Está en la página 1de 10

Jared Gragg

James Yang
1
e-mail: james.yang@ttu.edu
Human-Centric Design Research Laboratory,
Department of Mechanical Engineering,
Texas Tech University,
Lubbock, TX 79409
Digital Human Forward
Kinematic and Dynamic
Reliabilities
Probabilistic methods have been applied to many problems in various elds of study.
There are many distinct applications of probabilistic design in the biomechanics eld, in
particular. Traditionally, deterministic methods have been applied in digital human mod-
eling (DHM). Transforming the deterministic approach of digital human modeling into a
probabilistic approach is natural since there is inherent uncertainty and variability asso-
ciated with DHM problems. Typically, deterministic studies in this eld ignore this uncer-
tainty or try to limit the uncertainty by employing optimization procedures. Often, inverse
kinematics or dynamics techniques are introduced to point the system to the desired solu-
tion, or best solution. Due to the variability in the inputs, a deterministic study may not
be enough to account for the uncertainty in the system. Probabilistic design techniques
allow the designer to predict the likelihood of an outcome while also accounting for
uncertainty, in contrast to deterministic studies. The purpose of this study is to incorpo-
rate probabilistic approaches to a deterministic DHM problem that has already been
studied, analyzing human forward kinematics and dynamics. The problem is transformed
into a probabilistic approach where the human forward kinematic and dynamic reliabil-
ities are determined. The forward kinematic reliability refers to the probability that the
human end-effector position (and/or orientation) falls within a specied distance from
the desired position (and/or orientation) in an inverse kinematics problem. The forward
dynamic reliability refers to the probability that the human end-effector position (and/or
velocity) falls within a specied distance from the desired position (and/or velocity) along
a specied trajectory in the workspace. The dynamic equations of motion are derived by
the Lagrangian backward recursive dynamics formulation. [DOI: 10.1115/1.4024234]
Keywords: probabilistic design, forward kinematics, human dynamics, digital human
modeling
1 Introduction
In many engineering applications, it is necessary to include
uncertainty and variability into the design process. Typically,
employing a deterministic study for a DHM problem or design is
not enough to adequately account for the inherent uncertainty and
variability. Uncertainty and variability often enters DHM prob-
lems through human anthropometry or interaction with the envi-
ronment. Certain human anthropometric parameters, such as size,
shape, material properties, and inertial properties, are variable by
nature. Human beings are constantly interacting with their envi-
ronment. Parameters involved in the interaction between humans
and the environment also have uncertainty. Other factors in DHM
simulations, such as redundancy and human performance meas-
ures, will introduce further uncertainty and variability. In previous
works [13], kinematics and physics-based posture prediction
simulations were deterministic problems where uncertainty in the
inputs was not considered.
This study presents a novel probabilistic approach to DHM that
has not been addressed in literature. A deterministic study gives
outputs that are exactly determined from a set of inputs, but if
there is variability in the inputs, then a deterministic study may
not be enough to account for the uncertainty in the system. In con-
trast to deterministic studies, a probabilistic approach allows one
to predict the likelihood of the outputs while also accounting for
uncertainty.
Probabilistic methods have been applied to many elds, but, in
particular, the biomechanics eld has many applications where
the approach is well suited. Strickland et al. [4] employed the
Monte Carlo simulation technique to vary factors, such as malpo-
sitioning and soft-tissue constraint, in a study of two knee
implants under ISO-gait and passive laxity loading. Using a prob-
abilistic analysis allowed the designers to introduce uncertainty
into the design through the inherent variability of the input factors,
such as component positioning and material properties. Hughes
and Nelson [5] investigated a mathematical model for estimating
the investment worthiness of ergonomic intervention that would
prevent occupational lower back pain. The net present value dis-
tributions were modeled using the Monte Carlo simulation. Guo
and de Vita [6] presented a new probabilistic constitutive equation
that describes the damage evolution process in ligaments. They
employed the Weibull distribution to generate random data in
the determination of a constitutive equation describing the accu-
mulated damage in parallel-bered collagenous tissues, such as
ligaments. Galbusera et al. [7] studied the effect of axial, antero-
posterior, and lateral positioning of the center of rotation of a ball-
and-socket disc prosthesis for the C5 and C6 spine units. The
study is signicant in the fact that the input parameters involved
were not predictable in a deterministic way, yet through probabil-
istic techniques the effects that the geometric variables had on the
spine biomechanics were estimated. Rohlmann et al. [8] per-
formed a sensitivity analysis for the pedicle-screw-based motion
preservation system that supports slightly degenerative discs in
the lumbar spine. They employed the Latin hypercube sampling
method using the program OPTISLANG to perform a sensitivity anal-
ysis for a motion preservation system for degenerative discs in the
lumbar spine.
1
Corresponding author.
Contributed by the Design Automation Committee of ASME for publication in
the JOURNAL OF MECHANICAL DESIGN. Manuscript received October 29, 2012; nal
manuscript received April 4, 2013; published online May 24, 2013. Assoc. Editor:
Matthew B Parkinson.
Journal of Mechanical Design JULY 2013, Vol. 135 / 071008-1 Copyright VC
2013 by ASME
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 11/27/2013 Terms of Use: http://asme.org/terms
Probabilistic design methods have also been applied in the
robotics eld. Hu et al. [9] employed probabilistic methods for
optimizing bipedal gait in a nine-link humanoid robot. Bipedal
locomotion was formulated as a multiconstraint optimization
problem with two objective functions: zero-moment point, a com-
mon stability criterion for positioning of a robot or digital human;
and torque, a common human performance measure in DHM or
performance characteristic in robotics. Another example from
robotics that employs probabilistic methods for position tracking
is from a series of studies conducted by Koshizen [1012]. In Ref.
[10], the author presented an enhanced mobile robot localization
technique based on sensor selection. A key point in the problem
was how to model the uncertainty in the problem and then reduce
the uncertainty. Koshizen [10] presented a Gaussian mixture of
Bayes with regularized expectation maximization (GMB-REM)
technique that uses a sonar sensor. Bayes theorem was employed
in order to reduce the uncertainty of the sonar sensor. Koshizen
[11] enhanced the GMB-REM technique by incorporating a sen-
sor selection technique combined with sensor fusion scheme.
Koshizen [12] further enhanced the technique developed in Ref.
[10]. The main achievement of the nal work was constraining
the error no matter how much noise was introduced to the sensory
models. In Noohi et al. [13], probabilistic methods were employed
to ensure collision-free robot motion planning and stability of the
robots ngertip grasp. Uncertainty arose in the problem due to
the uncertainty of the position of the contact points in the robot
ngertip gripper.
Although there have been applications of probabilistic methods
in the biomechanics eld and the robotics eld, the authors are
unaware of any such studies in the DHM eld. Rao and Bhatti
[14] presented a probabilistic approach to robot manipulator kine-
matics and dynamics. In the study, forward kinematic and
dynamic reliabilities were found for two robot manipulators: a 2
degree of freedom (DOF) two-link manipulator and a 6 DOF Stan-
ford arm manipulator. This paper extends the approach to a 21
DOF digital human model of the spine and right arm. The focus
of this paper is on forward kinematics and dynamics. For forward
kinematics, joint angles are given and the goal is to determine the
end-effector position. For forward dynamics, joint torque proles
are given and the goal is to determine joint angle proles and end-
effector locations. First, the digital human kinematics and dynam-
ics are given. Second, the probabilistic approach is explained and
the forward kinematic and dynamic reliabilities dened. Finally,
the forward kinematic and dynamic reliabilities are calculated for
a human pulling task. A pilot study [15] was presented at the 2012
ASME IDETC conference. In the previous work, the dynamic
reliability results were limited to one single time step. The current
study expands the dynamic results and includes an additional sen-
sitivity analysis.
2 Forward Kinematic and Dynamic Reliabilities
2.1 Human Kinematics and Dynamics. In DHM, the body
is often modeled as a kinematic joint chain that represents a
humans skeletal system. The skeletal structure is constructed
according to the DenavitHartenberg (DH) convention [16]. All
joints in the chain are single DOF revolute joints described by a
single real number, or joint angle. For the ith joint, we associate
the joint variable q
i
and rigidly attach a coordinate system. Each
subsequent joint is connected by a link of varying length. For
joints with multiple DOFs, such as the shoulder, the human joint
is represented as multiple DH joints connected by links of length
zero. q q
1
q
n

T
2 R
n
is the set of all joint angles and
represents a specic posture. xq 2 R
4
is the augmented position
vector of the end-effector in Cartesian space with respect to the
global coordinate system. x(q) is determined through the DH
method and expressed as the product of the individual transforma-
tions associated with each joint. The 4 4 homogeneous transfor-
mation matrix T
i
denotes the position and orientation of joint i
with respect to joint i 1. x(q) and T
i
are found by the DH
method as follows:
xq

n
i1
T
i
_ _
x
n
(1)
T
i

cos h
i
cos a
i
sin h
i
sin a
i
sin h
i
a
i
cos h
i
sin h
i
cos a
i
cos h
i
sin a
i
cos h
i
a
i
sin h
i
0 sin a
i
cos a
i
d
i
0 0 0 1
_

_
_

_
(2)
where x
n
is the position of the end-effector with respect to the nth
frame and n is the number of DOFs. The rotational displacement
q
i
changes the value of h
i
. The constants a
i
, a
i
, and d
i
are the DH
parameters link twist, link length, and link offset and along with
h
i
are used to dene the geometric relationships between the joint
i 1 and joint i. For this paper, a 21 DOF digital human model of
the spine and right arm is presented and detailed in Fig. 1.
Due to the tree structure nature of digital human models, the
Lagrangian backward recursive dynamics formulation was used to
derive human dynamics. The forward recursive kinematics for
j 1; ; n can be written as follows:
Fig. 1 Twenty one DOF human model
071008-2 / Vol. 135, JULY 2013 Transactions of the ASME
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 11/27/2013 Terms of Use: http://asme.org/terms
A
j

n
j1
T
j
A
j1
T
j
with A
0
U
B
j

_
A
j
B
j1
T
j
A
j1
@T
j
@q
j
_ q
j
with B
0
0
C
j

_
B
j
C
j1
T
j
2B
j1
@T
j
@q
j
_ q
j
A
j1
@
2
T
j
@q
2
j
_ q
2
j
A
j1
@T
j
@q
j
q
j
with C
0
0
r
0
j
A
j
r
j
_ r
0
j
B
j
r
j
r
0
j
C
j
r
j
(3)
where r
0
j
, _ r
0
j
, and r
0
j
are the global augmented position, velocity,
and acceleration vectors of a point in Cartesian space, respec-
tively. r
j
is the augmented position vector of the point in the jth
local coordinate frame and U is an identity matrix. The backward
recursive dynamics for j n; ; 1 can be written as follows:
D
j
I
j
C
T
j
T
j1
D
j1
with D
n1
0
E
j
m
j
r
j
j
T
j1
E
j1
with E
n1
0
F
j
r
k
f
d
jk
T
j1
F
j1
with F
n1
0
G
j
h
k
d
jk
G
j1
with G
n1
0
(4)
I
j

I
xx
I
yy
I
zz
=2 I
xy
I
xz
m
j
~ x
j
I
xy
I
xx
I
yy
I
zz
=2 I
yz
m
j
~ y
j
I
xz
I
yz
I
xx
I
yy
I
zz
=2 m
j
~z
j
m
j
~ x
j
m
j
~ y
j
m
j
~z
j
m
j
_
_
_
_
_
_
_
_
(5)
where I
j
is the inertia matrix for the jth link dened according to Eq.
(5); m
j
is the mass of the jth link; ~ x
j
; ~ y
j
; ~z
j
are the x, y, z coordinates
of the center of mass of the jth link in the local coordinate system;
r
j
j
~ x
j
~ y
j
~z
j
1 is the location of the center of mass of the jth
link in the local coordinate system; r
k
f
is the position of the external
force in the local kth frame; d
jk
is the Kronecker delta; and h
k
is the
external moment vector for the kth link in global coordinates. The
equations of motion for j n; ; 1 can then be written as follows:
s
j
tr
@A
j
@q
j
D
j
_ _
g
T
@A
j
@q
j
E
j
f
T
k
@A
j
@q
j
F
j
G
T
j
A
j1
z
0
(6)
where g is the gravity vector in global coordinates, f
k
is the exter-
nal force vector for the kth link in global coordinates, and
z
0
0 0 1 0
T
for a revolute joint.
2.2 Monte Carlo Method. There are many simulation tech-
niques employed for probabilistic studies. Some of the techniques
require a background in probability and statistics while others do
not. The basic simulation procedure is as follows: each random
variable in the problem is sampled several times in order to gain
understanding of its probabilistic characteristics, a set of numbers
are generated that indicates one realization of the problem itself,
the problem is solved deterministically for each realization (one
simulation cycle), and the probabilistic characteristics of the prob-
lem are determined by using many simulation cycles [17]. The
most common simulation technique employed in probabilistic
studies is the Monte Carlo simulation technique. The Monte Carlo
simulation uses repeated random sampling of variables according
to their distributions and allows the designer to study risk or reli-
ability of complicated engineering systems with limited back-
ground knowledge of probability or statistics [17,18]. The Monte
Carlo simulation involves six steps: (1) dening the problem in
terms of all of the random variables; (2) determining probability
density functions (PDFs) or probability mass functions for all of
the continuous or discrete random variables, respectively; (3) gen-
erating values for the random variables; (4) performing simulation
cycles, that is, solving the problem deterministically for each set
of random variables; (5) using the generated data to extract proba-
bilistic information; and (6) determining the associated accuracy
and efciency for the M simulation cycles [17]. The popularity of
the Monte Carlo simulation is tied to the greatest strength of the
Monte Carlo simulation, its ability to converge to the correct solu-
tion. Though the Monte Carlo simulation will always converge on
the correct solution, the simulation technique may take thousands
of trials to obtain useful results [18]. The accuracy of the Monte
Carlo method is dependent on the number of trials or simulation
cycles, and that is the weakness of the Monte Carlo method, its ef-
ciency. If the probabilities involved are small, the number of
simulation cycles required to converge on the solution will
increase. Since the probability of failure of a part is not known
beforehand, it is often required to take a trial-and-error approach
with the Monte Carlo simulation to ensure accuracy. The Monte
Carlo simulation is a popular simulation technique that will con-
verge on the correct solution, but may require extensive computa-
tion time due to the fact the accuracy is tied to the number of trials.
Other simulation techniques include variance reduction techni-
ques (VRTs), response surface methods (RSM), and most proba-
ble point (MPP) methods. VRTs aim to increase the efciency of
the Monte Carlo technique by reducing the variance of the esti-
mated output variable without disturbing the expected value or
increasing the sample size [17]. VRTs aim to reduce the number
of trials necessary to converge on an accurate solution, but tend to
require increased computational difculty not present in the sim-
ple Monte Carlo simulations. RSMs use a simple analytical func-
tion of the input variables, typically a low-order polynomial
equation, to approximate the output parameter for the entire sam-
ple space [18]. RSMs tend to work well for linear models, whereas
highly nonlinear functions may not represent the true function
well or require increased simulation cycles. MPP methods aim to
reduce analysis time when the deterministic model used in the
simulation cycle has a long run time. MPP methods typically try
to determine the most probable point by using optimization of a
rst-order Taylor series approximation of the performance func-
tion [18]. MPP methods have been shown to converge quickly on
an accurate solution with a small number of simulation cycles but
the disadvantage of the methods is that they only provide informa-
tion for a single point, thus the method must be reapplied several
times for each point of interest. The Monte Carlo simulation is
reliable and requires little background knowledge of probability
and statistics, but computational time can be a concern depending
on the accuracy needed and the probabilities involved. The other
methods are valid methods that can reduce the simulation time
and achieve accurate results, but typically require further knowl-
edge of probability and can often be computationally difcult.
Due to its ease of application, the Monte Carlo technique was
employed in this paper in calculating the forward kinematic and
dynamic reliabilities for a digital human model simulation.
2.3 Forward Kinematic and Dynamic Reliability. Forward
kinematic and dynamic reliabilities for a robotic manipulator are
dened in Ref. [14] and extended to digital human models as fol-
lows. The forward kinematic reliability refers to the probability
that the human end-effector position (and/or orientation) falls
within a specied range from the desired position (and/or orienta-
tion) [14]. The specied range can be thought of as the maximum
allowable position error that the end-effector can have and still be
considered reliable. The specied range is called the permissible
region and its size and shape can vary depending on the applica-
tion and the required accuracy. Based on the application, the two
types of forward kinematic reliability can be dened as the posi-
tional reliability and the orientation reliability. For the purposes of
this study, only the positional reliability was studied.
The forward dynamic reliability refers to the probability that
the human end-effector position (and/or velocity) falls within a
Journal of Mechanical Design JULY 2013, Vol. 135 / 071008-3
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 11/27/2013 Terms of Use: http://asme.org/terms
specied range from the desired position (and/or velocity) along a
specied trajectory in the workspace. According to this denition,
two major types of forward dynamic reliability are dened as fol-
lows: single pass forward dynamic reliability (SPFDR) and cumu-
lative forward dynamic reliability (CFDR). The SPFDR is dened
as the probability that the maximum deviation in the end-effector
location (and/or velocity) from the desired location (and/or veloc-
ity) will not exceed a prescribed value at any point along a given
trajectory [14]. That is, if the SPFDR is equal to 1 (reliability
equal to 1 means certain success) for a given path, at no point in the
path will the end-effector position (or velocity) error exceed the
maximum allowable error. The CFDR is dened as the probability
that the end-effector location (and/or velocity) lies within the speci-
ed range from the desired location (and/or velocity) at every point
along the given trajectory [14]. That is, the CFDR indicates the
probability that the end-effector position (or velocity) error is less
than the maximum allowable error at any point along the trajectory.
CFDR and SPFDR can be expressed mathematically as
CFDR
number of acceptable trajectory points
total number of trajectory points examined
(7)
SPFDR
number of acceptable trajectories
total number of trajectories examined
(8)
according to Rao and Bhatti [14]. In addition to CFDR and SPFDR,
dynamic reliability can be further categorized according to the
application. For the purpose of this paper, only CFDR was studied.
2.4 Human Forward Kinematic Reliability. The forward
kinematic reliability was then calculated using the Monte Carlo
method. The following steps were used to calculate the forward
kinematic reliability, based on Refs. [14] and [17]:
(a) Dene the problem in terms of the systems random var-
iables, RV
k
. The random variables in this case were the
set of joint angles, q, and link lengths (or offsets,
depending on whether the value refers to the DH param-
eter a or d), L L
1
L
10

T
. Thus, RV
k
q; L f g.
(b) Determine PDFs for the continuous random variables.
All the random variables in the simulations were
assumed to follow normal (Gaussian) distributions. The
PDFs for the random variables can therefore be dened
in terms of their mean values, fl
q
; l
L
g, and standard
deviations, r
q
; r
L
_ _
.
(c) Generate values for the random variables. Random val-
ues, q
v
; L
v
f g, for the random variables were assigned
through the MATLAB random number generator according
to their distribution.
(d) Solve the problem deterministically for each set of ran-
dom variables. Each deterministic solution constitutes a
simulation cycle. The position of the end-effector was
calculated in MATLAB as follows:
X
v
xq
v
(9)
(e) Examine the performance measure, i.e., check to see
whether the end-effector distance error, e
v

_
_
xl
q
X
v
k,
is less than the maximum allowable error, e
max
. If the error
is less than the maximum allowable error the simulation
cycle is counted as a success, u
v
1; e
v
e
max
, other-
wise, it is noted as a failure, u
v
0; e
v
> e
max
.
(f) Conduct a large number, k
k
, of these experiments by
repeating steps (c) through (e) and nd the forward kine-
matic reliability, R
k
.
The forward kinematic reliability was calculated as follows:
R
k

kk
v1
u
v
k
k
(10)
Section 3 details the values used in the simulations and Sec. 4
details the results of the experiments.
2.5 Human Forward Dynamic Reliability. In order to cal-
culate the reliability, it was assumed that the torque was a stochas-
tic input to the system and caused the system to accelerate.
Generally for human inverse dynamics, the joint angles, veloc-
ities, and accelerations are predicted and then the torque is calcu-
lated to achieve the desired position, velocity, and acceleration
proles. In order to treat the torque as an input to the system, it
was necessary to integrate the equations of motion to solve for the
acceleration. The matrix C
j
can be rewritten as
C
j
w
j

@A
j
@q
j
q
j
w
j
C
j1
T
j
2B
j1
@T
j
@q
j
_ q
j
A
j1
@
2
T
j
@q
2
j
_ q
2
j
(11)
Solving the equations of motion, Eq. (6), for q
j
yields
q
j

s
j
g
T
@A
j
@q
j
E
j
f
T
k
@A
j
@q
j
F
j
G
T
j
A
j1
z
0
tr
@A
j
@q
j
I
j
w
T
j
T
j1
D
j1

_ _
tr
@A
j
@q
j
I
j
@A
j
@q
j
T
_ _ (12)
This equation is straightforward to solve, but due to the nature of
digital human simulations, the denominator term in Eq. (12) leads
to a solution of lim
Ij !0
q
j
1. For human joints with multiple
DOFs, such as the shoulder, there are DH joints where I
j
0, which
leads to an unrealistic solution. Thus, Eq. (12) must be solved as
tr
@A
j
@q
j
I
j
@A
j
@q
j
T
_ _
q
j
s
j
g
T
@A
j
@q
j
E
j
f
T
k
@A
j
@q
j
F
j
G
T
j
A
j1
z
0
tr
@A
j
@q
j
I
j
w
T
j
T
j1
D
j1

_ _
where q qt
T
; s st
T
; _ q _ qt
T
Dt; q qt
T
Dt
t t
0
t
F
for T 1; ; F
(13)
Equation (13) can be solved with MATLAB, but an explicit form of
the solution is difcult to obtain for situations where I
j
0. The
numerical integration of the equations of motion was carried out
as follows:
_ qt
T
Dt _ qt
T
qt
T
DtDt
qt
T
Dt qt
T
_ qt
T
Dt 1=2 qt
T
DtDt
2
(14)
where the initial conditions q(0) and _ q0 are specied [14].
The forward dynamic reliability was then calculated using the
Monte Carlo method. The following steps were used to calculate
the forward dynamic reliability, based on Refs. [14] and [17]:
071008-4 / Vol. 135, JULY 2013 Transactions of the ASME
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 11/27/2013 Terms of Use: http://asme.org/terms
(a) Dene the problem in terms of the systems random var-
iables, RV
d
. The random variables in this case were: the
joint torques, s; the link lengths (or offsets, depending
on whether the value refers to the DH parameter a or d),
L; the link masses, m; and link inertias, I
j
. Thus,
RV
d
fs; L; m; I
j
g.
(b) Determine PDFs for the continuous random variables.
All the random variables in the simulations were
assumed to follow normal (Gaussian) distributions. The
PDFs for the random variables can therefore be dened
in terms of their mean values, fl
s
; l
L
; l
m
; l
Ij
g, and
standard deviations, r
s
; r
L
; r
m
; r
Ij
_ _
.
(c) Generate values for the random variables. Random val-
ues, s
v
; L
v
; m
v
; I
jv
_ _
, for the random variables were
assigned through the MATLAB random number generator
according to their distribution.
(d) Solve the problem deterministically for q
v
at the current
time step, T, according to Eq. (13). Each deterministic
solution constitutes a simulation cycle. The simulation
cycling was performed in MATLAB.
(e) Integrate the dynamic equations of motion according to
Eq. (14) to nd q
v
.
(f) Use the kinematic model, Eq. (1), to determine the end-
effector position as
X
v
xq
v
(15)
(g) Examine the performance measure, i.e., check to see
whether the end-effector distance error,
e
v

_
_
xl
q
X
v
k, is less than the maximum allowable
error, e
max
. If the error is less than the maximum allow-
able error the simulation cycle is counted as a success,
u
v
1; e
v
e
max
, otherwise, it is noted as a failure,
u
v
0; e
v
> e
max
.
(h) Conduct a large number, k
d
, of these experiments by
repeating steps (c) through (g) and nd the forward
dynamic reliability for the Tth time step.
(i) Continue to the next time step and repeat.
The forward dynamic reliability for the Tth time step and the
CFDR were calculated as follows:
R
d;T

kd
v1
u
v
k
d
(16)
CFDR

c
T1
R
d;T
c
(17)
where c is the number of time steps. Section 3 details the values
used in the simulations and Sec. 4 details the results of the
experiments.
3 Forward Kinematic and Dynamic Monte Carlo
Simulations
The forward kinematic and dynamic parameters for the 21 DOF
digital human model, Fig. 1, were assumed to be independent ran-
dom variables with normal distributions. The mean values of the
kinematic parameters, RV
k
, were taken from a simulation of a
50th percentile (stature) pregnant female during a 3 s pulling sim-
ulation as shown in Fig. 2 [19]. Joint angles from the rst 1 s of
the simulation were taken at 0.01 s intervals, thus 100 sets of joint
angles were used in the forward kinematic reliability calculations.
The link lengths had the same mean values throughout the simula-
tions. The mean values and standard deviations used in the for-
ward kinematic reliability calculation are shown in Table 1. Only
the rst set (rst time step) of values of the joint angles are given.
k
k
10,000 Monte Carlo cycles were performed for each of the
100 sets of joint angles. The maximum allowable position error
for the end-effector was e
max
0.4 mm.
The mean values of the dynamic parameters, RV
d
, were taken
from a simulation of a 50th percentile (stature) pregnant female
during a 3 s pulling simulation [19]. Joint torques from the rst 1 s
of the simulation were taken at 0.01 s intervals, thus 100 sets of
joint torques were used in the dynamic reliability calculations.
The link lengths, link masses, and link inertias had the same mean
values throughout the simulations, respectively. c 100 time steps
and k
d
1000 Monte Carlo cycles were performed for the calcula-
tion of the dynamic reliability. For 100 time steps and 1000 Monte
Carlo simulations per time step, the calculation of forward
dynamic reliability took roughly 15 days on a Dell Precision
T7500 with 2 Intel Xeon X5660 processors at 2.80 GHz. Two sep-
arate simulation trials were studied. The coefcient of variation
(COV) is dened as the ratio of the standard deviation to the
mean, COV r=l, for a particular random variable. The effect of
Fig. 2 Pregnant female pulling simulation
Table 1 Mean values, l, and standard deviations, r, of the sto-
chastic parameters for calculation of forward kinematic
reliability
l r l r
L
i
(m) i q
i
(rad)
0.0562670 0.00005627 1 3.000000 0.0035
0.0544274 0.00005443 2 2.995049 0.0035
0.0599962 0.00006000 3 3.000000 0.0035
0.1738560 0.00017386 4 1.000000 0.0035
0.0170799 0.00001708 5 2.000000 0.0035
0.0557820 0.00005578 6 1.000000 0.0035
0.1505700 0.00015057 7 0.017390 0.0035
0.2586380 0.00025864 8 0.000287 0.0035
0.2473740 0.00024737 9 1.000000 0.0035
0.1650990 0.00016510 10 1.000000 0.0035
11 0.000287 0.0035
12 1.000000 0.0035
13 1.743110 0.0035
14 16.409511 0.0035
15 89.923698 0.0035
16 8.423747 0.0035
17 128.70124 0.0035
18 92.641144 0.0035
19 20.291050 0.0035
20 7.318476 0.0035
21 11.906292 0.0035
Journal of Mechanical Design JULY 2013, Vol. 135 / 071008-5
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 11/27/2013 Terms of Use: http://asme.org/terms
the coefcient of variation on the forward dynamic reliability was
studied. The rst simulation trial considered small coefcients of
variation for the dynamic parameters, RV
d
. For the second trial,
large coefcients of variation were considered. The large coef-
cients of variation were found as double the small coefcients of
variation. The coefcients of variation for the dynamic parameters
are given in Table 2. The mean values and standard deviations
used in the forward dynamic reliability calculation for trial 1 are
shown in Appendix (Tables 4 and 5). Only the rst set (rst time
step) of values of the joint torques are given. An external force of
1 N in the global y direction was applied at the wrist throughout
the pulling simulation. No external moments were applied. The
maximum allowable position error for the end-effector was
e
max
0.4 mm.
4 Results
The results for the calculation of forward kinematic reliability
are summarized in Fig. 3. The gure gives the forward kinematic
reliability, according to Eq. (10), for each of the 100 sets of joint
angles. The joint PDF of the position error for the rst set of joint
angles (rst time step) is given in Fig. 4.
From Fig. 3, it is shown that the forward kinematic reliabilities
for the 100 sets of joint angles range between 0.9529 and 0.9641.
From Fig. 4, it is shown that the PDF of the position error (rst
set) was a normal distribution with mean of 0.1785 mm and a
standard deviation of 0.1076 mm.
The results for the forward dynamic reliability simulations are
shown in Figs. 5 and 6. Figure 5 gives the forward dynamic reli-
abilities, according to Eq. (16), calculated for 100 time steps for
both trials. The joint PDFs (both trials) of the position error for
the rst time step are given in Fig. 6.
From Fig. 5, it is shown that the forward dynamic reliabilities
for 100 time steps ranged between 0.931 and 0.964 for trial 1 and
0.529 and 0.703 for trial 2. From Fig. 6, it is shown that the PDF
of the position error (rst set) was a normal distribution with
mean of 0.1623 mm and a standard deviation of 0.1091 mm for
trial 1 and a normal distribution with mean of 0.3245 mm and a
standard deviation of 0.2183 mm for trial 2. The forward dynamic
reliability for the rst time step of the pulling simulation was cal-
culated as R
d,1
0.965 for trial 1 and R
d,1
0.693 for trial 2. The
CFDR for trials 1 and 2 was 0.9588 and 0.6708, respectively.
5 Discussion
The forward kinematic reliabilities over the 100 time steps
remained fairly consistent. Also, there are no distinct patterns in
the forward kinematic reliabilities across the 100 time steps. The
number of simulations, k
k
for each time step, was relatively high
and therefore it can be assumed that the calculated reliabilities
were accurate. Unlike the forward dynamic reliabilities, there was
little difference in the sensitivity to the random variables,
RV
k
{q, L}, across time.
The forward dynamic reliabilities over the 100 time steps, how-
ever, exhibited certain trends for both trials. The reliabilities
remained relatively constant for the rst 33 time steps and then
experienced a dramatic decrease in the reliability for the 34th time
step, R
d,34
0.529 (trial 2). The reliabilities for the 35th time step
through the 61st time step exhibited similar behavior as the rst
33 time steps. Then there was a steady decline in reliability until
the 76th time step, R
d,76
0.539 (trial 2), followed by a steady
incline in reliability until the 85th time step. After the 85th time
step, the trend from the rst 33 time steps continued. It was specu-
lated that the lower number of simulations, k
d
, could have led to
numerical error for certain time steps. A separate simulation for
the 34th time step was conducted with k 5000 simulation cycles.
The forward dynamic reliability for the 34th time step was then
calculated as 0.517. Therefore, it was concluded that the lower
number of simulations was not responsible for the drastic decrease
Table 2 Coefcients of variation COV r=l for dynamic
parameters
RV
d
COV (trial 1) COV (trial 2)
s 0.005 0.01
L 0.0025 0.005
m 0.0025 0.005
I
xx
0.002 0.004
I
yy
0.002 0.004
I
zz
0.002 0.004
I
xy
0.002 0.004
I
yz
0.002 0.004
I
xz
0.002 0.004
Fig. 3 Forward kinematic reliability results for 100 sets of joint angles
071008-6 / Vol. 135, JULY 2013 Transactions of the ASME
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 11/27/2013 Terms of Use: http://asme.org/terms
in reliability for certain time steps. Since numerical error was
ruled out as a possible source of the decrease in reliability for cer-
tain time steps, other factors were considered. Given that the cal-
culation of forward dynamic reliability was carried out in the
same way for each of the time steps, the only cause for the
decrease in reliability for certain time steps could be higher sensi-
tivity to the input random variables, RV
d
fs; L; m; I
j
g.
A sensitivity analysis for the dynamic reliability was performed
to see which of the random variables had the highest sensitivity.
For the sensitivity analysis, the COV for each of the random varia-
bles was varied and compared to a baseline reliability for time steps
T2 and T34 with k 500. The baseline reliability, case 1, was
determined by setting the COV for each random variable to 0.0025.
For cases 25, the COV for one of the random variables was
doubled while the COV for the other variables was held xed. Sev-
eral values were calculated in order to compare the reliability
results for each time step. First, the percent change in dynamic reli-
ability was calculated as % change R
RV
R
baseline
=R
baseline
,
Fig. 4 PDF of the position error (rst time step) of the forward kinematic reliability
simulation
Fig. 5 Forward dynamic reliability results for 100 time steps: In trial 2, the stand-
ard deviations, r
s
; r
L
; r
m
; r
Ij
_ _
, of the input random variables, RV
d
fs; L; m; I
j
g,
are double that of trial 1
Journal of Mechanical Design JULY 2013, Vol. 135 / 071008-7
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 11/27/2013 Terms of Use: http://asme.org/terms
where R
RV
was the dynamic reliability for cases 25 and R
baseline
was the baseline reliability. Next, the mean, x
e
, and standard devia-
tion, s
e
, of the end-effector distance error, e
v
, was determined for
each of the cases. A statistical Z-test was performed to see if the dif-
ference in the results was signicant or not [17]. The Z-values were
calculated according to
Z-value
x
A
x
B

s
2
A
s
2
B
N
_ (18)
where x
A
and s
A
were the mean and standard deviations of the dis-
tance error for cases 25, x
B
and s
B
were the mean and standard
deviations of the distance error for case 1, and N was the number
of sample values, Nk 500. Critical Z-values can be found
from a cumulative standard normal distribution table [17]. For the
case in question, the one-tailed critical Z-value at a signicance
level of p 0.001 is 3.090. The critical Z-value at a signicance
level of p 0.025 is 1.96. One-tailed critical Z-values were con-
sidered since it was expected that the mean of the distance error
for cases 2-5 would be greater than the mean of the distance error
of case 1. Thus, if the Z-value is higher than the critical Z-value,
then the cases are signicantly different. Table 3 provides the
detailed results of the sensitivity analysis.
As seen from Table 3, changing the COV of the link lengths
had a signicant effect on the end-effector distance error for both
time steps and changing the COV of the mass had a signicant
effect for the 34th time step. None of the other COV changes
were deemed signicant. However, changing the COVs of the ran-
dom variables at the 34th time step had a more drastic effect on
the dynamic reliability compared to the second time step. There-
fore, it can be said that certain time steps had a higher sensitivity
to the input random variables, causing a decrease in the dynamic
reliability for that time step. It should also be noted that the
dynamic reliability at the 34th time step was lower than at the sec-
ond time step for all cases, including the baseline.
Similar trends in the forward dynamic reliability can be seen
for both trials. This reinforces the idea that the trends were due to
a higher sensitivity to the input random variables for certain time
steps. The CFDR for trials 1 and 2 was calculated as 0.9588 and
0.6708, respectively. Thus, doubling the uncertainty in the input
random variables caused a decrease in the CFDR of 0.288 or
roughly 30%.
The difference in CFDR due to sensitivity to the input random
variables was not surprising, given the results published in Ref.
[14]. They reported cumulative dynamic reliability values of
0.7812, 0.9587, 0.8279, and 0.3982 for four cases of initial joint
angles and joint torques. The kinematic chain used in their simula-
tions was the Stanford arm, a 6 DOF robotic manipulator. If the
calculated dynamic reliabilities of the Stanford arm varied as
much as reported from simply changing the initial state of the
robot, it was not surprising that there was variation in dynamic
reliability across time for a 21 DOF kinematic chain.
Fig. 6 PDFs of the position error (rst time step) of the forward dynamic reliability
simulations
Table 3 Sensitivity analysis
T 2 T34
RV
d
R
d,T
% change in R x
e
(mm) s
e
(mm) Z-value R
d,T
% change in R x
e
(mm) s
e
(mm) Z-value
Baseline 0.96 1.5958 1.1194 0.946 1.9834 1.1319
s 0.948 1.25% 1.6231 1.1587 0.38 0.934 1.27% 2.0888 1.1605 1.45
L 0.674 29.79% 3.2525 2.2135 14.93 0.624 34.04% 3.6581 2.1506 15.41
m 0.954 0.63% 1.5782 1.1162 0.25 0.874 7.61% 2.4444 1.2900 6.01
I
j
0.96 0.00% 1.6043 1.1280 0.12 0.926 2.11% 2.1074 1.1662 1.71
071008-8 / Vol. 135, JULY 2013 Transactions of the ASME
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 11/27/2013 Terms of Use: http://asme.org/terms
6 Conclusion
Forward kinematic and dynamic reliabilities for a 21 DOF
human model were calculated using mean values taken from a
50th percentile female pulling simulation. The minimum and
maximum forward kinematic reliability over the 100 sets of joint
angles were calculated as 0.9529 and 0.9641, respectively. The
joint PDF of the position error (rst time step) for the forward ki-
nematic reliability was a normal distribution with mean of
0.1785 mm and a standard deviation of 0.1076 mm. The mini-
mum and maximum forward dynamic reliability over the 100
sets of joint angles were calculated as 0.931 and 0.964 for trial 1
and 0.529 and 0.703 for trial 2, respectively. The forward
dynamic reliability of the rst time step of the pulling simulation
was 0.965 for trial 1 and 0.693 for trial 2. The joint PDF of the
position error for the forward dynamic reliability (rst time step)
was a normal distribution with mean of 0.1623 mm and a stand-
ard deviation of 0.1091 mm for trial 1 and a normal distribution
with mean of 0.3245 mm and a standard deviation of 0.2183 mm
for trial 2. The CFDR for trials 1 and 2 was 0.9588 and 0.6708,
respectively.
For robotic manipulators, the forward kinematic and dynamic
reliabilities provide the user with a performance criterion inherent
to the manipulator. Reliability for digital human simulations takes
on a different meaning. The procedure detailed here for the calcu-
lation of the forward kinematic and dynamic reliability is useful,
but not as useful as the inclusion of uncertainty into future kine-
matic and dynamic simulations. One of the goals of DHM is to
provide realistic simulations of human beings in real world sce-
narios. One of the criticisms of DHM in the past is that uncer-
tainty is not taken into account. There is inherent uncertainty in
the inputs for DHM problems. Some of the uncertainty is due to
anthropometry, such as height, weight, size, shape, inertia, etc.
Further uncertainty is added to the problem when posture predic-
tion and motion prediction are simulated with human performance
measures as objective functions in an optimization problem. The
inclusion of uncertainty into physical human characteristics and
human performance measures is a critical step in addressing the
limitations of DHM. The calculation of forward kinematic and
dynamic reliabilities presented in this study addresses the
inclusion of uncertainty into a DHM simulation but mainly
addresses the uncertainty in human anthropometry. The next step
is to include uncertainty into posture or motion prediction simula-
tions where human performance measures play a critical role.
Acknowledgment
This work is partly supported by AT&T Chancellors Fellow-
ship, and College of Engineering Deans Fellowship, Texas Tech
University.
Appendix
Table 4 Mean values, l, and standard deviations, r, of the sto-
chastic parameters for calculation of dynamic reliability
l r l r l r
L
i
(m) i s
i
(N m) m
i
(kg)
0.0562670 0.00014067 1 5.695879 0.02848 0.00 0.00000
0.0544274 0.00013607 2 30.723036 0.15362 0.00 0.00000
0.0599962 0.00014999 3 0.183908 0.00092 2.32 0.00580
0.1738560 0.00043464 4 3.956545 0.01978 0.00 0.00000
0.0170799 0.00004270 5 28.218962 0.14109 0.00 0.00000
0.0557820 0.00013946 6 0.170469 0.00085 2.32 0.00580
0.1505700 0.00037643 7 3.075714 0.01538 0.00 0.00000
0.2586380 0.00064660 8 25.106595 0.12553 0.00 0.00000
0.2473740 0.00061844 9 0.178328 0.00089 2.32 0.00580
0.1650990 0.00041275 10 2.266776 0.01133 0.00 0.00000
11 21.848867 0.10924 0.00 0.00000
12 0.202748 0.00101 5.78 0.01445
13 5.482136 0.02741 0.50 0.00125
14 0.976275 0.00488 0.50 0.00125
15 1.412666 0.00706 0.00 0.00000
16 2.034052 0.01017 0.00 0.00000
17 0.040898 0.00020 1.90 0.00475
18 3.011437 0.01506 0.00 0.00000
19 0.052206 0.00026 1.34 0.00335
20 0.321654 0.00161 0.00 0.00000
21 0.130163 0.00065 0.50 0.00125
Table 4 (continued)
l r l r l r l r l r l r
i I
xx
(kg m
2
) I
yy
(kg m
2
) I
zz
(kg m
2
) I
xy
(kg m
2
) I
xz
(kg m
2
) I
yz
(kg m
2
)
1 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
2 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
3 0.029029 0.00005806 0.027871 0.00005574 0.033158 0.00006632 0.000000 0.00000000 0.000000 0.00000000 0.000257 0.00000051
4 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
5 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
6 0.029029 0.00005806 0.027871 0.00005574 0.033158 0.00006632 0.000000 0.00000000 0.000000 0.00000000 0.000257 0.00000051
7 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
8 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
9 0.029029 0.00005806 0.027871 0.00005574 0.033158 0.00006632 0.000000 0.00000000 0.000000 0.00000000 0.000257 0.00000051
10 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
11 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
12 0.056000 0.00011200 0.021600 0.00004320 0.063400 0.00012680 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
13 0.001600 0.00000320 0.000700 0.00000140 0.001100 0.00000220 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
14 0.003100 0.00000620 0.007600 0.00001520 0.007600 0.00001520 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
15 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
16 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
17 0.068014 0.00013603 0.002769 0.00000554 0.067283 0.00013457 0.000172 0.00000034 0.000000 0.00000000 0.000000 0.00000000
18 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
19 0.040714 0.00008143 0.004162 0.00000832 0.040676 0.00008135 0.000324 0.00000065 0.000000 0.00000000 0.000000 0.00000000
20 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000 0.000000 0.00000000
21 0.000769 0.00000154 0.007652 0.00001530 0.007195 0.00001439 0.000171 0.00000034 0.000122 0.00000024 0.000137 0.00000027
Journal of Mechanical Design JULY 2013, Vol. 135 / 071008-9
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 11/27/2013 Terms of Use: http://asme.org/terms
References
[1] Howard, B., Cloutier, A., and Yang, J., 2012, Physics-Based Seated Posture
Prediction for Pregnant Women and Validation Considering Ground and Seat
Pan Contacts, ASME J. Biomech. Eng., 134(7), p. 071004.
[2] Howard, B., and Yang, J., 2012, Calculating Support Reaction Forces in
Physics-Based Seated Posture Prediction for Pregnant Women, Int. J. Rob.
Autom., 27(2), pp. 247260.
[3] Gragg, J., and Yang, J., 2011, Effect of Obesity on Seated Posture Inside a Ve-
hicle Based on Digital Human Models, SAE Int. J. Mater. Manuf., 4(1), pp.
516526.
[4] Strickland, M., Browne, M., and Taylor, M., 2009, Could Passive Knee Laxity
be Related to Active Gait Mechanics? An Exploratory Computational Biome-
chanical Study Using Probabilistic Methods, Comput. Methods Biomech.
Biomed. Eng., 12(6), pp. 709720.
[5] Hughes, R., and Nelson, N., 2009, Estimating Investment Worthiness of an
Ergonomic Intervention for Preventing Low Back Pain From a Firms
Perspective, Appl. Ergon., 40, pp. 457463.
[6] Guo, Z., and de Vita, R., 2009, Probabilistic Constitutive Law for Damage in
Ligaments, Med. Eng. Phys., 31, pp. 11041109.
[7] Galbusera, F., Anasetti, F., Bellini, C., Costa, F., and Fornari, M., 2010, The
Inuence of Axial, Antero-Posterior and Lateral Positions of the Center of
Rotation of a Ball-and-Socket Disc Prosthesis on the Cervical Spine Bio-
mechanics, Clin. Biomech., 25, pp. 397401.
[8] Rohlmann, A., Boustani, H., Bergmann, G., and Zander, T., 2010, Effect of a
Pedicle-Screw-Based Motion Preservation System on Lumbar Spine Biome-
chanics: A Probabilistic Finite Element Study With Subsequent Sensitivity
Analysis, J. Biomech., 43, pp. 29632969.
[9] Hu, L., Zhou, C., and Sun, Z., 2006, Biped Gait Optimization Using Spline
Function Based Probability Model, Proceedings of the 2006 IEEE Interna-
tional Conference on Robotics and Automation, Orlando, FL.
[10] Koshizen, T., 2000a, Sensor Selection by GMB-REM in Real Robot Position
Estimation, J. Intell. Robotic Syst., 27, pp. 275289.
[11] Koshizen, T., 2000b, Improved Sensor Selection Technique by Integrating
Sensor Fusion in Robot Position Estimation, J. Intell. Robotic Syst., 29, pp.
7992.
[12] Koshizen, T., 2001, The Architecture of a Gaussian Mixture Bayes (GMB)
Robot Position Estimation System, J. Syst. Archit., 47, pp. 103117.
[13] Noohi, E., Moradi, H., Noori, N., and Ahmadabadi, M., 2011, Manipulation of
Polygonal Objects With Two Wheeled-Tip Fingers: Planning in the Presence of
Contact Position Error, Rob. Auton. Syst., 59, pp. 4455.
[14] Rao, S., and Bhatti, P., 2001, Probabilistic Approach to Manipulator Kinemat-
ics and Dynamics, Reliab. Eng. Syst. Saf., 72, pp. 4758.
[15] Gragg, J., Yang, J., and Yang, G., 2012, Probabilistic Approach for Digital
Human Kinematic and Dynamic Reliabilities, Proceedings of the ASME 2012
International Design Engineering Technical Conferences & Computers and In-
formation in Engineering Conference, Chicago, IL, August 1215.
[16] Denavit, J., and Hartenberg, R., 1955, A Kinematic Notation for Lower-Pair
Mechanisms Based on Matrices, ASME J. Appl. Mech., 22, pp. 215221.
[17] Haldar, A., and Mahadevan, S., 2000, Probability, Reliability, and Statistical
Methods in Engineering Design, John Wiley & Sons, Inc., New York, NY.
[18] Laz, P., and Browne, M., 2010, A Review of Probabilistic Analysis in Ortho-
paedic Biomechanics, Proc. Inst. Mech. Eng., Part H: J. Eng. Med., 224(H8),
pp. 9273943.
[19] Zou, Q., and Yang, J., 2013, Motion Synthesis for Digital Pregnant Women
Multibody System, Int. J. Rob. Autom., 28(2) (in press).
071008-10 / Vol. 135, JULY 2013 Transactions of the ASME
Downloaded From: http://mechanicaldesign.asmedigitalcollection.asme.org/ on 11/27/2013 Terms of Use: http://asme.org/terms

También podría gustarte