You are on page 1of 12

CHINESE JOURNAL OF MECHANICAL ENGINEERING

Vol. 26, No. 5, 2013 ·839·

DOI: 10.3901/CJME.2013.05.839, available online at www.springerlink.com; www.cjmenet.com; www.cjmenet.com.cn

Dynamic Modeling and Simulation of Multi-body Systems


Using the Udwadia-Kalaba Theory

ZHAO Han1, ZHEN Shengchao1, 2, *, and CHEN Ye-Hwa2


1 School of Mechanical Engineering, Hefei University of Technology, Hefei 230009, China
2 School of Mechanical Engineering, Georgia Institute of Technology, Atlanta 30332, USA

Received December 20, 2012; revised March 20, 2013; accepted March 27, 2013

Abstract: Laboratory experiments were conducted for falling U-chain, but explicit analytic form of the general equations of motion was
not presented. Several modeling methods were developed for fish robots, however they just focused on the whole fish’s locomotion
which does little favor to understand the detailed swimming behavior of fish. Udwadia-Kalaba theory is used to model these two
multi-body systems and obtain explicit analytic equations of motion. For falling U-chain, the mass matrix is non-singular. Second-order
constraints are used to get the constraint force and equations of motion and the numerical simulation is conducted. Simulation results
show that the chain tip falls faster than the freely falling body. For fish robot, two-joint Carangiform fish robot is focused on.
Quasi-steady wing theory is used to approximately calculate fluid lift force acting on the caudal fin. Based on the obtained explicit
analytic equations of motion (the mass matrix is singular), propulsive characteristics of each part of the fish robot are obtained. Through
these two cases of U chain and fish robot, how to use Udwadia-Kalaba equation to obtain the dynamical model is shown and the
modeling methodology for multi-body systems is presented. It is also shown that Udwadia-Kalaba theory is applicable to systems
whether or not their mass matrices are singular. In the whole process of applying Udwadia-Kalaba equation, Lagrangian multipliers and
quasi-coordinates are not used. Udwadia-Kalaba theory is creatively applied to dynamical modeling of falling U-chain and fish robot
problems and explicit analytic equations of motion are obtained.

Key words: Udwadia-Kalaba equation, multi-body systems, falling U-chain, fish robot

developed equations of constrained motion when the


1 Introduction ∗ constraints satisfy D’ Alember’ s principle. In the treatise
on the analytical dynamics, PARS[5] refers to the
The general problem of obtaining equations of motion Gibbs-Appell equations as “probably the most
for constrained discrete mechanical systems has been an comprehensive equations of motion so far discovered”. But
area of considerable interest among scientists and engineers. the Gibbs-Appell equations require a “lucky” choice of
It is also one of the central issues in multi-body dynamics. problem-specific quasi-coordinates and they suffer from
The problem has been aggressively and continuously similar problems when dealing with systems with a large
pursued by many scientists, engineers and mathematicians number of degrees of freedom and many non- integrable
since constrained motion was initially described by constraints. DIRAC[6] used Poisson brackets, a recursive
LAGRANGE[1]. He invented the special Lagrange scheme for determining the Lagrange multipliers, for
multiplier method to deal with constrained motion. singular Hamiltonian systems where the constraints do not
However, the Lagrange multiplier method relies on exactly depend on time.
problem-specific approaches to determine the multipliers; it On the other hand, UDWADIA, et al[7–8], obtained a
is often very difficult to find the multipliers to obtain the concise, explicit set of equations of motion for constrained
explicit equations of motion for systems with large discrete dynamic systems which lead to a simple and new
numbers of degrees of freedom and a mass of fundamental view of Lagrangian mechanics. They derived
non-integrable constraints. GAUSS[2] introduced a new the fundamental equation of motion that describes the
general principle of mechanics for handling constrained dynamics of constrained systems from Gauss’ s principle
motion. Gauss’s Principle gives a clear description of the which seems somewhat less popular than the principles of
general nature of constrained motion through minimization Lagrange, Hamilton, Gibbs and Appell. The equations can
of a function of the accelerations of the particles of a deal with holonomic and also non-holonomic constraints.
system. GIBBS[3] and APPELL[4] have independently UDWADIA, et al[9–10], observed that all the research above
has used D’ Alember’ s principle as their starting point. D’
Alember’ s principle assumes that the forces of constraints
* Corresponding author. E-mail: zhenshengchao@gmail.com
© Chinese Mechanical Engineering Society and Springer-Verlag Berlin Heidelberg 2013
are considered to be ideal and the total work done by the
ZHAO Han, et al: Dynamic Modeling and Simulation of Multi-body Systems
·840· Using the Udwadia-Kalaba Theory

forces of constraints under virtual displacement is always Boltzmann, we are not required to develop a system of
zero. This assumption works well in many situations and is proper quasi-coordinates.
regarded as the core of classical analytical dynamics, but it The significant advantages of Udwadia and Kalaba’ s
is not applicable when the constraints are nonideal. Thus, new theory are threefold. First, it allows descriptions of
UDWADIA and KALABA generalized their previous mechanical systems that use more than the minimum
equations to constrained mechanical systems that may not number of required coordinates which speeds and eases the
satisfy D’ Alember’ s principle. Systems with singular setup of the equations of motion of complex systems.
mass matrices are not common in classical dynamics when Second, it applies to systems whether or not their mass
dealing with unconstrained motion. PARS[5] proposed that matrices are singular, is applicable to systems with
when the minimum number of coordinates is employed for holonomic and/or nonholonomic constraints and to systems
describing the unconstrained motion of mechanical systems, whose constraint forces may or may not be ideal. Third, the
the corresponding Lagrange equations usually yield new theory opens up a new way of modeling complex
non-singular, symmetric and positive definite mass multi-body systems. It permits decomposition of such
matrices. However, singular mass matrices can arise when systems into sub-systems whose equations of motion are
one wants greater flexibility in modeling complex known, and then combine the sub-system equations to get
mechanical systems by using more than the minimum the composite system’ s equations of motion in a
number of required generalized coordinates. Thus, straightforward and simple manner.
UDWADIA, et al[11], developed general and explicit
equations of motion to handle systems whether or not their 2 Explicit General Equations of Motion
mass matrices are singular. We call these discoveries the for Multi-body System
Udwadia-Kalaba theory. About 20 years has passed, we
seldom realize its importance only because it is abstract. In Using Udwadia and Kalaba’ s approach, we first
fact, the Udwadia-Kalaba theory can be applied to consider an unconstrained discrete dynamical system
constrained discrete mechanical systems with unmatched whose configuration is described by the n generalized
ease, clarity and elegance, especially to multi-body systems coordinates q : [q1 ,q2 , ,qn ]T . Its equation of motion
such as mechanical manipulators, space vehicles, robots can be obtained, using Newtonian or Lagrangian mechanics,
and similar systems. by the relation:
As we know, the principles of mechanics[12–16] are so
perfect that it is impossible to create a totally new M (q, t )q  Q (q, q , t ), (1)
fundamental principal for the theory of motion and
equilibrium of discrete, dynamical systems. However, an where M (q, t )  R nn is symmetric and positive definite
additional perspective has been proposed by Udwadia and inertia matrix, q  R n is the velocity and q  R n is the
Kalaba which is useful to help us understand nature’ s law acceleration, and Q (q, q , t )  R n is the force imposed on
from new points of view[7–11]. the system whose constraints are released. The imposed
Using Udwadia and Kalaba’ s marvelous theory, we can forces could include centrifugal force, gravitational force
formulate the explicit, general equations of motion for and control input. The generalized acceleration of the
constrained discrete dynamic systems in three steps:First, unconstrained system, which we denote by a (q, q , t ) is
in terms of the generalized coordinates, we consider the thus given by
unconstrained discrete dynamic system whose equations of
motion can be written using Newtonian or Lagrangian q  M 1 (q, t )Q (q, q , t )  a (q, q , t ). (2)
mechanics. Then we form the constraint equations which
Second, constraints present in the system should be
could include the usual holonomic, non-holonomic,
considered. We shall assume that the system is subjected to
scleronomic, rheonomic, catastatic and acataetatic variaties
h holonomic constraints and m  h nonholonomic
of constraints or even combinations of such constraints.
constraints of the form:
Finally, we impose the additional generalized forces of
constraint on the system. That is to say, we add the force ϕi (q, t )  0, i  1, 2, , h, (3)
resulting from the presence of constraints to the
unconstrained system’ s force. ϕi (q, q , t )  0, i  h  1, h  2,  , m. (4)
The advantage of determining the explicit equations of
motion for constrained discrete dynamical system through We can differentiate standard constraint equations in
this new theory is obvious. Compared to the Lagrange Lagrangian mechanics which are usually in pfaffian form.
multiplier method, there is no need for this new approach to Under the assumption of sufficient smoothness, we take the
determine the multiplier which is often very difficult to
time derivative once on nonholonomic constraints and
obtain for systems having a large number of degrees of
freedom and many non-integrable constraints. Unlike the twice on holonomic constraints, to derive the constraint
formulations offered by Gibbs, Volterra, Appell and equations of the matrix form:
CHINESE JOURNAL OF MECHANICAL ENGINEERING ·841·

UDWADIA and KALABA have proved that the ideal


A(q, q , t )q  b(q, q , t ), (5) constraint force takes the form

where A(q, q , t ) is referred to as constraint matrix and Qidc  M 1/2 B  (b  AM 1Q ), (11)
b(q, q , t ) is a m-vector.
The final step is to form the explicit equations of motion and the non-ideal constraint force takes the form
with constraints. Due to the presence of constraints,
c
additional “generalized forces of constraints” should be Qnid  M 1/2 (I  B + B )M 1/2 c, (12)
imposed on the system. So, the actual explicit equation of
motion of the constrained system could be assumed to take where B  AM 1/2 and the superscript “” denotes the
the form: Moore-Penrose generalized inverse [17–18].
From Eqs. (6), (7), (11), and (12), the explicit equation
M (q, t )q  Q (q, q , t )  Q c (q, q , t ), (6) of motion that governs the evolution of the constrained
system (including both ideal and non-ideal constraints) is
where Q (q, q , t ) , arising by virtue of the holonomic and
nonholonomic constraints, is the additional term of forces Mq  Q +M 1/2 B + (b  AM 1Q ) + M 1/ 2 (I  B + B )M 1/2 c,
imposed on the system. In Lagrangian mechanics,
(13)
Q c (q, q , t ) is considered to be ideal and it is governed by
D’ Alember’ s Principle which indicates that forces of where vector c is determined by the mechanician and could
constraint do zero work under virtual displacements. be obtained by experimentation and/or observation.
However, the constraints can also be non-ideal. Ideal When c is always zero, Eq. (13) reduces to D’
constraints generate ideal constraint forces subject to D’ Alember’ s Principle, which means the total work done
Alember’ s Principle, while non-ideal constraints generate under virtual displacement is zero, the constraints are ideal,
non-ideal constraint forces such as friction force, the constraint force is
electro-magnetic force, etc. If there exist both ideal and
non-ideal constraints in the system, Udwadia put Q c  Qidc  M 1/2 B + (b  AM 1Q ), (14)
Q c (q, q , t ) in the form of
and the explicit equation of motion of the constrained
Q c (q, q , t )  Qidc (q, q , t )  Qnid
c
(q, q , t ), (7)
system (only including ideal constraints) is

where Qidc (q, q , t ) is the ideal constraint force and


c Mq  Q  M 1/ 2 B + (b  AM 1Q ). (15)
Qnid (q, q , t ) is the non-ideal constraint force.
Udwadia generalizes D’ Alember’ s Principle to include
forces of constraint that may do positive, negative, or zero Thus, at each instant of time t, the constrained system is
work under virtual displacement at any instant time during subjected to an additional constraint force F c (t ) given by
the motion of the constrained system. That is, he extends
Lagrange’ s form of D’ Alember’ s Principle to include F c (t )  M 1/2 B + (b  AM 1Q ). (16)
non-ideal constraints. We denote constraint force as
c (q, q , t )  R n . The constraint force does the work When matrix M is a constant diagonal matrix so that
W  v T c (the same as the work done by Q c (q, q , t ) in M  mI, then Eq. (16) simplifies to
any displacement v which is subject to A(q, q , t )v  0.
So we write these equations in the form of F c (t )  MA+ (b  AM 1Q ). (17)

W  v T Q c  v T c, (8) When M 1Q is zero, Eq. (17) becomes

which is the extended Lagrange’ s form of D’ Alembert’ s F c (t )  MA+ b. (18)


Principle (We henceforth omit the arguments of functions
where there is no confusion in our notations). The work For Eq. (13), if the mass matrix M is singular, we should
done by the ideal constraint force Qidc under virtual use the Udwadia-Phohomsiri equation[11] instead of
displacements is Udwadia-Kalaba equation[7–10] to get the equation of
motion for the constrained system. Hence, the equation of
v T Qidc  0, (9) motion is

while the work done by non-ideal constraint force is   I  A+ A M + Q 


    .
q   
    (19)
v T Qnid
c
 0. (10)  A   b 
ZHAO Han, et al: Dynamic Modeling and Simulation of Multi-body Systems
·842· Using the Udwadia-Kalaba Theory

The above equation is valid when the matrix dependence of the velocity and the acceleration of the chain
T
 M | A has full rank[11]. This full rank condition can serve tip for several initial conformations of the chain and also
as a check on whether we have obtained a correct model, give us a simple analytical model of the system.
because it is also the condition required for the equation of As seen above, research on falling U-chain problem has
motion of the constrained system to be unique. seemingly been conducted well for years. However, they all
Remark. UDWADIA and KALABA have provided focus their attention on whether the system is energy
explicit general equations of motion for constrained
conservative or non-conservative, results of laboratory
discrete dynamical systems with their newly initiated
experiments and numerical simulations based on a simple
approach which is applicable to all holonomical and non-
analytical model and arguments with each other. None of
holonomical constrained systems no matter whether they
them attempt to present the explicit analytic form of the
satisfy D’ Alember’ s Principle. They generalize
Lagrangian mechanics to include both ideal and non-ideal general equations of motion for the falling U-chain system.
constraint forces by using a new fundamental principle In this paper, we creatively apply the Udwadia-Kalaba
governing the motion of constrained systems. The equation theory to solve the falling U-chain problem and to get its
of motion obtained by using Udwadia and Kalaba’ s theory explicit analytic form of the general equations of motion.
is general, simple and understandable.
4 Analytical-form Equations of Motion
3 Case 1: Falling U-chain of the Falling U-chain

The falling U-chain (Fig. 1) is a folded flexible heavy We apply Udwadia and Kalaba’ s theory to the falling
chain with one end suspended from a rigid support and the U-chain problem to formulate the analytical-form equation
other end lifted up to form a U-shaped fold at the bottom, of motion. We do not focus on the inclined angle of each
then the lifted end is released to move down in the manner link. Therefore, to simplify, we assume the U-chain
of a bungee fall[19]. The whole chain is in the gravitational consists of n particles which are connected with n–1
field while falling. This U-chain case is actually so old that
massless links. The length of each link is 1, and the mass of
it has been studied frequently by many dissertations[20–25].
each particle is mi. Through this simplification, we can
We are interested in it because it is an ideal constrained
describe the position of each particle with two degrees of
dynamic system. Also, compared to the bottom-pile chain
freedom, thus it will make the process of obtaining the
and top-pile chain, the U-chain is conceptually easier to
model because there are no actual collisions, no made or analytical-form equation of motion easier and speed up the
broken contacts and thus no explicit collisions even in a simulation. The chain is constrained to move only in the
discrete chain. This model provides ease and reliability for vertical plane with each particle denoted by (xi, yi), i1, 2,
numerical simulation[19].  , n. All links are considered to be rigid and cannot be
deformed. Consecutive particles are connected by massless
links regardless of the friction, since the friction in the
falling U-chain system is small and makes no difference to
the performance of the dynamic system.

4.1 Falling U-chain: unconstrained equation of motion


If there are no constraints placed on the links,
connections between links are removed and the last link is
Fig. 1. Falling U-chain not attached to the fixed support. The unconstrained
U-chain’ s equation of motion can be obtained easily as
WONG, et al[19], have given us a detailed and critical
review of the history of falling U-chain problems which Mq  Q (q, q , t ), i  1, 2,  , n, (20)
mainly include some erroneous approaches. They identify
 0 
the source of error and propose a fool-proof Lagrangian m1 0 0 0  0


approach to the falling U-chain problem. They imply that  0 m1 0 0  0 0 
the method of SOUSA, et al[23], is not reliable because its  
 0 0 m2 0  0 0 
solution is an erroneous for the falling U-chain and they  
also conclude that Lagrange’ s method gives definitive M   0 0 0 m2  0 0  ,
 
answers which are better and more easily developed.  
     0 0 

TOMASZEWSKI, et al[24], analyze the dynamics of the tip 
 0 0 0 0 0 mn 0 
of the falling U-chain with laboratory experiments and  
numerical simulations. They have determined the time 0 0 0 0 0 0 mn 
CHINESE JOURNAL OF MECHANICAL ENGINEERING ·843·

    0 
 x1     2x1 x2 − x12 − x22 +2 y1 y 2 − y12 − y 22 
     
  y1   1  m g 2 2 2 2
 2x2 x3 − x − x +2 y 2 y3 − y − y 
    2 3 2 3
  x2   0  
b 2x3 x4 − x32 − x42 +2 y3 y 4 − y32 − y 42

     . (27)
q    y  , Q  m2 g , (21)   
 2     
      2 2 2 2 2 2 
     2xn1 xn − xn1 − xn +2 y n1 y n − y n1 − y n − xn − y n 
    
 xn   0 
   
yn 
  mn g 4.3 Udwadia-Kalaba equation
Considering Eq. (15), detailed Udwadia and Kalaba
where m1, m2,  , mn are the mass of each link, x1, x2,  , eqution of the whole U-chain can be rewritten as
xn, y1, y2,  , yn are coordinates, Q is the external force.
Mq  Q  M 1/2 (AM −1/2 ) + (b − AM −1Q ), (28)
4.2 Falling U-chain: constraint equation
For i  1, 2,  , n 1 , the constraint between i and i1 is
where M , q, Q can be found in equation (21), A in equation
2 2 2
(xi+1  xi ) +(yi+1  yi )  l . (22) (26) and b in Eq. (27). This is the explicit analytical-form
equation of motion for the whole falling U-chain.
Taking the time derivative twice, we get To summarize, we can derive the analytical-form
equation of motion of the whole U-chain: form the
(xi  xi+1 )
xi +(yi  yi+1 )
yi +(xi+1  xi )
xi +1 +(yi+1  yi )
yi +1  unconstrained equation of motion at first, then write the
constraint equation so that the additional imposed
 y i2  y i2+1 +2yi+1 yi  x i2  x i2+1 +2xi+1 xi (23)
constrained force can be derived, and finally add the
constrained force to the equation of motion of the
For in, there is constraint from the n–1 link and the other unconstrained system.
constraint is

xn2 + yn2  l 2 . (24)


5 Numerical Simulation of the Falling
U-chain and analysis

Taking the time derivative twice, we get


Through simulations, insights into the dynamics of the
falling U-chain can be developed. To simplify, we assume
xn  yn  xn2  y n2 .
xn + yn  (25) li 1, g9.8, mi 1 for i1, 2,  , n and n17. We choose
n17 through trial and analysis. n17 is large enough to
simulate the dynamics and can get good results. Since we
Considering all the constraints which have been
know the whole U-chain’ s unconstrained equation of
described above, the whole U-chain’s constraint equation
motion and the constraint equation, we now know how to
Aq  b can be derived as follows:
calculate the constrained force, and we can simulate the
falling U-chain system in Matlab. We select the ode15i
A algorithm to implement the numerical simulation. It is a
 x1  x2 y1  y2 x2  x1 y2  y1
 0 0 0 0 variable order method that can solve fully implicit
 0 0 x2  x3 y2  y3 x3  x2 y3  y2 0 0 differential equations.

 0 0 0 0 x3  x4 y3  y4 x4  x3 y4  y3 We give five different initial separations xD between two
 ends of the chain: (a) x012, (b) x013, (c) x014 (d)
        
 x015, (e) x016.
 0 0 0 0 0 0 0 0
 Table 1 provides the maximum vertical fall distance hmax
 0 0 0 0 0 0 0 0 of the falling chain tip and the time it takes of each case, also,
it provides the vertical fall distance h of the freely falling
 0 0 0 0  body. From Table 1, we can determine if two ends of the

 0 0 0 0  chain get closer, the maximum vertical fall distance of the
 0 0 0 0  chain tip is reduced, and the time it takes gets less. In cases
, (26) (a), (b), (c), (d), (e), the vertical fall distance of the chain tip,
     
 xn -1 − xn yn -1 − yn xn − xn -1 yn − yn -1  up to the time tmax at which the vertical fall distance of the
 chain tip reaches its maximum value hmax, is seen to always
 0 0 xn yn 
be ahead of the vertical fall distance h of the freely falling
ZHAO Han, et al: Dynamic Modeling and Simulation of Multi-body Systems
·844· Using the Udwadia-Kalaba Theory

body. This can be summarized by the general statement that


the chain falls faster than a freely falling body.

Table 1. Chain tip’s five cases


Initial Maximum fall Corresponding Fall
Case separation distance time distance
x0mm hmaxmm tmaxmm hmm
a 12 15.943 1.635 13.099
b 13 15.990 1.685 13.912
c 14 16.070 1.749 14.989
d 15 16.323 1.802 15.911 Fig. 3. Vertical force Fy1 on the chain tip of case (b)
e 16 16.671 1.842 16.625

In the simulation, we also find that the vertical distance


of a freely falling ball is always less than that of the falling
U-chain tip before the time tmax in all the five cases which
reconfirm the conclusion above.
To further understand this behavior, we analyze the time
dependence of the vertical acceleration ay1 of the chain tip
(horizontal acceleration is not so important for the analysis).
Downward is the positive direction and Fy1 denotes the
vertical force imposed on the falling chain tip. We know
m1, so Fy1 is numerically equal to ay1. We performed Fig. 4. Vertical force Fy1 on the chain tip of case (c)
simulations also with five initial separations between the
ends of the chain (a) x012, (b) x013, (c) x014, (d) x0
15, (e) x016. Figs 2, 3, 4, 5, 6 show that as the initial
separation between the chain ends is reduced, the time of
the chain tip’s vertical acceleration larger than g (vertical
acceleration of the freely falling body, equal to 9.8) gets
more, and the value(about from 10 to 25) of chain tip’s
vertical acceleration ay1 larger than g generally gets bigger.
In other words, as the initial separation between chain ends
is reduced, Fy1 and ay1 grow and the vertical velocity of the
falling chain tip grows.
Fig. 5. Vertical force Fy1 on the chain tip of case (d)
Furthermore, two other characteristic features in Figs. 2
–6 are the negative peak heights of the vertical acceleration
and the time at which these maxima occur. As the initial
separation lessens, the negative peak heights lessen and the
time lessens. Next, we examine the values of the negative
peak heights of the vertical acceleration of the falling chain
tip as shown in Figs. 2–6. They are all large between –200
and –350 (far more than g). This means the tip is subjected
to a great pulling force when hitting the lowest position.

Fig. 6. Vertical force Fy1 on the chain tip of case (e)

6 Case 2: Fish Robot


There are over 28 000 species of fish and a wide variety
of propulsive systems used by fish for maneuvering in the
aquatic environment. Fish have numerous fins which act to
transfer momentum to the surrounding water and shed
reverse Karman vortex street in the wake. The undulating
Fig. 2. Vertical force Fy1 on the chain tip of case (a) motion of fish can generate forward thrust as well as
CHINESE JOURNAL OF MECHANICAL ENGINEERING ·845·

receive drag from the water[26]. They exhibit amazing waters, while anguilliform fish exhibit remarkable
swimming capabilities: high propulsive efficiency, maneuverability. We focus on the robot model of
extraordinary maneuverability, long duration and station carangiform fish which typically have large, high-aspect-
keeping ability[27]. Fish’ s excellent locomotion ability ratio tails. They swim primarily using the rear and tail,
under water has stimulated growing interests in biomimetic while the front body remains moving forward and
robot fish. Many researchers have attempted to study the swinging.
mechanism of fish swimming and have developed We can idealize the main body of the fish as a rigid
underwater fish robots which can be prospectively applied body. The body is connected to the tail by a peduncle (a
in military detection, undersea operation, reconnaissance slender region of generally negligible hydrodynamic
de-mining and so on[28]. influence). Fish tails are usually flexible, but the caudal fins
We review briefly related research work on the fish’ s of carangiform fish are very stiff. So, we treat the tail as a
swimming mechanism. Most of the early research focused rigid lifting surface. For simplicity, we analyze the
on building hydrodynamic models. TAYLOR[29] employed motion of the robotic fish only in two dimensions of the
steady-state flow theory to calculate the fluid force. horizontal plane. The fish-like mechanism can be modeled
LIGHTHILL[30] developed the elongated-body theory to as a three link system as shown in Fig. 7.
study the swimming of slender fish. A two-dimensional
waving plate theory was originally proposed by WU[31]. He
advised to study the swimming fish as an elastic plate.
Thereafter, a large-amplitude elongated-body theory was
developed to analyze rapid acceleration and steady
swimming. An overview of fish swimming modes for
aquatic locomotion was presented by SFAKIOTAKIS, et
al[32]. TRIANTAFYLLOU, et al[33], initiated significant
work to develop an artificial robot fish. They developed an
eight-link, foil-flapping robotic fish mechanism(RoboTuna)
Fig. 7. Fish-like mechanism model
and investigated drag forces experimentally. Subsequently,
more fish robots have been developed, e.g., the well-known To conduct a hydrodynamic analysis in a relatively
Mitsubishi robotic fish, the lamprey robot, the ISRobotics simple way, we model the three-link fish robot using
Ariel Robot, the robotic Blackbass, etc. carangiform propulsion in which the fish oscillates the
As seen above, research on swimming fish has caudal fin and peduncle. The drag acts on all the three
seemingly been conducted well for years. But they all focus links, but mainly on the body. We assume the flow is
on the whole fish’ s locomotion which does little to aid in inviscid and calm. The second link(peduncle) is usually
understanding the detailed swimming behavior of fish. hydrodynamically negligible, but here we only consider the
They do not present the explicit analytic form of the drag in the lateral direction. The third link(caudal fin) is
general equations of motion for the swimming fish system. regarded as a thin flat plate generating lift and lateral drag
In this paper, we show the detailed locomotion of each part according to quasi-static two-dimensional wing theory
of the fish robot and the explicit analytic form of the which is often used to approximately calculate fluid forces
general equations of motion and its derivation. acting on a high-aspect-ratio wing. Because z-component is
The fish robot can be regarded as a multi-body system usually not considered in steady swimming, the robotic
or a constrained discrete mechanical system. In order to fish’ s locomotion is analyzed only in two dimensions of
obtain explicit equations of motion of the fish robot system the horizontal plane. Thus, we refer to the direction of
and get the propulsive characteristics of each part, we intended locomotion longitudinal to the fish body as the x
choose Udwadia-Kalaba theory to model the multi-body direction and the lateral direction as the y direction. Fig. 8
system. Compared with former simple model, the model illustrates an analytical model of the fish robot’s dynamics.
built by applying Udwadia-Kalaba theory is more
reasonable and can clearly describe the detailed locomotion
of each part of the fish robot, not only the whole fish
robot’s locomotion described in the former simple model.

7 Fish Robot Model

The fish’ s nervous system controls muscle contraction


making the fish’ s body and fins undulate. These motions
generate thrust by transferring momentum to the Fig. 8. Analytical model of fish robot dynamics
surrounding water. Of all the fishes’ swimming modes,
carangiform fish maintain high-speed swimming in calm We introduce the following coordinate system,
ZHAO Han, et al: Dynamic Modeling and Simulation of Multi-body Systems
·846· Using the Udwadia-Kalaba Theory

coordinates, variables and symbols. We take the quasi-steady wing theory to approximately
x axis: longitudinal to the fish body; y axis: lateral calculate the lift force because the caudal fin can be
direction; x1, y1, x2, y2, x3, y3: coordinates of the mass center; regarded as a high-aspect-ratio thin plate. Although this
θ1 ,θ 2 ,θ3 : undulating angles of the three links; l2: length of model of calculating forces acting on the fish body is
the second link; FD1x: longitudinal drag acting on the first substantially simplified, it does a very good job of
link; FD1y, FD2y, FD3y: lateral drag acting on the three links; describing the qualitative behavior of the system.
Fx, Fy: x, y component of the lift force; T2, T3: input torques
exerted on the two joints; a: length from mass center of 9 Constrained Equation of Motion of Fish
first link to first joint; b: length from mass center of third Robot System
link to second joint.
We begin to derive the constrained equation of motion
8 Drag and Lift Force from this point by applying Udwadia-Kalaba theory. The
fish robot system has three links and two joints. It can be
The drag acting on the fish is empirically measured by regarded as a common multi-body system. So, the
towing the fish through the water and is found to be constrained equation of motion can be obtained through
approximately quadratic with velocity[34]. It can be Udwadia-Kalaba equation which has been introduced in
calculated by detail.

FD 0.5ρ CDV 2 S , (29) 9.1 Equations of Motion of Unconstrained fish robot


If there are no constraints exerted on the three links, in
other words, connections between links are removed.
where ρ is the mass density of water, V is the velocity of
Referring to the analytical model of the fish robot’ s
the link of fish robot relative to the water flow, it can be in
dynamics (Fig. 8), we can get equations of each link as
the direction of x axis or y axis. CD is the drag coefficient, S
is the projection area of the link of the fish robot on the
plane perpendicular to the velocity. m1  y1 FD1y , I1θ1 0,
x1 FD1x , m1  (35)
We assume that the caudal fin is in a quasi-steady
m2  y2 FD2y , I 2θ2 T2 ,
x2 0, m2  (36)
uniform flow. Following the notations of MASON[35] and
YU[28], lift force acting on the caudal fin is represented as m3  y3 Fy +FD3x , I 3θ3 T3 ,
x3 Fx , m3  (37)

L 2πρ lC (v3  le )  v3 , (30)


where m1, m2, m3 are the mass of each link and I1, I2, I3 are
the moment of inertia. Because we will constrain θ2 and θ3
where l is the span and C is half of the chord length, ρ is
respectively to a definite function later in the constraint
the water density. le is a unit vector pointing in the direction
of the leading edge of the caudal fin which can be equation, here equation I 2θ2 T2 and I 3θ3 T3 should be
represented as removed. After mathematical manipulation, the
unconstrained equation of motion for the robot system can
le  cosθ3 i  sinθ3 j , (31) be written in matrix form as

M (q, t )q Q (q, q , t ), (38)


where i is the unit vector along x axis and j is the unit
vector along y axis. v3 is the relative velocity at the center
of the caudal fin, which can be represented as M : M (q, t ) 
m1 0 0 0 0 0 0 0 0 
 
v3  x3 i  y3 j , (32) 
 0 m1 0 
 
 0 I1 0 0 
where x3 is x component velocity and y3 is y component  
velocity at the center of the caudal fin. Substituting Eqs. (31),  0 m2 0 
 
(32) into Eq. (30), we can derive x component lift force  0
 m2 0  ,


 0 0 I2 0 
Fx 2πρ lC (x3 y3sinθ3  y32 cosθ3 ), (33)  
 0 m3 0 
 
and y component lift force
 0 m3 0 
 
 0 
0 0 0 0 0 0 0 I 3 
Fy 2πρ lC (  x32 sinθ3   x3 y3cosθ3 ). (34)
CHINESE JOURNAL OF MECHANICAL ENGINEERING ·847·

 x1   FD1x  Aq b, (50)


    
   
 1  y
   F D1y

 1 0 asinθ1 1 0
l
 2 sin θ 2 0 0 0
      2 
θ1    0   
   
   l2
  x2   0  0 1 a cos θ1 0 1 cos θ 2 0 0 0 
     2 
   
q  y2  , Q :Q (q, q, t )  FD2y .
      (39)  l 
   1 0 b sin θ3  ,
    A  0 0 0 1 0  2 sin θ 2
θ2   0   2 
 

  
l2 
  
x
 3 
  Fx  0
 0 0 0 1 cos θ 2 0 1 b cos θ3 
   2 
 
 y3   Fy + FD3x   
  0 0 0 0 0 1 0 0 0 
     
θ3   0 
0 0 0 0 0 0 0 0 1 
9.2 Constraint Equations
 aθ 2 cosθ  (l  2)θ 2 cosθ 
For fish robot system, we get three links with two joints  1 1 2 2 2 
  2 
which generate four constraint equations. Also two more  1 aθ sin θ  (l  2)θ 2
sin θ 2


1 2 2 
constraint equations are used to constrain θ2 and θ3 .  (l2  2)θ2 2 cosθ 2  bθ32 cosθ3 
For the first joint, the constraint equations are b   . (51)
(l  2)θ 2 sin θ  bθ 2 sin θ 
 2 2 2 3 3
 
x1 +acosθ1  x2  (l22)cosθ 2 , (40)  0.3(2 π ) 2
sin(2π t )  
 
y1 asinθ1  y2  (l22)sinθ 2 . (41) 
 0.6(2π) 2 sin(2πt  0.25) 

Taking the time derivative twice, we get


9.3 Constrained equation of motion
l l
x1  x2  asinθ1θ1  2 sinθ 2θ2 aθ12 cosθ1  2 θ2 2 cosθ 2 , (42) Eq. (38) constitutes the unconstrained equation of
2 2
motion of the fish robot system, and Ep. (50) specifies the
l l
y1  y2 +acosθ1θ1  cosθ 2θ2 aθ1 sin θ1  θ2 2 sin θ 2 . (43)
2 2 2
constraints. Since matrix M is singular, we should use the
2 2 Udwadia-Phohomsiri equation to get the equation of
For the second joint, the constraint equations are motion for the constrained system. We have checked the
full rank condition and it is fulfilled: the rank of
x2  (l22)cosθ 2  x3  bcosθ3 , (44) [M | A]T is 9 (full rank). Hence, the acceleration of the fish
y2 (l22)sinθ 2  y3  bsinθ3 . (45) robot system is

Taking the time derivative twice, we get   I  A+ A M + Q 


    .
q  
    (52)
 A   b 
l l
x2  
 x3  2 sinθ 2θ2  bsinθ3θ3  2 θ2 2 cosθ 2 +bθ32 cosθ3 ,
2 2
(46) For this fish robot system, we can explicitly find closed
l2 l form expressions for M and Q in (39), A and b in (51).
y2  y3  cosθ 2θ2  bcosθ3θ3  2 θ2 2 sin θ 2  bθ32 sin θ3 .
2 2 Having thus obtained the matrices M, Q, A and b, the
(47) equations of motion of the system can be simply generated
There are two additional constraint equations: by substitution in the right-hand side of Eq. (52). We get
nine coupled equations which describe the motion of the
θ 2 0.3sin(2πt ), (48) constrained fish robot system. Using the explicit
expressions for M, Q, A and b, the nine coupled equations
θ3 0.6sin(2πt  0.25).
can be obtained directly in Maple. The results cover several
pages. The complexity of the results precludes explanation
Taking the time derivative twice, we get in this paper. However, through simulation, an insight into
the dynamics of the fish robot system can be obtained.
θ2  0.3(2π) 2 sin(2πt ), (49)
2 10 Numerical Simulation
θ3  0.6(2π) sin(2πt  0.25).

It is necessary to first verify the model we have


After mathematical manipulation, the constraint equations constructed. Therefore, numerical simulations have been
can be written in matrix form as made and the results detailed. Through a series of
ZHAO Han, et al: Dynamic Modeling and Simulation of Multi-body Systems
·848· Using the Udwadia-Kalaba Theory

manipulations, Eq. (52) is obtained which include nine fish’ s swimming behavior. The fish’ s main propulsive
coupled equations. The solution of the equations can be characteristics include forward velocity, acceleration and
obtained by solving the initial value problems for ODE swing angle of the body. We change the amplitude,
through ode15i function in Matlab. frequency and phase difference to get the different
simulation results and then compare them to draw some
10.1 Parameters and initial condition conclusions. Of course, if we change the amplitude,
We assume that the fish robot is one kind of the frequency and phase difference, initial condition should
carangiform fishes and that it is a fish of small size in a
also be changed accordingly.
calm water. Table 2 shows the parameter values used in the
The influence of the amplitude of caudal fin's oscillating
simulation. The initial conditions including initial
on forward velocity, acceleration and swing angle of fish’s
coordinate, velocity and acceleration of each link are
presented in Table 3. main body: We keep A2  0.3, ω  2π, φ  15o , but
A3 varies with the number of 0.1, 0.2, 0.4, 0.6. Figs. 9–12
show that the forward velocity increases very quickly in the
Table 2. Parameters in simulation beginning, but as the fish speeds up, the drag force
Object Variable Value increases. The velocity tends to become fixed velocity in
Moment of inertia I1  (kg  m 2 ) 1.23×10–3
the end. The forward velocity fluctuates slightly around a
Mass of 1st link m1kg 0.409
1st link Length of link 2am 0.1875 mean value. This phenomenon is caused by the unsteady
Projected area to vertical plane Sm 2 7.069×10–3 thrust force coming from inherent oscillations in fish itself.
Projected area to horizontal plane S1m 2 0.19 When the oscillating amplitude of caudal fin increases, the
Moment of inertia I 2  (kg  m 2 ) 3.7×10–5 forward velocity of fish robot increases accordingly which
2nd Mass of 2nd link m2kg 0.104
link Length of link l2m 0.062 5
are consistent with the swimming performance of this type
Projected area to horizontal plane S 2m 2 0.04 of fish. Figs. 13–16 show that the starting acceleration (i.e.
Moment of inertia I 3  (kg  m 2 ) 6.75×10–7 propulsive force) increases along with oscillating amplitude
3rd Mass of 3rd link m3kg 9.0×10–3 and tends to zero because of the increasing drag force. The
link Length of link 2bm 0.03
fluctuation around zero can be explained by the unsteady
Projected area to horizontal plane S3m 2 0.008
Density of water ρ  (kg  m3 ) 998
thrust force. The frequency of the oscillating swing angle of
Fluid Drag coefficient CD 0.5 the fish's main body is the same as that of caudal fin and
force Fin span lm 0.075 the amplitude of the oscillating swing angle increases
Fin chord length 2Cm 0.03
slightly as the caudal fin's oscillating amplitude increases.

Table 3. Initial condition of each link


Object Variable Value
(x1 , y1 , θ1 ) (0, 0, 0)
1st
(x1 , y1 , θ1 ) (0, 0, 0)
link
y1 , θ1 )
x1 , 
( (0, 0, 0)
(x2 , y2 , θ 2 ) (a l2 /2 , 0, 0)

2nd (  l2θ2 2  bsinθ3θ3  bθ32cosθ3 , bcosθ3θ3 


(x2 , y 2 , θ2 )
link bθ32sinθ3 ,  A3ω 2sin(  φ ))
y2 , θ2 )
x2 , 
( (  l22θ2 , 0,0)
(x3 , y3 , θ3 ) (a +l2 +bcosθ3 , bsinθ3 , A 3 sin(  φ ))
3rd (x3 , y3 , θ3 ) (  bθ3sinθ3 , l2θ2 +bcosθ3θ3 , A3ω cos(  φ )) Fig. 9. Forward velocity of amplitude 0.1
link (  l2θ2 2  bsinθ3θ3  bθ32cosθ3 , bcosθ3θ3 
y3 , θ3 )
x3 , 
(
bθ32sinθ3 ,  A3ω 2sin(  φ ))

10.2 Results and discussions


In this simulation, we assume the peduncle and caudal
fin’ s oscillating functions are θ 2  A2 sin(ωt ), θ3 
A3sin(ω t  φ ). The peduncle and caudal fin oscillate with
the same frequency but different amplitudes and there
exists a phase difference between the two oscillations. We
investigate the influence of caudal fin’ s oscillating on the Fig. 10. Forward velocity of amplitude 0.2
CHINESE JOURNAL OF MECHANICAL ENGINEERING ·849·

Fig. 11. Forward velocity of amplitude 0.4 Fig. 16. Acceleration of amplitude 0.6

The influence of the frequency of caudal fin’s oscillating


on forward velocity, acceleration and swing angle of fish’s
main body: in this simulation, A2  0.3, A3  0.6, φ  15o ,
but the oscillating frequency of the peduncle and caudal fin
ω changes with 2π, 2.5π, 3π, 3.5π . We find that the final
forward velocity and the oscillating frequency of the final
forward velocity increase with the frequency of the caudal
fin's oscillating increasing (figures omitted). Starting
acceleration increases significantly when ω increases and
Fig. 12. Forward velocity of amplitude 0.6 the frequency of final acceleration’ s oscillation around
zero also increases (figures omitted). ω has little influence
on the oscillating amplitude of swing angle. The oscillating
frequencies of peduncle, caudal fin and the main body are
the same.
The influence of phase difference of caudal fin’ s
oscillating on forward velocity, acceleration and swing
angle of fish’ s main body: the influence of phase
difference φ on the swimming behavior is not very
obvious as is the influence of oscillating amplitude and
frequency of caudal fin. The final forward velocity (figures
Fig. 13. Acceleration of amplitude 0.1 omitted), starting acceleration and oscillating amplitude of
swing angle of the fish’ s main body decreases as the phase
difference increasing.

11 Conclusions

Udwadia-Kalaba theory is applied to analyze the


dynamics of the falling U-chain tip and 2-joint Carangiform
fish robot.
(1) The explicit analytical form of general equations of
motion of the whole U-chain system is obtained.
Fig. 14. Acceleration of amplitude 0.2 (2) Through numerical simulation, ideal and interesting
insights are obtained: as the chain ends are moved closer
together, the maximum vertical fall distance of the chain tip
and the falling time decrease. The chain will fall faster than
freely falling body and the vertical acceleration will always
be higher than g9.8. As the separation becomes less, the
chain tip will fall faster. The chain tip is subjected to a
greater pulling force which may be up to 40 times larger
than one link's gravitational force when it hits the lowest
position.
(3) The explicit analytical form of general equations of
Fig. 15. Acceleration of amplitude 0.4 motion of the 2-joint Carangiform fish robot is obtained.
ZHAO Han, et al: Dynamic Modeling and Simulation of Multi-body Systems
·850· Using the Udwadia-Kalaba Theory

The dynamic model including the forward velocity, [24] TOMASZEWSKI W, PIERANSKI P, GEMINARD J C. The motion
of a freely falling chain tip[J]. American Association of Physics
acceleration and also the swing angle of fish body is Teachers, 2006, 74(5): 776.
derived which is a characteristic difference between this [25] GREWAL A, JOHNSON P. A chain that speeds up, rather than
dynamic analysis and other conventional analysis of the slows, due to collisions: How compression can cause tension[J].
American Association of Physics Teachers, 2011, 79(7): 723.
dynamics of Carangiform fish robot. [26] CHEN H, ZHU C A. Modeling the dynamics of biomimetic
(4) By solving Udwadia-Kalaba equation, the swimming underwater robot fish [C]//IEEE International Conference on
performance of the fish robot is obtained. The phenomenon Robotics and Biomimetics, Hong Kong, July 5–9, 2005: 478–483.
[27] LIU L Z, YU J Z, WANG L. Dynamic modeling of
that fish gain burst propulsive force, acceleration and a high three-dimensional swimming for biomimetic robotic fish[C]//
velocity through raising oscillating amplitude and International Conference on Intelligent Robots and Systems, Beijing,
frequency of caudal fin agrees with the simulation. Raising October 9–15, 2006: 3 916–3 921.
[28] YU J Z, LIU L Z, WANG L. Dynamic modeling and experimental
the oscillating amplitude and frequency of the caudal fin
validation of biomimetic robotic fish[C]//Proceedings of the 2006
will result in an increasing of the main body’s swing angle. American Control Conference, Minnesota, June 14–16, 2006:
4 129–4 134.
References [29] TAYLOR G. Analysis of the swimming of long narrow animals[J].
[1] LAGRANGE J L. Mechanique analytique[M]. Paris: Mme ve Proceedings of the Loyal Society A: Mathematical, Physical &
Courcier, 1787. Engineering Science, 1952, 214(1 117): 158–183.
[2] GAUSS C F. Uber ein neues allgemeines Grundgsetz der [30] LIGHTHILL M J. Note on the swimming of slender fish[J]. The
Mechanik[J]. Journal of die reine und angewandte Mathematik, Journal of Fluid Mechanics, 1960, 9(2): 305–317.
1829, 4: 232–235. [31] WU T Y. Swimming of a waving plate[J]. The Journal of Fluid
[3] GIBBS J W. On the fundamental formulae of dynamics[J]. The Mechanics, 1961, 10(3): 321–344.
American Journal of Mathematics, 1879, 2(1): 49–64. [32] SFAKIOTAKIS M, LANE D M, DAVIES J B C. Review of fish
[4] APPELL P. Sur une Forme Generale des Equations de la swimming modes for aquatic locomotion[J]. IEEE Journal of
Dynamique[J]. Comptes rendus de l'Academie des sciences, 1899, Oceanic Engineering, 1999, 24(2): 237–252.
129(1): 459–460. [33] TRIANTAFYLLOU M S, TRIANTAFYLLOU G S. An efficient
[5] PARS L A. A treatise on analytical dynamics[M]. Connecticut: Ox swimming machine[J]. Scientific American, 1995, 272(3): 64–70.
Bow Press, 1965. [34] CHAN W L, KANG T. Simultaneous determination of drag
[6] DIRAC P A M. Lectures in quantum mechanics[M]. New York: coefficient and added mass[J]. IEEE Journal of Oceanic
Yeshiva University, 1964.
Engineering, 2011, 36(3): 422–429.
[7] UDWADIA F E, KALABA R E. A new perspective on constrained
[35] MASON R. Fluid locomotion and trajectory planning for
motion[J]. Proceedings of the Royal Society, 1992, 439(1 906):
shape-changing robots[D]. California Institute of Technology, 2003.
407–410.
[8] UDWADIA F E, KALABA R E. Analytical dynamics: A new
approach[M]. Cambridge, UK: Cambridge University Press, 1996. Biographical notes
[9] UDWADIA F E, KALABA R E. Explicit equations of motion for ZHAO Han, born in 1957, is currently a vice-president and a
mechanical systems with nonideal constraints[J]. Journal of Applied doctoral supervisor at Hefei University of Technology, China. He
Mechanics, 2001, 68(3): 462–467. is a committee member of International IFToMM Education
[10] UDWADIA F E, KALABA R E. On constrained motion[J]. Applied
Commission and also a member of Editorial Board of Chinese
Mathematics and Computation, 2005, 164(2): 313–320.
[11] UDWADIA F E, PHOHOMSIRI P A. Explicit equations of motion
Journal of Mechanical Engineering. His research interests include
for constrained mechanical systems with singular mass matrices and mechanical transmission, magnetic machine, vehicles, digital
applications to multi-body dynamics[J]. Proceedings of the Royal design and manufacturing, information system, dynamics and
Society, 2006, 462(2 071): 2 097–2 117. control.
[12] ROUSEBALL W W. A short account of the history of Tel: 86-13805697995; E-mail: hanzhaoff@qq.com
mechanics[M]. New York, NY: Dover Publications, 1960.
[13] SHAMES I H. Engineering mechanics: dynamics[M]. New Jersey,
ZHEN Shengchao, born in 1988, is currently a PhD candidate at
NJ: Englewood Cliffs, 1960.
[14] SALETAN E J, CROMER A H. Theoretical mechanics[M]. New Hefei University of Technology, China and a visiting scholar at
York, NY: Wiley Press, 1971. Georgia Institute of Technology, USA. His research interests
[15] ROSENBERG R M. Analytic Mechanics of discrete systems[M]. include analytic mechanics, dynamics of multi-body systems,
New York, NY: Plenum Press, 1977. optimal control, robust control, adaptive control, fuzzy
[16] IRSCHIK H, HOLL H J. The equations of Lagrange written for a
engineering, uncertainty management.
non-material volume[J]. Acta Mechanica, 2002, 153(3–4): 231–248.
[17] MOORE E H. On the reciprocal of the general algebraic matrix[J]. Tel: 1-404-9531295; E-mail: zhenshengchao@gmail.com
Bulletin of American Mathematical Society, 1920, 26(5): 394–395.
[18] PENROSE R. A generalized inverse for matrices[J]. Proceedings Chen Ye-Hwa, got his PhD degree at the University of California,
of the Cambridge Philosophical Society, 1955, 51(3): 406–413. Berkely in 1985, is currently a professor at Georgia Institute of
[19] WONG C W, YASUI K. Falling chains[J]. American Association of
Technology, USA. He is a member of IEEE, ASME and Sigma Xi.
Physics Teachers, 2006, 74(7): 490.
[20] CALKIN M G. The dynamics of a falling chain:II[J]. American He is also a regional editor(North America) of Nonlinear
Journal of Physics, 1989, 57(2): 157–159. Dynamics and System Theory, and an associate editor of
[21] CALKIN M G, MARCH R H. The dynamics of a falling chain: I[J]. International Journal of Intelligent Automation and Soft
American Journal of Physics, 1989, 57(2): 154–157. Computing. His research interests include advanced control
[22] SCHAGERL M, STEINDL A, STEINER W, et al. On the paradox
methods for mechanical manipulators, neural networks and fuzzy
of the free falling folded chain[J]. Acta Mechanica, 1997, 125(1-4):
155–168. engineering, adaptive robust control of uncertain systems,
[23] SOUSA C A, RODRIGUES V H. Mass redistribution in variable uncertainty management.
mass systems[J]. European Journal Physics, 2004, 25(1): 41–49. Tel: 1-404-894-3210; E-mail: yehwa.chen@me.gatech.edu

You might also like