number of reviews for respected international journals and conferences, as well
as for domestic conferences. Within the field of robotics and automation he was
engaged in several projects (research and commercial), sometimes heading and
sometimes participating.

Dragan Kostic (S00) was born in 1969, Pristina,

Serbia. He received the B.Sc. and M.Sc. degrees
in automatic control and robotics from the Faculty
of Electronic Engineering, University of Nis, Nis,
Serbia, in 1994 and 2000, respectively. Since
2000 he has been pursuing the Ph.D. degree with
the Dynamics and Control Technology Group
at the Department of Mechanical Engineering,
Eindhoven University of Technology, Eindhoven,
The Netherlands.
Previously, he investigated problems in biomechanics and ergonomics. His current research interests include dynamic
modeling, identification, and robust and data-based control of nonlinear
mechanical systems, in particular, robots.

POTKONJAK: et al. MOTION IN HUMAN AND MACHINE: A VIRTUAL FATIGUE APPROACH

Milan Rasic (S00) was born in 1972 in Yugoslavia.

He received the B.Sc. and M.Sc. degrees in electrical
engineering from the Faculty of Electronic Engineering, University of Nis, Nis, Serbia in 1996 and
2000, respectively.
He is a Teaching Assistant at the Department
of Automatic Control and a Researcher with the
Robotics and Automation Lab at the Faculty of
the Electronic Engineering, University of Nis.
His research interests include robotics, intelligent
control, embedded control systems, and application
of PCs in control and data acquisition.

595

Goran S. ordevic (M95) received the B.Sc.,

M.Sc., and Ph.D. degrees in electrical engineering,
from the University of Nis, Nis, Yugoslavia, in 1988,
1992, and 1995, respectively.
From 1988 to 1995, he was a Research and
Teaching Assistant at the Department of Automatic
Control, Faculty of Electronic Engineering, University of Nis. Since 1996 he has been an Assistant
Professor of automatic control, robotics, robot
control and flexible automation at the University
of Nis. He is head of the Robotics and Automation
Lab. He also works with the Department of Biomedical Engineering, The
Johns Hopkins University, Baltimore, MD, developing intelligent controller
of biomimetic robots. His research interests include intelligent, biologically
inspired solutions to robot control, robotics in medicine, data modeling, and
PC-based control.