Multibody System Dynamics 8: 141159, 2002.

2002 Kluwer Academic Publishers. Printed in the Netherlands.

141
Augmented Lagrangian Formulation: Geometrical
Interpretation and Application to Systems with
Singularities and Redundancy
WOJCIECH BLAJER
Institute of Applied Mechanics, Technical University of Radom, ul. Krasickiego 54, PL-26-600
Radom, Poland; E-mail: wblajer@poczta.onet.pl
(Received: 28 November 2000; accepted in revised form: 2 October 2001)
Abstract. A geometric interpretation of the augmented Lagrangian formulation of Bayo et al. (Com-
put. Methods Appl. Mech. Engrg. 71, 1988, 183195), applied to equations of motion in relative and
Cartesian coordinates, is presented. Instead of imposing constraints on a system in the traditional
sense, large articial masses resisting in the constrained directions are added, and the system motion
is enforced to evolve primarily in the directions with smaller masses (in the unconstrained directions).
Then, the residual motion in the constrained directions is removed by applying the constraint reac-
tions to the system, estimated effectively in few iterations. The formulation is comparatively simple
and leads to computationally efcient numerical codes. Useful applications of the formulation to the
dynamic analysis of constrained multibody systems with possible singular congurations, massless
links and redundant constraints are shown. The theoretical background is followed by some remarks
on the modeling precautions and assisted computational peculiarities of the method. The results of
numerical simulation of motion of a parallel ve-bar and a parallel four-bar linkage are reported.
Key words: constrained multibody systems, redundant constraints, singularities.
1. Introduction
Multibody systems are often modeled by using a non-minimal set of, say n, de-
pendent coordinates q, interrelated through l kinematic constraints. Assumed the
constraints are holonomic and scleronomic, the constraint equations are
(q) = 0. (1)
Then, after differentiating with respect to time the original constraint equations (1)
at position level, the constraint conditions at velocity and acceleration levels follow
consecutively as:

C(q) q = 0, (2)

C(q) q (q, q) = 0, (3)

142 W. BLAJER
where C = /q is the l n constraint matrix, and =

C q is the l-vector of
constraint induced accelerations. The dynamic equations of the system arise then
in the form of Lagranges equations of the rst type
M(q) q +d(q, q) = f(q, q, t ) C
T
(q), (4)
where M is the generalized mass matrix, the n-vector d represents the centrifugal,
Coriolis and gyroscopic dynamic terms, f is the n-vector of generalized applied
forces, are l Lagrange multipliers associated with the constraints (1), and t is the
time. Handled together, (4) and (3) form n + l Differential-Algebraic Equations
(DAEs) in q and , which can be solved directly or, prior to the numerical integ-
ration, transformed to a smaller set of Ordinary Differential Equations (ODEs). A
variety of the DAE/ODE multibody formulations have so far been developed and
widespread in the literature, see e.g. [15].
There are two basic requirements that condition applicability of many of the
classical multibody codes [1] in the whole conguration range of a system:
M must be nonsingular (it is usually assumed to be positive denite), which
excludes the modeling of some of the system links as massless,
C must be of maximal row-rank, rank(C) = l, which yields
no redundant constraints on the system,
no singular positions encountered.
Special procedures must also be followed to model/simulate systems in the pres-
ence of constraint addition/deletion and/or changing topologies, which may be
especially cumbersome in the case of ODE formulations.
A variant method for constrained multibody systems, called augmented Lag-
rangian formulation, was proposed in [6] and then succeeded e.g. in [79]. Accord-
ing to the geometrical interpretation provided in the following, instead of imposing
constraints on a system in the conventional sense, large articial masses resisting
in the constrained directions are added, which enforce the system motion to evolve
primarily in the directions with smaller masses (unconstrained directions). Then,
the residual motion in the constrained directions is removed by applying the con-
straint reactions to the system, which are estimated effectively in few iterations.
The formulation is comparatively simple and leads to computationally efcient
numerical codes. As compared to [6-9], the main contributions and amendments of
this paper to this formulation can be itemized as follows:
an insightful geometrical/physical interpretation of the method is given, which
expounds the formulation as an augmented mass method;
a simple derivation of the formulation is demonstrated;
AUGMENTED LAGRANGIAN FORMULATION 143
Figure 1. The physical interpretation of Arnolds/penalty formulation.
an application of the formulation to equations of motion in any set of com-
monly used dependent coordinates q is shown (the original applications the
method are restricted to the so-called natural coordinates [2]);
the algorithm for efcient determination of q(t ), q(t ), and (T ) is recon-
sidered;
the method is shown applicable to systems with redundant constraints, chan-
ging topologies, and, with precautions, singular positions and/or massless
members;
some peculiarities concerned with the modeling and computational treatment
of the singularities and redundancy are emphasized;
some remarks on the computational applications of the method, its effective-
ness and accuracy, are incorporated.
As said, the method presented in this paper is not genuinely novel it coincides
with the augmented Lagrangian formulation [2, 69]. It is rather a more physically
insightful reformulation/interpretation of the previous approach, and an extension
of its application range. In fact, the original formulation is closely related to the
natural coordinates [2], ingenious but specic coordinates to describe a system
conguration. The present contribution is then an application of the approach to
systems described in the more commonly used coordinates, e.g. absolute and/or
joint coordinates.
2. The Physical Background
Instead of thinking about constraints in terms of hard surfaces, rigid links, slipless
rolling contacts, etc., a legitimate and old treatment, motivated then by Arnold [10],
is to replace the stiff constraints by a strong force eld in a neighborhood of the
manifold (q) = 0, directed towards the manifold. The effect can be interpreted
as the action of l elastic forces pointed in the constrained directions towards the
respective constraints and proportional to the constraint violations (Figure 1). As-
sumed the same stiffness coefcient k in the all constrained directions, the elastic
144 W. BLAJER
forces are k, and their projection onto the directions of q is C
T
k [4]. The
dynamic equation (4) can then be modied to
M q = f d C
T
k, (5)
which can also be seen as the scantiest penalty formulation [2, 6]. For the penalty
factor k large enough, the moving system will always be close to the constraint
manifold (q) = 0, and in the limit case of an innite force eld (k ), the
system must remain on the manifold. The latter case is equivalent to constraints (1)
put on the system [10].
While approved theoretically, the innite (very large) increase of k always leads
to instability in computations. More strictly, for a given integration time step t ,
moderate values of k allow usually an unacceptable constraint violation, while
increasing k leads rst to the oscillating and then unstable numerical solutions.
The situation can be improved by adding some damping and inertia along the
constrained directions the forces proportional to the violations of constraints
at the velocity and acceleration levels. The arising penalty forces that resist the
constraint violations are then

+ c

+ k, where stands for the value of

the constraining masses and c is the damping coefcient. The effect can also be
rewritten in a more familiar form (

+ 2

+
2
), where 2 = c/ and

2
= k/, and the condition of critical or over-critical damping is advisable.
It may be worth noting that the penalty factors, in this formulation taken identical
for each constraint, can possibly be different, e.g. , c and k can be replaced by
appropriate diagonal matrices of separate factors. Moreover, depending on the j-th
constraint
j
(j = 1, . . . , l) denotes the restricted relative translation or rotation
in a joint, the units of the penalty factors , c and k should respectively be: [kg]
or [kg m
2
], [kg s
1
] or [kg m
2
s
1
], and [kg s
2
] or [kg m
2
s
2
]. The latter aspects
can however be concealed in computations, and unied penalty factors are usually
chosen for simplicity. In the following we will therefore refer only to the values of
, c and k, granting that the units of

, c

and k are appropriately adjusted.

After applying Equation (3), the nal (full) penalty formulation is (see also [2,
6])
(M+C
T
C) q = f d +C
T
( 2

2
). (6)
Irrespective of some advantageous peculiarities of the leading matrix M+ C
T
C
(which will be discussed in more detail in the following sections), the formulation
still shows a tendency to instability in the numerical solutions for very large pen-
alty factors , c and k. More specically, referring to the modied penalty form
of constraint conditions, (

+ 2

+
2
), for given values of and such
that , setting (very large), and keeping the integration time step t
constant, always leads to computational instability. The value of must then be
limited to a certain (case dependent) value, usually ranging from 10
6
to 10
7
(see
also [69]). On the other hand, the penalty formulation (6) is very close related to
Baumgartes constraint violation stabilization method [11], with all its advantages
AUGMENTED LAGRANGIAN FORMULATION 145
Figure 2. The physical interpretation of the method.
and disadvantages. The solution to (6) will remain close but not exactly on the
constraint manifold (q) = 0.
The augmented Lagrangian formulation exploits the advantageous peculiarit-
ies of the leading matrix M+ C
T
C (shown in the following) and removes the
drawbacks of limited accuracy and/or instability of the penalty formulation.
3. The Present Method
For a mechanical system as described in Section 1, the physical interpretation of
the present formulation can be summarized in the following three steps:
1. Start from the unconstrained system whose dynamic equations are M q+d = f.
2. Add to the system l articial masses of value , which resist only in the
constrained directions with respect to (1). Assuming that are large enough
compared to the actual masses of the system represented in M (in Figure 2
denoted symbolically by m), the motion of the system (of the mass m) will
evolve primarily in the unconstrained directions, while the whole modied
system (both the masses and m) will tend to drift slightly in the constrained
directions as well.
3. Stop the residual motion in the constrained directions by imposing appropriate
forces

in these directions (applied to ); see Figure 2. Once the constraint

drift tendency is removed, the applied additional forces are equivalent to the
constraint reactions in the classical sense.
Due to the above physical interpretation, the code can be referred to as an aug-
mented mass method. Resembling mass-orthogonal projections of the augmented
Lagrangian formulation were previously observed in [9].
In order to introduce the mathematical formulation of the method, let us rst
reconsider the constraint equations (3) at the acceleration level using the geomet-
rical interpretation provided in [4]. The equality C q = means that the projections
of the system acceleration on the constrained directions, C q, are equal to the con-
146 W. BLAJER
straint induced accelerations in these directions. By assumption, the motion of
the constraining masses must be consistent with this constraint condition. The
dynamic generalized forces (in the constrained directions) related to the masses
are then C q = , and the projection of the dynamic forces into the directions of
q is nally C
T
C q = C
T
. The latter equality can be seen as the mathematical
representation of the imposition of the constraining masses on the unconstrained
system, as mentioned in step 2 above. The representation of step 3 can simply be
written as C
T

added to right-hand side of the arisen motion equations, where

are the l forces added in the constrained directions to resist the constraint drift
tendency. In sum, starting from the unconstrained dynamics (step 1) and following
steps 2 and 3, we obtain
M q = f d
C
T
C q = C
T

= C
T

(M+C
T
C) q = f d +C
T
C
T

(7)
The result (7) can also be obtained from the penalty formulation (6) after removing
the elastic and damping constraining forces, and adding the additional forces

of step 3, where

= if the system acceleration satises the constraint condition

C q = . Also, the result is equivalent to the augmented Lagrangian formulation
with projections [8, 9].
4. The Numerical Code
Assumed q
0
and q
0
satisfy the constraint conditions (1) and (2), for the
solution to (7) will tend to be consistent with the constraints even though

=
0 the motion in the constrained directions will be excluded due to the innite
constraining masses. In computations must be limited to a certain value, however,
and the residual motion in the constrained directions can be removed by applying
some additional forces

. Since

are not known a priori, the determination of the

constraint-consistent accelerations q and adequate values of

must be performed
iteratively. Denoting in Equation (7) H(q) = M+C
T
C and h(q, q, t ) = f d +
C
T
, at a given simulation time t , the iteration process is (setting

0
= 0).
q
i+1
= H
1
(q)(h(q, q, t ) C
T
(q)

i
);

i+1
=

i
+

i+1
; i = 0, 1, 2, . . . , (8)
where

i+1
= C(q) q
i+1
(q, q). For ranging from 10
5
to 10
7
usually only
13 iterations sufce to achieve | q
i+1
q
i
| , where is the required numerical
accuracy.
The numerical code can be explained as follows. For the current state variables
q and q that satisfy the constraints (1) and (2), initially (i = 0) we set

0
= 0 to
AUGMENTED LAGRANGIAN FORMULATION 147
Figure 3. The physical interpretation of the numerical code.
obtain q
1
= H
1
h in the rst (i = 1) step. Evidently q
1
may not be consistent with
the constraint condition (3), and the inconsistency measure is

1
= C q
1
, which
denotes the difference between the actual and the constraint induced accelerations
in the constrained directions. To resist the constraint drift tendency the forces

1
=

1
are applied on the system, equal to the inertial forces of the masses due to
the drift accelerations

1
. The acceleration estimated in the next (i = 2) iteration,
q
2
= H
1
(h C
T

1
), is then much more consistent with the constraint condition
(3). The consistency is not exact however since both the masses and m (Figure 3)
are affected by the drift accelerations, while only the inertial forces of the masses
are resisted by

1
=

1
. The consistency is then improved by correcting the
reaction force to

2
=

1
+

2
, where

2
= C q
2
, which follows in still
more consistent accelerations q
3
= H
1
(h C
T

2
) in the next (i = 3) step. Then

3
=

2
+

3
, . . . , etc. Note that for much larger than the actual masses of
the system (for m), the iteration process converges very quickly. Note also
that all the used quantities H
1
, h, C and remain constant in the iteration process
(are determined only once for the current state of the system). To illustrate how the
method works, let us investigate the following simple example.
4.1. ILLUSTRATIVE EXAMPLE
A particle of mass m moves on a circle of radius . The constraint equation and its
rst and second time-derivatives are:
=
_
x
2
+y
2
= 0;

=
x x

+
y y

= 0;

=
x x

+
y y

+
x
2
+ y
2

= 0,
which yields the constraint matrix C = [x/ y/] and the constraint induced
acceleration = ( x
2
+ y
2
)/. The arising formulation (7) of the method is then
148 W. BLAJER
Figure 4. The illustrative example.
_
_
m+
x
2

2

xy

xy

2
m+
y
2

2
_
_
_
x
y
_
=
_
_
F
x

x( x
2
+ y
2
)

2
F
y

y( x
2
+ y
2
)

2
_
_

_
x

,
where F
x
and F
y
are the components of the external force on the particle.
For simplicity, let us assume that the current particle position is on the axis Ox
(Figure 4). The above formulation of the method simplies then to
_
m+ 0
0 m
_ _
x
y
_
=
_
F
x

y
2

F
y
_

_
1
0
_

.
The iteration process at this position is demonstrated in Table I. As seen, assumed
m, the acceleration x
i
of mass m and the force

i
that resists the constraint
drift, both in the constrained direction, converge very quickly respectively to the
constraint consistent acceleration y
2
/ and the constraint reaction F
x
+m y
2
/.
5. The Range of Applications and Computational Aspects
The dependent coordinates q of the present formulation (7) can be any set of the
coordinates that are commonly used to describe a multibody system congura-
tion, namely the absolute (Cartesian) and joint (relative) coordinates [2, 3, 12],
or mixed. For a slider-crank mechanism shown in Figures 5a, 5b, and 5c, these
coordinates are:
q
a
= [x
1
y
1

1
x
2
y
2

2
x
3
y
3
]
T
; q
b
= [
1

2
s]
T
;
q
c
= [ x
2
y
2

2
s]
T
.
In all the three cases, the constraints = [
1
. . .
l
]
T
on the respective systems
can be modeled so that to express the prohibited relative translations/rotations in
the cut joints. The corresponding constraint reactions = [
1
. . .
l
]
T
will then
be the physical forces/moments in the joints,
i
counteracting
i
, which is stated
AUGMENTED LAGRANGIAN FORMULATION 149
Table I. The illustration of iteration process.
i = 0

0
= 0
i = 1
x
1
=
F
x
m+

m+
y
2

y
1
=
F
y
m

1
=

m+
_
F
x
+
m y
2

_
i = 2
x
2
=
mF
x
(m+)
2

(m+)
2
m
2
(m+)
2
y
2

y
2
=
F
y
m

2
=
(m+)
2
m
2
(m+)
2
_
F
x
+
m y
2

_
i = 3
x
3
=
m
2
F
x
(m+)
3

(m+)
3
m
3
(m+)
3
y
2

y
3
=
F
y
m

3
=
(m+)
3
m
3
(m+)
3
_
F
x
+
m y
2

_
(. . .) (. . .) (. . .)
i
( m)
x

=
y
2

=
F
y
m

= F
x
+
m y
2

Figure 5. A slider-crank mechanism modeled with (a) absolute, (b) joint, (c) mixed
(joint-absolute), and (d) natural coordinates.
150 W. BLAJER
in the classical formulation (4). Though the drift resisting forces

in the present
formulation (7) have notionally different meaning from , after some N iterations
described in Equation (9),

N
are equal to with a required numerical accuracy.
More strictly,

N
are applied to the augmented masses . However, since the drift
of the augmented system (both and m masses) in the constrained directions is
removed, i.e. the acceleration of the actual system (mass m) is consistent with the
constraint condition (3), the constraint reactions are the reactions between the
actual system (mass m) and the augmented masses (mass ), and therefore =

N
.
The determination of constraint reactions as physical forces/moments in the
joints is an important advantage of such setting the constraint equations (1) to ex-
press the prohibited relative translations/rotations in the cut joints, and this relates
the present application of the formulation (7) to equations of motion in abso-
lute/joint coordinates. The situation is different when the so-called natural co-
ordinates [2] are used. For the slider-crank mechanism seen in Figure 5d these
are
q
d
= [x
1
y
1
x
2
y
2
]
T
,
which are the coordinates of points 1 and 2, to which, followed a legitimated pro-
cedure [2], the kinematical and inertial properties of the system are transferred.
While the arising constraint equations
_
(x
1
x
A
)
2
+(y
1
y
A
)
2
l
1
= 0;
_
(x
2
x
1
)
2
+(y
2
y
1
)
2
l
2
= 0; y
2
= 0
have evident physical meaning, the associated constraint reactions

N
, due to the
articiality of the mass points 1 and 2, do not express any physical forces in the
system joints. Therefore, though the natural coordinates have many advantageous
peculiarities, the loss of the physical insight may be discouraging. Moreover, the
augmented Lagrangian formulation is usually introduced together with these spe-
cic coordinates [2] (the examples are so restricted), and this is why this interesting
and useful method of dynamics seems to be hidden behind the natural coordinate
formulation. The application of the approach to systems described in the more
commonly used absolute/joint coordinates is shown in the following.
Irrespectively of the coordinate type used, the intrinsic advantages of formu-
lation (8) are due to the inherent features of the leading matrix M+ C
T
C. The
matrix remains invertible even in the presence of
changing topologies (varying number of constraints),
redundant constraints,
singular positions encountered,
possible singularity of M.
For any (possibly varying) number of constraints on the system, including the case
m n (redundant constraints), the dimension of C
T
C is n n. Of paramount
AUGMENTED LAGRANGIAN FORMULATION 151
importance is that the diagonal elements of this matrix are either positive or equal
to zero (they are never negative). This holds true also in those cases when the
maximal row-rank of the constraint matrix C is lost, either permanently (redundant
constraints) or occasionally (at singular positions). As a consequence, the matrix
M+C
T
C remains positive denite even in the presence of changing topologies,
redundant constraints and singular positions. Moreover, some singularity of the
mass matrix M is admissible as well. More strictly, selected links of a multibody
system modeled with the present method can be set massless. The correspond-
ing null diagonal entries of M must then be completed with the positive denite
diagonal entries of C
T
C. This excludes the last links of open chains from being
modeled as massless. Moreover, the internal massless links must rst be entirely cut
off from the system, and then modeled in absolute coordinates. This excludes the
modeling of mass singular systems in joint coordinates, and absolute coordinates
of (at least) massless links must be used. As an illustration, massless can be any
link in Figure 5a (absolute coordinate formulation), and the link 2 in Figure 5c
(mixed joint-absolute formulation).
An important virtue of the present formulation applied to systems with redund-
ant constraints is that the iterated values

N
involve the correct values of reactions
of all the constraints on the system, including the redundant constraints. Many
of the existing multibody codes are unable neither to analyze the systems with
redundant constraints nor to determine the redundant constraint reactions.
Singular positions are encountered when the motion cannot continue beyond
some particular congurations of the system (the lockup congurations) or when
the system reaches a conguration in which there is a sudden change in the number
of degrees of freedom [3]. In the following we will limit ourselves only to the latter
singularities. Depending on the nature of the driving input, the system motion in
such singular positions may proceed non-smoothly and/or more than one possible
motion can occur. For the slider-crank mechanism shown in Figure 5, and designed
so that l
1
= l
2
and b = 0, a singular position is reached when the two links are in
vertical position.
A singular position is mathematically detected by an instantaneous fall of rank
of the constraint matrix C. The simulation may then crash not because of the
physics of the problem but because of the inability of the dynamics formulations to
overcome the rank change. This is not the case of the present formulation, however.
The leading matrix M + C
T
C in Equation (7) can be invertible irrespective of
the rank of C. The mathematical impact of the singularities is thus removed. The
physical reasons of the singularity may still remain, however. Getting through the
singular congurations smoothly is then conditioned upon the additionally arising
degrees of freedom are activated or not. If it is so, the iteration procedure (8) for
the constraint consistent accelerations may not converge, and special algorithms
have to be used. Moreover, the closed loops of the singular position mechanisms
analyzed with the present method should not be cut in the joints but must be disjoint
152 W. BLAJER
by rending off the whole members, modeled then as free bodies. These problems
will be illustrated in the next section.
It may be worth comparing the present formulation (7) with the standard DAE
formulation [1, 2, 4, 12] of equations of motion,
_
M C
T
C 0
_ _
q

_
=
_
f d

_
, (9)
where both the n dynamic equations (4) and k constraint equations (3) are explicitly
involved. Both the formulations are simple to obtain and then implement in digital
computations. In the case of DAEs (9), the n +l leading matrix has to be inverted
(or an equivalent numerical code has to be applied) to solve the DAEs for the
exact accelerations and constraint forces at the consistent position and velocity
of the system. By contrast, in formulation (7) only n differential equations are
solved, while the constraint equations (3) are involved in the iteration process (8)
for the determination of the consistent accelerations and reaction forces. In fact,
the same number of unknowns and equations is thus used in both formulations, the
numerical efciency of both the codes seem to be similar. The solvability of DAEs
(9) is however dependent on det(CM
1
C
T
) = 0, or more strictly: det(M) = 0 and
rank(C) = l = max, which excludes the direct analysis of systems with massless
links and redundant constraints, and difculties may arise at the singular positions.
The iteration process (8) usually converges very quickly (in 2 or 3 steps), and
H
1
(q), h(q, q, t ), C(q) and (q, q) used in the process are determined only once
for the current state variables. In principle, the bigger the better convergence
of the code (8). Too large s usually cause numerical instability, however, which
limits to a certain case dependent maximal value. The source of the instability
are the random errors due to the cancellation of leading digits in subtraction of
numbers of the same order of magnitude. The computations of terms like [1
/(m + )] are involved, and increasing above a certain level may not only
enlarge the random errors but lead to incorrectness in computations as well. The
value of [1/(m+)], which should always be smaller then one, for very large
and a moderate m, may be computed as larger than one. These computational
phenomena lead to instability of the present code. The authors experience is that
for many typical applications and calculations performed in FORTRAN (double
precision), the reasonable values of range from 10
5
to 10
7
(see also [2, 69]).
As in the case of the DAE formulation (9), the solution to (7) may suffer from
the constraint violation problem. The constraint violations = (q) = 0 and

=

(q,

q) = 0 and for the current state values q and

q obtained from the
numerical integration can be eliminated by correcting the state variables according
to the following schemes [8, 9]:
q = H
1
C
T
; q = H
1
C
T

, (10)
where q = q q and q = q

q, and q and q are the corrected state variables.
This is also an iterative process, converging usually in 2 or 3 steps to achieve
AUGMENTED LAGRANGIAN FORMULATION 153
Figure 6. The parallel ve-bar linkage with one redundant constraint.
= 0 and

= 0 with a numerical accuracy. Using the correction scheme
after each integration step (or a sequence of steps), constant H
1
and C, used
previously in the integration, can be applied. The variant correction schemes are
q = M
1
C
T
(CM
1
C
T
)
1
and q = M
1
C
T
(CM
1
C
T
)
1
, which are
motivated in [4, 13]. These, however, depend on det(CM
1
C
T
) = 0, and as such
do not apply to systems with massless links, redundant constraints and singular
positions.
6. The Case Studies
Consider a planar parallel ve-bar linkage (Figure 6a) with one redundant con-
straint (the kinematics of the mechanism is unaffected by the removal of one of the
links 2 or 4). Opening the closed loops by rending off the links 2 and 4, the eight
mixed Lagrangian and Cartesian coordinates are q = [
1
x
2
y
2

2

3
x
4
y
4

4
]
T
.
The eight constraint equations, the 8 8 constraint matrix C, and the 8-vector of
the constraint accelerations are:

1
= x
2
0.5l cos
2
l cos
1
= 0,

2
= y
2
0.5l sin
2
l sin
1
= 0,

3
= x
2
+0.5l cos
2
l cos
3
= 0,

4
= y
2
+0.5l sin
2
l sin
3
= 0,

5
= x
4
0.5l cos
4
0.5l cos
1
= 0,

6
= y
4
0.5l cos
4
0.5l sin
1
= 0,

7
= x
4
+0.5l cos
4
0.5l cos
3
= 0,
154 W. BLAJER

8
= y
4
+0.5l sin
4
0.5l sin
3
= 0,
=
_
_
_
_
_
_
_
_
_
_
_
_
_
_
0.5l

2
2
cos
2
l
2
1
cos
1
0.5l

2
2
sin
2
l
2
1
sin
1
0.5l

2
2
cos
2
l
2
3
cos
3
0.5l

2
2
sin
2
l
2
3
sin
3
0.5l

2
4
cos
4
0.5l
2
1
cos
1
0.5l

2
4
sin
4
0.5l
2
1
sin
1
0.5l

2
4
cos
4
0.5l
2
3
cos
3
0.5l

2
4
cos
4
0.5l
2
3
cos
3
_

_
,
C =
_
_
_
_
_
_
_
_
_
_
_
_
_
_
l sin
1
1 0 0.5l sin
2
0 0 0 0
l cos
1
0 1 0.5l cos
2
0 0 0 0
0 1 0 0.5l sin
2
l sin
3
0 0 0
0 0 1 0.5l cos
2
l cos
3
0 0 0
0.5l sin
1
0 0 0 0 1 0 0.5l sin
4
0.5l cos
1
0 0 0 0 0 1 0.5l cos
4
0 0 0 0 0.5l sin
3
1 0 0.5l sin
4
0 0 0 0 0.5l cos
3
0 1 0.5l cos
4
_

_
.
The maintenance of the constraints yields then
2
=
4
= 0 and
1
=
2
. For

1
=
3
= n (n = 0, 1, . . .) the rank of the matrix C is equal to seven (one
redundant constraint), while for the singular positions
1
=
3
= n the rank
of C decreases to six the number of degrees of freedom of the system switches
from one to two. Assuming the links are thin and homogeneous bars, and that the
applied forces on the system are only the gravity forces, the dynamics of the system
is dened by:
M = diag
_
m
1
l
2
3
, m
2
, m
2
,
m
2
l
2
12
,
m
3
l
2
3
, m
4
, m
4
,
m
4
l
2
12
_
; d = 0,
f =
_

m
1
gl cos
1
2
0 m
2
g 0
m
3
gl cos
4
2
0 m
4
g 0
_
T
.
Using the data m
1
= m
2
= m
4
= 1 kg, l = 1 m, = 10
6
kg (as all
the constraints express the prohibited relative translations in the joints, the unit
of can be stated in kg), and the initial state values dened by
10
= 0 deg
(singular position) and
10
= 2 s
1
, the obtained results of numerical simula-
tion are shown in Figure 7 (solid lines). The dashed lines in the graphs illustrate
the results obtained for the case m
2
= 0 kg. A standard ODE solver based on
RungeKutta fourth-order algorithm was used, and the integration time step was
t = 0.02 s. As seen, the simulated motion goes smoothly through the singular
AUGMENTED LAGRANGIAN FORMULATION 155
Figure 7. The parallel ve-bar linkage time-variations of the state variables and constraint
reactions.
positions, and the singularity of the mass matrix M does not affects the solution
attainability. In calculations, throughout the whole simulation time, at most three
iterations according to the scheme (9) were required to achieve numerical accuracy
| q
i+1
q
i
| 10
15
m s
2
, and at most two iterations according to the algorithm
(10), used after each integration step, were required to keep the constraint viola-
tions || 10
15
m and |

| 10
15
m s
1
. It may also be worth noting that
the reactions of all constraints are determined, including the redundant constraints.
These are obtained for the idealized case, however, when all the link lengths are
exactly the same. In reality, even small variations in the link lengths may cause
large deviations in the joint reactions. On the other hand, assumed the link length
variations are negligible compared to the joint allowances, and the impact and
stick-slip phenomena can be omitted due to the viscous lubrication, the present
approach provides one with a method for at least estimation of the reaction forces
in an over-constrained rigid-body model.
The smooth motion through the singular positions is possible since the addi-
tional degree of freedom occurring in these positions (the rank of C decreases from
seven to six) is not activated. More strictly, the applied gravitational forces do not
cause any difference between the accelerations
1
and
3
at the singular positions,
156 W. BLAJER
Figure 8. The parallel four-bar linkage.
even though the links 1 and 3 are not for a moment constrained from accelerating
independently. The physical singularity is thus not revealed, and the mathematical
singularity (the rank loss of C) is removed by the present formulation. The situation
is different when applying a driving torque to the link 1 that is discussed in the
sequel.
Let us focus on the problem of the physical singularity referring to a simplied
version of the hitherto mechanism, the parallel four-bar linkage shown in Figure 8a.
The torque causes interim
1
=
3
when going through the singular positions

1
=
3
= n (n = 0, 1, . . .), and this yields
1
=
3
and
1
=
3
at the next
instant of time (Figure 8c). The state of the system just after the singular positions
will thus be in conict with the constraints on the system (at the singular positions
the constraints vanish and allow
1
=
3
, and then immediately retrieve and restrict
from
1
=
3
). As a consequence, when encountered a singular position in numer-
ical simulation (or for a position very close to the singular position), the iteration
process described in Equation (9) may not converge (or may converge very slowly).
The lack of converge was also observed in [9], provided an explanation in terms of
time step.
To proceed with the numerical simulation through the singular positions when
the arising additional degrees of freedom are activated, the following procedure
is proposed. When encountered the physical singularity, which is detected by di-
vergence of the iteration process (8) (or a large number of required iterations), the
process is stopped after some reasonable number of iterations (for example, ve),
and Equations (7) are integrated to pass the singular position using the inconsist-
AUGMENTED LAGRANGIAN FORMULATION 157
Figure 9. The parallel four-bar linkage time-variations of the state variables and constraint
reactions.
ent accelerations. The arising state of the system (just after the singular position)
may then be in conict with the constraint conditions (2) and (3) the integrated
position and velocity of the system may cause constraint violations = 0 and

= 0. The constraint violations are then removed by using the state correction
scheme (10) since the invertibility of H is unaffected by the singular positions (the
rank loss of C). The effect of the algorithm is seen in Figure 9, and the results
of numerical simulation were obtained for m
1
= m
2
= m
3
= 1 kg, l = 1 m,
= 10
6
kg,
10
= 90 deg,
10
= 0 s
1
, and = 1.5 Nm. The motion goes
smoothly through the singular positions, while the abrupt changes in the values of

1
and
2
at the singular positions are observed, which are due to the numerical
effects (the lack of converge).
Let us note nally that a bifurcation of the system motion in the singular pos-
ition may happen. The bifurcation is most likely if the system reaches a singular
position with vanishing velocity (which is not the case of the present illustration). A
more thorough study on the problem should involve an elastic-body model and/or
release the constraints by introducing elasticities.
7. Concluding Remarks
A geometrical interpretation of the augmented Lagrangian formulation is given,
followed by a physically grounded derivation of the method. The application of
this formulation to equations of motion in commonly used absolute and/or joint
coordinates is then shown, by contrast to the original examples of implementations
in the natural coordinates [2]. Unlike many other multibody codes, the formulation
is capable of treating systems with changing topologies (not shown in this paper;
158 W. BLAJER
refer to [2, 69] for details), redundant constraints and singular positions. Some
of the bodies of the system can also be modeled massless. The physical constraint
reactions, including the reactions of redundant constraints, can also be estimated.
The formulation can be applied without any restrictions to systems modeled
in absolute coordinates. Either open-loop or closed-loop systems, with possible
redundant constraints and/or singular congurations can be treated this way. When
modeling a closed-loop system in a smaller set of coordinates, it is recommen-
ded to open the loops by rending off the whole bodies instead of cutting off the
joints, which leads to a mixed join/absolute coordinate formulations of constrained
systems (illustrated in Figure 5c). This is especially important for systems with
redundant constraints and/or singular positions. The bodies modeled massless must
also be rst rent off from the other links, and modeled as free bodies in absolute
coordinates. These cannot be the last links of the open chains, however. There are
also some numerical precautions in proceeding the motion through the singular
positions.
Acknowledgments
The bulk part of the work was prepared during the authors stay in the Institute B of
Mechanics, University of Stuttgart, Germany, on behalf of the Alexander von Hum-
boldt Fellowship. The research has also been supported by the State Committee for
Scientic Research, Poland, under grant 9 T12C 060 17.
References
1. Schiehlen, W. (ed.), Multibody Systems Handbook, Springer-Verlag, Berlin, 1990.
2. Garca de Jaln, J. and Bayo, E., Kinematic and Dynamic Simulation of Multibody Systems:
The Real Time Challenge, Springer-Verlag, New York, 1994.
3. Shabana, A.A., Computational Dynamics, John Wiley & Sons, New York, 1994.
4. Blajer, W., A geometric unication of constrained system dynamics, Multibody System
Dynamics 1, 1997, 321.
5. Schiehlen, W., Multibody system dynamics: Roots and perspectives, Multibody System
Dynamics 1, 1997, 149188.
6. Bayo, E., Garca de Jaln, J. and Serna, M.A., A modied Lagrangian formulation for the dy-
namic analysis of constrained mechanical systems, Computer Methods in Applied Mechanics
and Engineering 71, 1988, 183195.
7. Bayo, E. and Avello, A., Singularity free augmented Lagrangian algorithms for constrained
multibody dynamics, Nonlinear Dynamics 5, 1994, 247255.
8. Bayo, E. and Ledesma, R., Augmented Lagrangian and mass-orthogonal projection methods
for constrained multibody dynamics, Nonlinear Dynamics 9, 1996, 113130.
9. Cuadrado, J., Cardenal, J. and Bayo, E., Modeling and solution methods for efcient real-time
simulation of multibody dynamics, Multibody System Dynamics 1, 1997, 259280.
10. Arnold, V.I., Mathematical Methods of Classical Mechanics, 2nd edn., Springer-Verlag, New
York, 1989.
11. Baumgarte, J., Stabilization of constraints and integrals of motion in dynamical systems,
Computer Methods in Applied Mechanics and Engineering 1, 1972, 116.
AUGMENTED LAGRANGIAN FORMULATION 159
12. Nikravesh, P.E., Computer-Aided Analysis of Mechanical Systems, Prentice-Hall, Englewood
Cliffs, NJ, 1988.
13. Blajer, W., State and energy stabilization in multibody system simulations, Mechanics
Research Communications 26, 1999, 261268.