Stable Synchronization of

Mechanical System Networks

Sujit Nair and Naomi Ehrich Leonard∗
Mechanical and Aerospace Engineering
Princeton University
Princeton, NJ 08544 USA
nair@alumni.princeton.edu, naomi@princeton.edu

September 12, 2007

Abstract
In this paper we address stabilization of a network of underactuated mechanical systems with
unstable dynamics. The coordinating control law stabilizes the unstable dynamics with a term
derived from the Method of Controlled Lagrangians and synchronizes the dynamics across the
network with potential shaping designed to couple the mechanical systems. The coupled system
is Lagrangian with symmetry, and energy methods are used to prove stability and coordinated
behavior. Two cases of asymptotic stabilization are discussed, one that yields convergence
to synchronized motion staying on a constant momentum surface and the other that yields
convergence to a relative equilibrium. We illustrate the results in the case of synchronization of
n carts, each balancing an inverted pendulum.

1 Introduction
Coordinated motion and cooperative control have become important topics of late because of grow-
ing interest in the possibility of faster data processing and more efficient decision-making by a net-
work of autonomous systems. For example, mobile sensor networks are expected to provide better
data about a distributed environment if the sensors can be made to cooperate towards optimal
coverage and efficient coordination.
Much of the recent work explores coordination and cooperative control with very simple dy-
namical systems, e.g., single or double integrator models (e.g., [10, 16, 17]) or nonholonomic models
(e.g., [4]). For example, in some of these and related works, stabilization of coordinated group dy-
namics is studied in the case of limited, time-varying communication topologies. These authors
deliberately choose to focus on the coordination issues independent of issues in the stabilization of
individual dynamics.
However, for networks of autonomous systems such as unmanned helicopters or underwater
vehicles, stability of individual dynamics can be important and challenging, and it may not always
be possible (or desirable) to decouple the stabilization problem of individual dynamics from the
coordination problem. In [21] the authors consider stability of a group with dynamics that satisfy
a leader-to-formation stability (LFS) condition based on input-to-state stability [19]. Examples
∗
Research partially supported by the Office of Naval Research under grants N00014–02–1–0826 and N00014-04-1-
0534. A preliminary version of some parts of this paper appeared in [15].

1
include linear dynamical systems and kinematic nonholonomic robots; in the latter case feedback
linearization is used for stabilization. Using the LFS property, the authors are able to quantify how
leader inputs and disturbances affect group stability. In [6], an extension to a previous work ([5])
on unmanned aerial vehicle motion planning is presented for identical multiple-vehicle stabilization
and coordination. The single vehicle motion planning is based on the interconnection of a finite
number of suitably defined motion primitives. The problem is set in such a way that multiple-vehicle
motion coordination primitives are obtained from the single-vehicle primitives. The technique is
applied to motion planning for a group of small model helicopters.
Networks of rigid bodies are addressed in [8]. Reduction theory is applied in the case that
control inputs depend only on relative configuration (relative orientation or position). The reduction
results are used to study coordinated behavior of satellite and underwater vehicle network dynamics.
Stability of a network of rotating rigid satellites is proved in [14].
In this paper, we investigate the problem of coordination of a network of underactuated me-
chanical systems with unstable dynamics. As a first step we make use of the Method of Controlled
Lagrangians to stabilize the unstable dynamics of each mechanical system. The Method of Con-
trolled Lagrangians and the equivalent Interconnection and Damping Assignment Passivity-Based
Control (IDA-PBC) method use energy shaping for stabilization of underactuated mechanical sys-
tems (see [1, 18] and references therein). The Method of Controlled Lagrangians provides a control
law for underactuated mechanical systems such that the closed-loop dynamics derive from a La-
grangian. The approach is to choose the control law to shape the controlled kinetic and potential
energy for stability.
The class of underactuated mechanical systems we consider in this paper satisfy the simplified
matching conditions defined in [2, 1]. This class includes the planar or spherical inverted pendulum
on a (controlled) cart. The goal of the development in this paper is to stabilize unstable dynamics for
each individual mechanical system in the network and stably synchronize the actuated configuration
variables across the network. For example, for a network of pendulum/cart systems, the problem
is to stabilize each pendulum in the upright position while synchronizing the motion of the carts.
For stabilization of individual unstable dynamics we use the approach in [1]. To simulta-
neously synchronize the dynamics across the network, we show that potentials that couple the
individual systems can be prescribed so that the complete coupled system still satisfies the simpli-
fied matching conditions. Accordingly, we can choose potentials, find a Lagrangian for the coupled
system and prove Lyapunov stability of the stabilized and synchronized network. Since the con-
trolled Lagrangian has a symmetry, we use Routh reduction and Routh’s criteria to prove stability.
We then design additional dissipative control terms and prove asymptotic stability. We show,
on the one hand, how to apply a dissipative control term that yields convergence to synchronization
staying on a constant momentum surface. In the pendulum/cart system example, this corresponds
to a synchronized motion of the carts such that all the carts move together with a common velocity
that is the sum of a constant plus an oscillation. Likewise, the pendula synchronize and oscillate at
the same frequency as the carts. The oscillation frequency for the carts and pendula is determined
by the control parameters. On the other hand, we show how to apply a dissipative control term
that yields convergence to a relative equilibrium. In the example, this corresponds to steady,
synchronized motion of n carts, each balancing its inverted pendulum.
In this paper we consider a homogeneous group of mechanical systems, i.e., no leaders, and
a fixed, bi-directional, connected communication topology. Possibilities for extension include inte-
gration of the results with prior works cited above that address time-varying and directed commu-
nication topologies and/or the presence of leaders in the group.
The organization of the paper is as follows. In §2 we define notation and the different kinds
of stabilization studied. In §3, we give a brief background on the class of underactuated mechanical

2
systems that satisfy the simplified matching conditions defined in [2, 1]. We discuss how unstable
dynamics are stabilized with feedback control that preserves Lagrangian structure. In §4, we study
a network of n systems, each of which satisfies the simplified matching conditions. We choose
coupling potentials in §5, and we prove stability and coordination of the network. Asymptotic
stabilization is investigated in §6 and §7. We illustrate the theory with the example of n planar,
inverted pendulum/cart systems in §8. In §9 we conclude with a few remarks.

2 Definitions
In [1] the Method of Controlled Lagrangians is used to derive a control law that asymptotically
stabilizes a class of underactuated mechanical systems with otherwise unstable dynamics. This
class of systems satisfies a set of “simplified matching conditions”, and we denote such systems as
SMC systems. SMC systems lack gyroscopic forces; the planar inverted pendulum on a cart and
the spherical inverted pendulum on a 2D cart are two such systems.
Consider an underactuated mechanical system with an (m + r)-dimensional configuration
space. Let xα denote the coordinates for the unactuated directions with index α going from 1 to m.
θa denotes the coordinates for the actuated directions with index a going from 1 to r. In the case of
a network of n mechanical systems, each with the same (m+r)-dimensional configuration space, xαi
and θia are the corresponding coordinates for the ith mechanical system, i = 1, . . . , n. Beginning in
§5, we will assume that the configuration space for the actuated variables for each individual system
is Rr . Note that we only require the configuration space for the individual mechanical systems to
be the same and do not require that each system is identical, e.g., the individual systems can have
different mass and inertia values. We will need to make the assumption of individual systems being
identical only in §6.
The goal of coordination is to synchronize the actuated variables θia with the variables θja for
all i, j = 1, . . . , n. We define stable synchronization of these variables as stabilization of θia − θja = 0
for all i 6= j.
We define the following stability notions for the mechanical system network.

Definition 2.1 (SSRE) A relative equilibrium of the mechanical system network dynamics is a
Stable Synchronized Relative Equilibrium (SSRE) if it is defined by θia − θja = 0 for all i 6= j,
xαi = 0 for all i and if it is Lyapunov stable. This implies that the unactuated dynamics are stable
and the actuated dynamics are stably synchronized.

Definition 2.2 (ASSRE) A relative equilibrium of the mechanical system network dynamics is an
Asymptotically Stable Synchronized Relative Equilibrium (ASSRE) if it is SSRE and asymptotically
stable.

Definition 2.3 (ASSM) An asymptotically stable solution of the mechanical system network dy-
namics is an Asymptotically Stable Synchronized Motion (ASSM) if it is defined by xαi − xαj = 0
and θia − θja = 0 for all i 6= j and the dynamics of the network evolve on a constant momentum
surface.

We note that an ASSRE is a special case of an ASSM. In the example of the network
of pendulum/cart systems, the relative equilibrium of interest corresponds to the carts moving
together at the same constant speed with each pendulum at rest in the upright position. In §8 we
asymptotically stabilize this synchronized relative equilibrium as well as a family of synchronized
motions which exhibit a synchronized steady motion plus an oscillation of the carts and pendula.

3
3 Simplified Matching Conditions
Let the Lagrangian for an individual mechanical system be given by
1 1
L(xα , θa , ẋβ , θ̇b ) = gαβ ẋα ẋβ + gαa ẋα θ̇a + gab θ̇a θ̇b − V (xα , θa )
2 2
where summation over indices is implied, g is the kinetic energy metric and V is the potential energy.
It is assumed that the actuated directions are symmetry directions for the kinetic energy, that is,
we assume gαβ , gαa , gab are all independent of θa . The equations of motion for the mechanical
system with control inputs ua are given by

Exα (L) = 0
Eθa (L) = ua

where Eq (L) denotes the Euler-Lagrange expression corresponding to a Lagrangian L and general-
ized coordinates q, i.e.,
d ∂L ∂L
Eq (L) = − . (3.1)
dt ∂ q̇ ∂q
For such a system, following [1], the simplified matching conditions (SMC) are
• gab = constant

∂gαa ∂gβa
• =
∂xβ ∂xα
∂2V ad ∂2V
• ∂xα ∂θa g gβd = ∂xβ ∂θa
g ad gαd .
Satisfaction of these simplified matching conditions allows for a structured feedback shaping of
kinetic and potential energy. In particular, a control law ua = ucons a is given in [1] such that
the closed-loop system is a Lagrangian system. The controlled Lagrangian Lc , parametrized by
constant parameters κ and ρ and by a potential term V , is given by
 
α a β b 1 ρ−1
Lc (x , θ , ẋ , θ̇ ) = gαβ + ρ(κ + 1)(κ + )gαa g gbβ ẋα ẋβ + ρ(κ + 1)gαa ẋα θ̇a
ab
2 ρ
1
+ ρgab θ̇a θ̇b − V (xα , θb ) − V (xα , θb )
2
where V must satisfy
 
∂V ∂V ρ − 1 ad ∂V
− a
+ a (κ + )g gαd + α = 0. (3.2)
∂θ ∂θ ρ ∂x

The results in [1] further give conditions on ρ, κ and V that ensure stability of the equilibrium
in the full state space. Without loss of generality, we assume that the equilibrium of interest is the
origin. We further assume that it is a maximum of the original potential energy V (the case when
the origin is a minimum can be handled similarly). The inverted pendulum systems fall into this
category. In this case, κ > 0 and ρ < 0 and the potential V can be chosen such that the energy
function Ec for the controlled Lagrangian has a maximum at the origin of the full state space.
Asymptotic stability is obtained by adding a dissipative term udissa to the control law, i.e.,
1
ua = ucons
a + udiss
ρ a

4
which drives the controlled system to the maximum value of the energy Ec .
In [1], it is also shown how to select new, useful coordinates (xα , y a , ẋα , ẏ a ). In particular,
for any SMC system, there exists a function ha (xα ) defined on an open subset of the configuration
space of the unactuated variables such that
∂ha
 
ρ−1
= κ+ g ac gαc , ha (0) = 0.
∂xα ρ
The new coordinates are defined as
(xα , y a ) = (xα , θa + ha (xα )).
Note that if the origin is an equilibrium in the original coordinates it is also an equilibrium in the
new coordinates. In these coordinates, the closed-loop Lagrangian takes the form
 
1 ρ−1 1
Lc = gαβ − (κ + )gαa g ab gbβ ẋα ẋβ + gαa ẋα ẏ a + ρgab ẏ a ẏ b − V (xα , y a − ha (xα )) − V (y a )
2 ρ 2
1 1
= g̃αβ ẋα ẋβ + g̃αa ẋα ẏ a + g̃ab ẏ a ẏ b − V (xα , y a − ha (xα )) − V (y a ), (3.3)
2 2
where
 
ρ−1 ab
g̃αβ = gαβ − (κ + )gαa g gbβ ,
ρ
g̃αa = gαa ,
g̃ab = ρgab . (3.4)
Further, after adding dissipation udiss
a , the Euler-Lagrange equations in the new coordinates become

Exα (Lc ) = 0
Eya (Lc ) = udiss
a .

4 Matching for Network of SMC Systems

In this section we examine a network of n systems each of which satisfies the simplified matching
conditions. We determine what control design freedom remains under the constraint that the
complete network dynamics are Lagrangian and satisfy the simplified matching conditions.
Consider n SMC systems and let the ith system have dynamics described by Lagrangian Li
where
1 i α β 1 i a b
Li (xαi , θia , ẋβi , θ̇ib ) = gαβ i
ẋi ẋi + gαa ẋαi θ̇ia + gab θ̇i θ̇i −Vi (xαi , θia ), (4.1)
2 2
and the index i on every variable refers to the ith system.
1
The Lagrangian for the total (uncontrolled, uncoupled) system is L = ni=1 Li = ẋT M ẋ −
P
2
α , θ a ), where x = (xα , . . . , xβ , θ a , . . . , θ b )T , and
Pn
V (x
i=1 i i i 1 n 1 n
 1 1 
gαβ 0 gαa 0
 .. .. 
 . . 
n n
 
 0 gαβ 0 gαa 
M = 1 
1
.
 gaα 0 gab 0  
 . . . .

 . . 
0 n
gaα 0 n
gab

5
Since each system satisfies the simplified matching conditions, gab i = constant for each i = 1, . . . , n.

It can be easily verified that the simplified matching conditions are satisfied for the total system
L, since they are satisfied for each individual system.
For the total system, the symmetry coordinates are (θ1a , . . . , θnb ). As in [1], we can find a
control law and a change of coordinates x = (xα1 , . . . , xβn , θ1a , . . . , θnb ) 7→ x0 = (xα1 , . . . , xβn , y1a , . . . , ynb )
such that the closed-loop system is equivalent to another Lagrangian system with
1
L0c = (ẋ0 )T Mc ẋ0 − V0 (x0 ) (4.2)
2
and  1
g̃αβ 0 1
g̃αa 0

 .. .. 
 . . 
n n 
   
 0 g̃αβ 0 g̃αa
Mc =   := M11 M12 , (4.3)
 g̃ 1 1 T
M12 M22
 aα 0 g̃ab 0 
 .. .. 
 . . 
0 n
g̃aα 0 n
g̃ab

n 
X 
V0 = Vi (xαi , yia − hai (xαi )) + Vi (xαi , yia ) .
i=1

i , g̃ i , and g̃ i are defined as in (3.4) with all variables replaced with those corresponding
Here, g̃αβ αa αa
to the ith system, e.g., g̃ab i = ρ g i , etc.
i ab
The control gains κi and ρi and control potentials Vi can be chosen such that the mass
matrix Mc is negative definite and the potential V0 has a maximum when the configuration of each
system, i.e., (xαi , θia ), is at the origin. This means the control law brings each system independently
to the origin without coordination.
To determine what additional freedom exists in the choice of the control, notably in the choice
of control potentials Vi , such that the network system satisfies the simplified matching conditions,
we specialize to a network of SMC systems which each satisfy the following condition.

AS1. The potential energy for each system in the original coordinates satisfies Vi (xαi , θia ) =
V1i (xαi ) + V2i (θia ).

The inverted pendulum examples satisfy this assumption in the general case that the cart moves
on an inclined plane. In the case that the cart moves in the horizontal plane, V2 = 0.
As shown in [1], given the assumption AS1, Vi in the new coordinates for i = 1, . . . , n can
be chosen to take the form

Vi (xαi , yia ) = −V2i (yia − hai (xαi )) + V̄i (yia )

where V̄i is an arbitrary function and hai (xαi ) exists that satisfies

∂hai
 
ρi − 1
= κ i + giac gαc
i
, hai (0) = 0. (4.4)
∂xαi ρi

We show next that a more general potential V can be used in V0 in place of the sum of
potentials Vi (xαi , yia ).

6
Proposition 4.1 Under assumption AS1, the potential V0 = V + V satisfies the simplified match-
ing condition with
n
X
V = (V1i (xαi ) + V2i (yia − hai (xαi ))
i=1
n
!
X
V = − V2i (yia − hai (xαi )) + Ṽ (y1a , . . . , yna )
i=1
(4.5)

and Ṽ an arbitrary function.

Proof. Recall that the potential V0 = V + V given by (4.5) satisfies the simplified matching
condition if (3.2) holds. Following [1], we can use the definition of hai (xαi ) given by (4.4) to write
the simplified matching condition (3.2) for the potential as

∂V ∂V ∂ha (xα )

α = a i αi , i = 1, . . . , n . (4.6)
∂xi ∂yi ∂xi

∂V2i ∂via
By a direct computation, one can check that each side of the equation (4.6) is equal to
∂via ∂xαi
where via = yia − hai (xαi ).
Proposition 4.1 implies that we can couple the n vehicles in the network using the freedom in
our choice of Ṽ = Ṽ (y1a , . . . , yna ), and the network dynamics will still satisify the simplified matching
conditions. This result is completely independent of the degree of coupling, i.e., it extends from a
network of uncoupled systems to a network of completely connected systems.

5 Stable Coordination of SMC Network

In this section we make use of Proposition 4.1 to design coupling potentials Ṽ for stable coor-
dination of the network of SMC systems. We prove that the relative equilibrium of interest is a
Stable Synchronized Relative Equilibrium (SSRE). Recall from §2 that to be an SSRE, a relative
equilibrium should be defined by θia − θja = 0 for all i 6= j and xαi = 0 for all i and should be
Lyapunov stable. We note that this is equivalent to showing that yia − yja = 0 for all i 6= j and
xαi = 0 for all i is Lyapunov stable. In the remainder of the paper we assume that the configuration
space for the actuated variables for each individual system is Rr .
To synchronize the actuated variables we use the results of Proposition 4.1 and design coupling
potentials for stabilization of yia − yja = 0, for all i 6= j. Note that the condition yia − yja = 0 for
all i 6= j by itself is necessary but not sufficient for θia − θja = 0 for all i 6= j and xαi = 0 for all i.
We have yia − yja = 0 for all i 6= j under more general conditions, e.g., if θia − θja = 0 for all i 6= j
and hi (xαi ) = hj (xαj ) 6= 0, i 6= j. This more general case makes possible interesting synchronized
dynamics, when we add dissipation for asymptotic stability, as will be discussed in §6.
We choose Ṽ such that the closed-loop potential V0 , defined in Proposition 4.1, has a maxi-
mum when xαi = 0 and yia − yja = 0 for all i 6= j. This is possible since from (4.5), the closed-loop
potential is V0 = ni=1 (V1i (xi )) + Ṽ (y1a , . . . , yna ) and the V1i are assumed to already be maximized
P
at xαi = 0. We choose in this paper Ṽ to be quadratic in (yia − yja ) with maximum at yia − yja = 0
for all i 6= j. In this case, consider a graph with one node corresponding to each individual system
in the network. There is an (undirected) edge between nodes k and l if the term (yka − yla ) appears

7
in the quadratic function Ṽ . Then, V0 has a strict maximum when xαi = 0 and yia − yja = 0 for
all i 6= j, if the (undirected) graph is connected. Figure 5.1 illustrates an example of a connected,
undirected communication graph for four vehicles.

Figure 5.1: Connected, undirected communication graph for four vehicles.

With coupling of the individual systems using terms that depend only on yia − yja , the net-
work system has a translational symmetry. Specifically, the system dynamics are invariant under
translation of the center of mass of the network. Consider a new set of coordinates given by

xc = (xα1 , . . . , xβn , z1a , . . . , znb )T (5.1)

where

zia = y1a − yi+1

a
, i = 1...,n − 1
znb = y1b + ... + ynb .

In this coordinate system, the controlled Lagrangian for the total system (with abuse of notation
for V0 ) is
1
L̃c = ẋTc M̃c ẋc − V0 (xr ) (5.2)
2
where xr = (xα1 , . . . , xβn , z1a , . . . , zn−1
b )T and
 
M̃11 M̃12
M̃c = T . (5.3)
M̃12 M̃22

The transformation which takes the coordinates xc to the coordinates x0 = (xα1 , . . . , xβn , y1c , . . . , ynd )
is given by the matrix  
Imn×mn 0
B= (5.4)
0 B22
where  
Ir×r Ir×r ... Ir×r
1 (1 − n)Ir×r Ir×r ... Ir×r 
B22 = (5.5)
 
.. .. ..
n
 
. . ··· . 
Ir×r . . . (1 − n)Ir×r Ir×r
and Il×l denotes an l × l identity matrix and B22 is an rn × rn matrix. The expression for M̃c in
terms of Mc from (4.3) is
M̃c = B T Mc B. (5.6)

8
 We can compute the block elements in M̃c to be

M̃11 = M11 , (5.7)

1 1 1 1
 
g̃αa g̃αa ... g̃αa g̃αa
2
 (1 − n)g̃αa 2
g̃αa ... 2
g̃αa 2
g̃αa 
1 ..

M̃12 =  , (5.8)
 
.
n n−1 n−1 n−1 n−1

 g̃αa g̃αa ... g̃αa g̃αa 
n
g̃αa n
g̃αa n
. . . (1 − n)g̃αa n
g̃αa
1 T
B M22 B22
M̃22 = (5.9)
n2 22
where M11 and M22 are as defined in (4.3). From (5.5) and (4.3), we can calculate the lowermost
diagonal r × r block of M̃22 to be
n
1 X i
g̃ab = 2 (g̃ab ). (5.10)
n
i=1

Thus, we can define M̄22 = g̃ab and M̄11 and M̄12 in terms of M̃c such that
 
M̄11 M̄12
T = M̃c .
M̄12 M̄22

Then, we can rewrite (5.2) as

  
1 M̄11 M̄12 ẋr
ẋTr żTn − V0 (xr )


L̃c = T
2 M̄12 M̄22 żn

where zn = (zna )T .
Note that in these coordinates zna is the symmetry variable. We are interested in the relative
equilibria given by  
xr
vRE :=  ẋr  (5.11)
żn
where
xr = 0, ẋr = 0, żnd = ζ d
and ζ d corresponds to (n times) the constant velocity of the center of mass of the network.

Definition 5.1 (Amended Potential [13]) The amended potential for the Lagrangian system
with Lagrangian (5.2) is defined by

1
Vµ (xr ) = V0 (xr ) + g̃ cd µc µd
2
where V0 is given by (4.5) and g̃ab is given by (5.10). If Ja is the momentum conjugate to zna , then
µa is Ja evaluated at the relative equilibrium corresponding to żna = ζ a , i.e.,

∂ L̃c T ∂ L̃ c

Ja = = (M̄ ẋ
12 r + M̄ ż )
22 n a , µ a = = g̃ab ζ a . (5.12)
∂ żna ∂ żna 

xr =0,ẋr =0,żn =ζ
a a

9
 By the Routh criteria, the relative equilibrium is stable if the second variation of
1 −1
Eµ := ẋTr (M̄11 − M̄12 M̄22 T
M̄12 )ẋr + Vµ (xr ) (5.13)
2
evaluated at the origin is definite. Also, if Rµ (xr , ẋr ) is defined as
1 −1
Rµ := ẋTr (M̄11 − M̄12 M̄22 T
M̄12 )ẋr − Vµ (xr ), (5.14)
2
then the reduced Euler-Lagrange equations can be written as

Exαr Rµ = 0.

The Routhian Rµ plays the role of a Lagrangian for the reduced system in variables (xr , ẋr ). Since
i is a constant for each i ∈ {1, 2, . . . , n}, the second term in the amended potential V does not
g̃ab µ
contribute to the second variation. It follows that the relative equilibrium with momentum µa is
−1 T ) evaluated at the origin is negative definite, since the
stable if the matrix (M̄11 − M̄12 M̄22 M̄12
−1
potential V0 is already maximum at the equilibrium. But (M̄11 − M̄12 M̄22 T ) is negative definite
M̄12
because it is the Schur complement of the negative definite matrix M̃c [9].

Theorem 5.2 (SSRE) Consider a network of n SMC systems each satisfying Assumption AS1.
Suppose for each system that the origin is an equilibrium and that the original potential energy is
maximum at the origin. Consider the kinetic energy shaping defined in §4 and potential energy
coupling defined above with connected graph so that the closed-loop dynamics derive from the La-
grangian L̃c given by (5.2) and the potential energy V0 is maximized at the relative equilibrium
(5.11). The corresponding control law for the ith mechanical system is
  
1 i
cons i i δα i
ua,i = ua,i = − κi gβa,γ − gδa Ai gαβ,γ − gβγ,α − (1 + κi )gαd gi gβa,γi da i
ẋβi ẋγi
2
0 (5.15)
i ∂V i ∂V i 1 
db ∂V

+ κi gδa Aδα
i + − 1 + κ g i
A
i δa i
δα i
g g
αd i
∂xαi ∂θia ρi ∂θia
n   o
where Aiαβ = gαβ
i −(1+κ )g i g da g i , ρ < 0 and κ +1 > max λ| det g i − λg i g ab g i
i αd i βa i i αβ αa i bβ | x α =0 = 0 .
i

Then, the relative equilibrium (5.11) is a Stable Synchronized Relative Equilibrium (SSRE) for any
ζ d.
−1 T ) evaluated at the origin is negative definite, the second variation
Proof. Since (M̄11 − M̄12 M̄22 M̄12
of Eµ evaluated at the origin is definite. Hence, the relative equilibrium (5.11) is stable for the
total network system independent of momentum value µa .

6 Asymptotic Stability of Constant Momentum Solution

In this section we investigate asymptotic stabilization of the coordinated network to a solution
corresponding to a constant momentum Ja = µa . We prove that the solution is an Asymptotically
Stable Synchronized Motion (ASSM). Recall from §2 that an ASSM is an asymptotically stable
solution of the mechanical system network defined by xαi = xαj and θia = θja for all i 6= j and
dynamics that evolve on a constant momentum surface. An ASSM describes a fully synchronized
motion, i.e., one in which each degree of freedom is synchronized across the whole network. If
xαi = xαj = 0 then the solution is a relative equilibrium. However, in general, an ASSM is not a

10
relative equilibrium. For example, in the case of a network of pendulum/carts presented in §8, the
ASSM correspond to periodic solutions (synchronized oscillations of pendula and carts). In this
section we prove a control law that yields ASSM where the constant value of the momentum is given
by the initial conditions. Equivalently, given an arbitrary momentum value µa , initial conditions on
the corresponding momentum surface converge to ASSM on the same momentum surface. In the
pendulum/cart example of §8, we show that control gains can be used to determine the frequency
of the periodic solution (ASSM). We discuss at the end of the section a second case in which a
momentum value is prescribed and a control term is added to drive the ASSM to the prescribed
constant momentum surface.
In this section we apply no dissipative control in the xαi directions for all i and as our Case
I we use no control in the zna direction. Recall that for our closed-loop system, zna is the symmetry
direction. If there is no control applied in this direction, Ja remains a constant, i.e., the system
evolves on a constant momentum surface. On this surface, Eµ as defined in (5.13) is a conserved
quantity and can be chosen as a Lyapunov function to prove stability. By choosing appropriate
dissipation in the non-symmetry directions z1a , . . . , zn−1
b , we prove that solutions on a constant
momentum surface, corresponding to xi − xj = 0 and θia − θja = 0 for all i 6= j, are asymptotically
α α

stable, i.e., they are ASSM.

Let the control input for the ith mechanical system be
1 diss
ua,i = ucons
a,i + u (6.1)
ρi a,i
where ucons diss
a,i is the “conservative” control term given by (5.15) and ua,i is the dissipative control term
to be designed. The Euler-Lagrange equations in the original coordinates for the ith uncontrolled
system are
1 diss
Exαi (Li ) = 0 ; Eθia (Li ) = ucons
a,i + u
ρi a,i
where Li is given by (4.1).
In the new coordinates given by (5.1), we have for i = 1, . . . , n
1 diss
Exαi (L̃c ) = 0 ; Ezia (L̃c ) = ũ (6.2)
n a,i
where L̃c is given by (5.2) and
n
X
ũdiss
a,i = udiss diss
a,j − (n − 1)ua,i+1 , i = 1, . . . , n − 1
j=1,j6=i+1
Xn
ũdiss
a,n = udiss
a,j .
j=1

Case I: ũdiss
a,n = 0.
Let Ẽc be the energy function for the Lagrangian L̃c . Given momentum value µa , let ξ b =
g̃ µa . Then, the function Ẽcξ defined by
ab

Ẽcξ = Ẽc − Ja ξ a
has the property that its restriction to the level set Ja = µa = g̃ab ξ b of the momentum gives Eµ
(5.13). We can use this fact to calculate the time derivative of Eµ as follows. From (6.2), we get
n
d 1 X a diss
Ẽc = (żi ũa,i ). (6.3)
dt n
i=1

11
 d 1
Using (6.3) and the fact that Ja = ũdiss , we get
dt n a,n
n
d ξ 1 X a diss 1
Ẽc = (żi ũa,i ) − ( ũdiss ξ a ). (6.4)
dt n n a,n
i=1

d ξ
The expression for time derivative of Eµ is obtained by restricting Ẽc to the set Ja = µa . This
dt
and (5.12) gives us
n−1
d 1 X a diss 1
Eµ = (żi ũa,i ) + ũdiss ( ż a | − ξa)
dt n n a,n n Jb =µb
i=1
n−1
1 X 1 diss ab
= (żia ũdiss
a,i ) +
T
ũ (g̃ (µb − (M̄12 ẋr )b ) − ξ a )
n n a,n
i=1
n−1
1 X 1 diss
= (żia ũdiss
a,i ) + ũ (−g̃ ab (M̄12
T
ẋr )b ).
n n a,n
i=1

T ẋ is a covariant vector just like a momentum. Hence, its components are denoted by
Here, M̄12 r
subscripts. Since ũdiss
a,n is chosen to be zero, we get

n−1
d 1 X a diss
Eµ = (żi ũa,i ). (6.5)
dt n
i=1

Expressing ũdiss diss

a,i in terms of ua,i , we can write the expression for Ėµ as
 
n−1 n−1 n−1
d X X X
n Eµ = udiss
a,1 ( żja ) + udiss
a,j
a
−(n − 1)żj−1 + żka  (6.6)
dt
j=1 j=2 k=1,k6=j−1

and choose
 
n−1
X
udiss
a,1 = dab
 żjb 
j=1
 
n−1
X
udiss b
a,j = dab −(n − 1)żj−1 +
 żkb 
k=1,k6=j−1

j = 2, . . . , n − 1, (6.7)

where dab is a positive definite control gain matrix, possibly dependent on xαi , i = 1, . . . , n, and zia ,
d
j = 1, . . . , n − 1. With the dissipative control term (6.7), Eµ ≥ 0.
dt
We note that this dissipative control term requires that each individual system can measure
the variables żia of all other vehicles. Recall that for Lyapunov stability the interconnection among
individual systems need only be connected for the coupling potential Ṽ which is a function of the yka ,
k = 1, . . . , n. That is, for Lyapunov stability, each individual system need only measure its relative
position with respect to some subset of the other individual systems. However, for asymptotic
stability (ASSM) we require complete interconnection in the dissipative control term which is a

12
function of the variables żn . That is, each individual system feedbacks relative velocity with respect
to every other individual system. Figure 6.1 illustrates a complete interconnected graph for the
case of four vehicles. Complete interconnection is not needed for stabilization of group dynamics in
the simpler dynamical models used more typically in the literature as described in §1. It is hoped
that the interconnection limitation here in stabilization of networks of underactuated mechanical
systems can likewise be overcome in future work.

Figure 6.1: Complete interconnected communication graph for four vehicles.

We next study convergence of the system using the LaSalle Invariance Principle [11]. For
c > 0, let Ωc = {(xr , ẋr )|Eµ ≥ c}. Ωc is a compact and positive invariant set with integral curves
starting in Ωc staying in Ωc for all t ≥ 0. Define the LaSalle surface
  
d
E = (xr , ẋr )  Eµ = 0 .

dt
On this surface, udiss a
a,j = 0, i = 1, . . . , n which implies that żi = 0 for i = 1, . . . , n − 1. Let M be the
largest invariant set contained in E. By the LaSalle Invariance Principle, solutions that start in Ωc
approach M. The relative equilibrium (5.11) is contained in M; however, there are other solutions
in this set.
We now proceed to analyze in more detail the structure of solutions on the LaSalle surface
E. Using the condition żia = 0 for i = 1, . . . , n − 1, we get ẏia = ẏja for all i, j ∈ {1, . . . , n}. This
gives yia − yja = constant. Since we have chosen Ṽ to be a quadratic function of the terms yia − yja ,
∂ Ṽ
we get = constant =: ∆ia . The equations of motion for the yia restricted to the LaSalle surface
∂yia
are Eyia (L0c ) = 0, where L0c is given by (4.2). Equivalently,

d  ab i α  ∂ Ṽ
ÿia + g̃ g ẋ = −g̃iab b = −g̃iab ∆ib . (6.8)
dt i αb i ∂yi
As illustrated in [1], for SMC systems, there is a function lia (xαi ) for each vehicle i defined on
an open set of the configuration space for the ith vehicle’s unactuated variables such that
∂lia
= g̃iac gαc
i
. (6.9)
∂xαi
We can assume, by shrinking Ωc if necessary, that (6.9) holds in Ωc .
Let Kc be the projection of Ωc onto the coordinates (xn , ẋn ) where xn = (xα1 , . . . , xαn ). Then,
since lia is continuous and Kc is compact, there exist constants mi and ni such that
mi ≤ ||li (xi )|| ≤ ni (6.10)

13
for all xαi such that xn ∈ Ωc . Using (6.8), (6.9) and the condition ẏia = ẏja on E, we get

d ˙a ˙a
(li − lj ) = g̃jab ∆jb − g̃iab ∆ib . (6.11)
dt
Therefore, on E
1
lia − lja = (g̃jab ∆jb − g̃iab ∆ib )t2 + ν1a t + ν2a (6.12)
2
for some constant vectors ν1a and ν2a . The only way (6.10) can also be satisfied is if g̃jab ∆jb −g̃iab ∆ib = 0
and ν1a = 0.
To simplify our calculations, we assume the n individual mechanical systems to be identical.
In this case, g̃jab = g̃jab for any i, j ∈ {1, . . . , n}. This gives, ∆ia = ∆ja for any i, j ∈ {1, . . . , n} and so
for a connected network with potential V0 having a maximum at xαi = 0 and yia = yja for all i 6= j,
we get that yia = yja on E for all i, j ∈ {1, . . . , n}.
Using the definition (6.9) and the assumption that the individual systems are identical, the
fact that l˙ia − l˙ja = 0 on E yields
i j
gαb ẋαi = gαb ẋαj , (6.13)
k = g (xα ), for all k = 1, . . . , n. Therefore, on the LaSalle surface E, we see that
where gαb αb k
where yia (t) = yja (t) for any
solutions are of the form (xn (t), ẋn (t), y1a (t), . . . , ynb (t), ẏ1c (t), . . . , ẏnd (t))P
a n a
i, j ∈ {1, . . . , n}, Ja = µa and condition (6.13) holds. Since zn = i=1 yi and the individual
systems are identical, we have
n
∂ L̃c X i α
Ja = = (gαa ẋi + g̃ab ẏib )
∂ żna
i=1
n
X
= g̃ab (g̃ bc gαc
i
ẋαi + ẏib )
i=1
= ng̃ab (g̃ bc gαc
i
ẋαi + ẏib )

for any i ∈ {1, . . . , n}, where we have used the facts that ẏia = ẏja and (6.13) holds on E. Therefore,
for each i we get
1
ẏia = g̃ ab µb − g̃ ab gαb
i
ẋαi . (6.14)
n
Substituting (6.14) into the closed-loop equations for the Lagrangian L0c (4.2), we get the
following equations for the xαi variables,
d ∂Lµ ∂Lµ
= (6.15)
dt ∂ ẋαi ∂xαi
where
n  
1
)ẋαi ẋβi
X
µ i
L = (g̃αβ − g̃ ab gαa
i i
gβb − V1i (xαi )
2
i=1
n  
X 1 i ab i i α β α
= (g − (κ + 1)g gαa gβb )ẋi ẋi − V1i (xi ) , (6.16)
2 αβ
i=1

and V1i is defined by assumption AS1. Here, κi = κ for all i = 1, . . . , n.

Lµ is just the Routhian Rµ for a mechanical system with abelian symmetry variables without
a linear term in velocity and without the amended part of the potential. This follows because, for

14
SMC systems, these latter terms do not contribute to the dynamics of the reduced system. We also
see that the xαi dynamics completely decouple from the xαj dynamics on the LaSalle surface E for
all i and j. The yia dynamics given by (6.14) can be thought of as a reconstruction of dynamics in
the symmetry variables, obtained after solving the reduced dynamics in the xαi variables. We now
make the following assumption.
AS2. Consider two solutions (xα (t), y a (t)) and (x̃α (t), ỹ a (t)) of the Euler-Lagrange
equations corresponding to the Lagrangian given by (3.3). If y a (t) = ỹ a (t) and gαa (xα (t))ẋα (t) =
gαa (x̃α (t))x̃˙ α (t) then xα (t) = x̃α (t).
Note that checking this condition does not require extensive computation since we already know
the expression for the closed-loop Lagrangian. Consider two solutions xα (t) and x̃α (t) such that
gαa (xα (t))ẋα (t) = gαa (x̃α (t))x̃˙ α (t). This is equivalent to la (xα ) = la (x̃α ) + ca where la is defined
by (6.9) and ca is a constant, i.e., xα (t) and x̃α (t) are two solutions in (xα , ẋα ) space satisfying the
Euler-Lagrange equation corresponding to the Lagrangian Lµ given by (6.16) and differing by a
constant. For mechanical systems with symmetries, it may be possible to prove that ca is zero, as
is done for the pendulum/cart case in §8. Then, AS2 is equivalent to assuming that the function la
is injective, i.e., gαa is one-to-one in a neighborhood about the equilibrium. For the pendulum/cart
example in §8, this holds in the neighborhood defined by pendulum angles which are above the
horizontal plane. As mentioned in [1], AS2 is equivalent to the (local) strong inertial coupling
assumption in [20] and internal/external convertible system in [7].
Using (6.13) and the fact that yia = yja on the LaSalle surface, we get from AS2 that xαi = xαj
and θi = θja for all i, j ∈ {1, . . . , n}. So we get that the dissipation control law given by (6.7) yields
a

asymptotic convergence to synchronized motion on a constant momentum surface (ASSM).

Theorem 6.1 (ASSM) Consider a network of n identical SMC systems that each satisfy AS1
and AS2. Suppose for each individual system that the origin is an equilibrium and that the original
potential energy is maximum at the origin. Consider the kinetic energy shaping defined in §4 and
potential energy coupling Ṽ defined in §5 where the terms in Ṽ are quadratic in yia − yja and the
corresponding interconnection graph is connected. The closed-loop dynamics (6.2) derive from the
Lagrangian L̃c given by (5.2) and the potential energy V0 is maximized at the relative equilibrium
(5.11). The control input takes the form (6.1) where ucons
a,i is given by (5.15) and ρi = ρ, κi = κ.
The dissipative control term given by equation (6.7) asymptotically stabilizes the solution in which
all the vehicles have synchronized dynamics such that θia = θja and xαi = xαj for all i and j, and each
has the same constant momentum in the θia direction. The system stays on the constant momentum
surface determined by the initial conditions.

Remark 6.2 Consider Case II in which we choose ũdiss a,n = −λ(Ja −µa ) and ua,i for i = 1, . . . , n−1
as in Case I. Then Ja = (Ja (0) − µa ) exp(−λt) + µa and we can rewrite the reduced system in
(xr , ẋr ) coordinates as follows:
!
0
Exr (Rµ ) = 1 diss + λM̄12 M̄ 22 (J(0) − µ) exp(−λt). (6.17)
ũ
n
Here, ũdiss = (ũdiss diss
a,1 , . . . , ũa,n−1 ) is an rn-dimensional vector, J and µ are r-dimensional vectors with
components Ja and µb , respectively. When λ = 0, we get Case I. When λ 6= 0, the momentum Ja is
no longer a conserved quantity. This case needs to be analyzed more carefully since we are pumping
energy into the system now to drive it to a particular momentum value. Equation (6.17) can be
considered to be a parameter dependent differential equation with the parameter being λ. When

15
λ = 0, we already know the solution from Case I. From the continuity of dependence of solutions
upon parameters, we get that when 0 < λ < δ, the solution stays within an −tube of the solution
in Case I for time t ∈ [0, t1 ] for some t1 if the initial conditions are in a δ-neighborhood. Our
simulations for pendulum/cart systems suggests that this holds true for the infinite time interval.
We plan to investigate this case further in our future work.

Remark 6.3 The simplifying requirement for Theorem 6.1 that all systems be identical is a weak-
ness of the result and motivates the question of robustness to uncertainty in system parameters.
Simulations suggest that the stability of Theorem 6.1 is robust to model parameter uncertainty, but
a formal robustness analysis is warranted.

In Section §8 we illustrate the result of Theorem 6.1 and the dynamics of (6.16) in more detail
in the case of a network of inverted pendula/cart systems. Solutions for this example correspond
to synchronized balanced pendula on synchronized moving carts where the motion of the carts is
the sum of a constant velocity plus an oscillation and the motion of the pendula is oscillatory with
the same frequency as the carts.

7 Asymptotic Stabilization of Relative Equilibria

In the previous section, we proved asymptotic stability of the coordinated network in the case when
the network asymptotically converges to the momentum surface Ja = µa . This can lead to nontrivial
and interesting synchronized group dynamics as is discussed in §8. Stabilization was proved using
Eµ as a Lyapunov function on the reduced space. The dynamics after adding a dissipative control
term are given by θia = θja and xαi = xαj for all i, j = 1, . . . , n. The dissipative terms are chosen
such that the momentum is preserved.
In this section, we demonstrate how to isolate and asymptotically stabilize the particular
synchronized and constant momentum solutions corresponding to the relative equilibria given by
(5.11). The value of the momentum µa can be chosen arbitrarily. We use a different Lyapunov
function from that used in §6. We note that in the example of a network of inverted pendula/cart
systems, the relative equilibrium corresponds to the synchronized motion of all carts moving in
unison at steady speed with all pendula at rest in the upright position, i.e., it is the special case of
the motion proved in Theorem 6.1 without the oscillation.
Consider the following function:
1
ERE = (ẋc − vRE )T M̃c (ẋc − vRE ) + V0 (7.1)
2
where vRE is defined by (5.11). ERE is a Lyapunov function in directions transverse to the group
orbit of the relative equilibrium, i.e., ERE > 0 in a neighborhood of the Euler-Lagrange solution
given by (xr , zn , ẋr , żn ) where xr = 0, ẋr = 0, znd = ζ d t, żnd = ζ d and ζ d corresponds to (n times)
the constant velocity of the center of mass of the network.
The time derivative of ERE along the flow given by (6.2) can be computed to be
 
d 1 0
ERE = (ẋc − vRE ) · .
dt n ũdiss

See [3] for the steps involved in proving this identity. Choose
(
diss nσi żia for i = 1, . . . , n − 1
ũa,i = a a
(7.2)
nσn (żn − ζ ) for i = n

16
where control parameters σi are positive constants. Then,
n−1
d X
ERE = σi (żja )2 + σn (żnb − ζ b )2 ≥ 0.
dt
j=1

We note here, that unlike the case of asymptotic stabilization in the previous section where
a complete interconnection was required to realize dissipative control term (6.7), the dissipative
control term (7.2) only requires a connected interconnection graph.
Let ΩRE
c = {(xr , ẋr , żna )|ERE ≥ c} for c > 0. ΩRE
c is a compact set, i.e., ERE is a proper Lya-
punov function. Assume that the Euler-Lagrange system (6.2) satisfies the following controllability
condition.
AS3. The system (6.2) is linearly controllable at each point in a neighborhood of the
relative equilibrium solution manifold.
Note that checking this condition does not require extensive computation since we already know
the expression for the closed-loop Lagrangian.
We now use a result from nonlinear control theory which is stated as Lemma 2.1 in [3] and
the remark after that to conclude that the system (6.2) with dissipative control terms given by
(7.2) converges exponentially to the set

ERE = {(xr , ẋr , żna ) |ERE = 0 } .

On this set, the solution is given by (5.11). Thus, we have shown that the solutions of the controlled
1 1
system will exponentially converge to (xαi , θia , ẋβi , θ̇ib ) = (0, ζ a t + γ a , 0, ζ b ), with γ a constant.
n n
Theorem 7.1 (ASSRE) Consider a network of n (not necessarily identical) individual SMC sys-
tems that each satisfy Assumption AS1. Suppose for each individual system that the origin is an
equilibrium and that the original potential energy is maximum at the origin. Consider the kinetic
energy shaping defined in §4 and potential energy coupling Ṽ defined in §5 where the terms in Ṽ
are quadratic in yia − yja and the corresponding interconnection graph is connected. The closed-loop
dynamics (6.2) derive from the Lagrangian L̃c given by (5.2) and the potential energy V0 is maxi-
mized at the relative equilibrium (5.11). The control input takes the form (6.1) where ucons
a,i is given
by (5.15) and ρi = ρ. If (6.2) satisfies AS3, then the dissipative control term given by equation
(7.2) exponentially stabilizes the relative equilibrium given by (5.11) in which xαi = ẋαi = 0 for all
1
i = 1, . . . , n and θia = θja and θ̇ia = θ̇ja = ζ a for all i and j.
n

8 Coordination of Multiple Inverted Pendulum/Cart Systems

As an illustration, we now consider the coordination of n identical planar inverted pendulum/cart
systems. For the ith system, the pendulum angle relative to the vertical is xi and the position of
the cart is θi . Let the Lagrangian for each system shown in Figure 8.1 be
1 1
Li = αẋ2i + β cos(xi )ẋi θ̇i + γ θ̇i2 + D cos(xi ) ; i = 1, . . . , n
2 2
where l, m, M are the pendulum length, pendulum bob mass and cart mass, respectively. g is the
acceleration due to gravity. The quantities α, β, γ and D are expressed in terms of l, m, M, g as
follows:
α = ml2 , β = ml, γ = m + M, D = −mgl.

17
 m
g
x
l

M
θ

Figure 8.1: The planar pendulum on a cart.

The equations of motion for the ith system are

Exi (Li ) = 0
Eθi (Li ) = ui

where ui is the control force applied to the ith cart.

One can see that θi is a symmetry variable. Further, it can be easily verified that each
pendulum/cart system satisfies the simplified matching conditions [1, 2]. The n inverted planar
pendulum/cart systems lie on n parallel tracks corresponding to the θi directions. The coordi-
nation problem is to prescribe control forces ui , i = 1, . . . , n, that asymptotically stabilize the
solution where each pendulum is in the vertical upright position (in the case of ASSRE) or moving
synchronously (in the case of ASSM) and the carts are moving at the same position along their
respective tracks with the same common velocity. The relative equilibrium vRE (5.11) corresponds
to xi = ẋi = 0 for all i, θi = θj for all i 6= j and θ̇i = n1 ζ for some constant scalar velocity ζ.
Following (5.2), the closed-loop Lagrangian for the total system in the coordinates xc =
(x1 , . . . , xn , z1 , . . . , zn ) where zi = y1 − yi+1 for i = 1, . . . , n − 1, zn = y1 + · · · + yn , yi = θi + p sin xi
1
and p = βγ (κ + 1 − ) is
ρ
1
L̃c = ẋT M̃c ẋ − V0 (x1 , . . . , xn , z1 , . . . , zn−1 ). (8.1)
2
M̃c is as in (5.6) and Mc is as in (4.3),

i 1 β2
g̃αβ = α − (κ + 1 − ) cos2 (xi ), i
g̃αa = β cos(xi )
ρ γ
n−1
X 1 γ2 2

i 0
g̃ab = ργ, V = −D cos(xi ) −  2 zi − D cos(xn ) (8.2)
2 β
i=1

with  > 0. The control law (6.1) for the ith system is

∂V0
  
κβ sin xi αẋ2i + cos(xi )D − Bi − udiss


i
∂θi
ui = 2 (8.3)
α − βγ (1 + κ) cos2 (xi )

18
 β 2 cos2 (xi )
 
1
where Bi = α− . Note that we have chosen ρi = ρ and κi = κ. In the case
ρ γ
udiss
i = 0, by Theorem 5.2, we get stability of the relative equilibrium vRE (SSRE) if we choose
β2
ρ < 0,  > 0 and κ such that mκ := α − (κ + 1) < 0. The choice of udiss
i depends upon what kind
γ
of asymptotic stability we want, i.e, convergence to a synchronized constant momentum solution
or to a relative equilibria.
The dependence of V0 on zi2 in (8.2) implies that coupling between the pendulum/cart systems
introduced by the control is a function of terms yi − yj rather than θi − θj . That is, our approach to
simultaneous stabilization and synchronization of a network of planar pendulum/cart systems yields
coupling not simply as a function of relative cart positions but rather more subtly as a function of
the horizontal component of relative positions of pendulum bobs (where pendulum length is scaled
by p). Numerical simulations show that naively coupling the positions of the carts for the purpose
of synchronization in fact destabilizes the network. This particular example illustrates the need to
integrate synchronization and stabilization tasks.

8.1 Asymptotic stability on constant momentum surface (ASSM)

Following (6.7), we let udiss
1 be
n−1
!
X
udiss
1 = d1 (żk )
k=1

and udiss
i for i = 2, . . . , n be
 
n−1
X
udiss
i = di −(n − 1)żi−1 + żk 
k=1,k6=i−1

where coefficients di are constant positive scalars.

We now analyze the dynamics on the LaSalle surface. On this surface, we have ẏi = ẏj for
all i, j ∈ {1, . . . , n} and J = µ where momentum µ is determined by the initial conditions. From
the calculations made in §6, we also get yi = yj and cos(xi )ẋi = cos(xj )ẋj . The xi dynamics are
given by (6.15) with
n  
β2

X 1 
Lµ = α − (κ + 1) cos2 (xi ) ẋ2i + D cos(xi ) . (8.4)
2 γ
i=1

To verify AS2 we need to check that if cos(xi )ẋi = cos(xj )ẋj about the origin for a system
corresponding to the Lagrangian Lµ , then xi = xj identically. This condition can also be written
as sin(xi ) = sin(xj ) + c, where c is a constant. Note that if xi (t) is an Euler-Lagrange solution
corresponding to Lµ for the ith vehicle, then −xi (t) is also a solution. Since we have a stable
pendulum oscillation about the upright position, xi (t) and therefore | sin(xi (t))| oscillates with
mean zero for all i. This can also be concluded from the fact that the solution curves are closed
level curves in the (xi , ẋi ) plane of Lµ given by (8.4) and Lµ is invariant under the sign change
(xi , ẋi ) 7→ −(xi , ẋi ). Since | sin(xi )| oscillates with zero mean for all i, the constant c must be zero.
Hence, xi (t) = xj (t) for all i, j identically and AS2 is verified. Thus by Theorem 6.1 the pendulum
network asymptotically goes to an ASSM.
From (8.4), it can be seen that on the LaSalle surface, the dynamics of xi are decoupled
from the dynamics of xj for all i 6= j. For small xi , the dynamics of each individual term in Lµ

19
corresponds to the stable dynamics of a spring-mass system with a κ-dependent mass −mκ > 0
and spring constant −D > 0. The mass −mκ , which determines the oscillation frequency of the
pendulum for each individual cart, can be controlled by choice of κ. For the nonlinear system also,
constant energy curves are closed curves in the (xi , ẋi ) plane. Hence, we have a periodic orbit for
the angle made by each pendulum with the vertical line with a κ−dependent frequency. On the
LaSalle surface, J = ργ θ̇i + (β + pργ) cos(xi )ẋi = constant. Therefore, the velocity of the cart θ̇i
oscillates about a constant velocity with the same frequency as the pendulum oscillation.
Figure 8.2 shows the results of a MATLAB simulation for the controlled network of pen-
dulum/cart systems using the following values for the system parameters. The pendulum/cart
systems have identical pendulum bob masses, lengths and cart masses. The pendulum bob mass is
chosen to be m = 0.14 kg, cart mass is M = 0.44 kg, pendulum length is l = 0.215 m. The control
gains are ρ = −0.27, κ = 40, di = d = 0.2 and  = 0.0005. We compute mκ = −0.058 kgm2 < 0 as
required for stability. The initial conditions for the two systems shown are

( x1 (0) ẋ1 (0) θ1 (0) θ̇1 (0) x2 (0) ẋ2 (0) θ2 (0) θ̇2 (0) )
= ( 0.48 0.99 0.37 0.53 0.18 0.50 0.42 0.66 ).

Figure 8.2 shows plots of the pendulum angle, cart position and cart velocity as a function of time
for two of the coupled pendulum/cart systems. Convergence to an ASSM is evident. The frequency
of oscillation of the pendula can be observed to be the same as the frequency
p of oscillation in the
cart velocities. This frequency of oscillation can be computed as ω = D/mκ and the period
of oscillation as T = 2π/ω = 2.8 s which is precisely the period of the oscillations observed in
Figure 8.2.

8.2 Asymptotic stability of relative equilibria (ASSRE)

In this case, we want to asymptotically stabilize the relative equilibrium vRE , i.e., xi = ẋi = 0 for
all i, θi = θj for all i 6= j and θ̇i = n1 ζ for all i and any constant scalar velocity ζ. Recall that
this corresponds to each pendulum angle at rest in the upright position and all carts aligned and
moving together with the same constant velocity n1 ζ. Following (7.2), we let

udiss
i = ndi żi

for i = 1, . . . , n − 1 and
udiss
n = ndn (żn − ζ)
where the control parameters di are positive constants.
Figure 8.3 shows the results of a MATLAB simulation for the controlled network of pendu-
lum/cart systems with this dissipative control. We choose ζ = 2n m/s and the remaining system
and control parameters are as above in the ASSM case. The initial conditions for the two systems
shown are

( x1 (0) ẋ1 (0) θ1 (0) θ̇1 (0) x2 (0) ẋ2 (0) θ2 (0) θ̇2 (0) )
= ( 0.53 1.12 0.56 0.50 1.02 0.63 0.24 0.81 ).

Figure 8.3 shows convergence to the relative equilibrium; the pendula are stabilized in the upright
position, the cart positions become synchronized and the cart velocities converge to 2 m/s.

20
 2

x 1 , x 2 (rad)
0

−2
0 20 40 60 80 100

200
θ1 , θ2 (m)

100

0
0 20 40 60 80 100
time (sec)
10
θ , θ (m/sec)

0
.
2

.
1

−10
0 20 40 60 80 100
time (sec)

Figure 8.2: Simulation of a controlled network of pendulum/cart systems with dissipation designed
for asymptotic stability of a synchronized motion on a constant momentum surface (ASSM). The
pendulum angle, cart position and cart velocity are plotted as a function of time for each of two
pendulum/cart systems in the network.

9 Final Remarks
We have derived control laws to stabilize and stably synchronize a network of mechanical systems
with otherwise unstable dynamics. We have proved stability of relative equilibria corresponding to
synchronization in all variables and common steady motion in the actuated directions. Using two
different choices of a dissipative term in the control law we prove two different kinds of asymptotic
stability. In the first case of dissipation, we show how to drive the network to a synchronized motion
on the constant momentum surface determined by the initial conditions. Such a synchronized
motion can be interesting when examined in physical space. In our example of a network of
planar pendulum/cart systems, we show that the synchronized motion is periodic and the period
of the oscillation can be controlled with a control parameter. In the second case of dissipation, we
show how to isolate and asymptotically stabilize the relative equilibrium for any choice of constant
momentum. We illustrate all of our results for a network of pendulum/cart systems. For this
example, our approach yields a subtle choice in the coupling variables: the coupling that leads
to stable synchronization is a function of relative positions of pendulum bobs not simply relative
positions of carts. Indeed, coupling as a function of relative cart positions destabilizes the network.
For asymptotic stabilization of the relative equilibrium we assume that the interconnection
graph for the network is connected. However, for asymptotic stabilization of a synchronized motion
on the constant momentum surface, we assume that the interconnection graph for the dissipative

21
 2

1
x 1 , x 2 (rad) 0

−1

−2
0 50 100 150

400

300
θ1 , θ2 (m)

200

100

0
0 50 100 150
time (sec)
10
θ , θ (m/sec)

0
.
2

. −5
1

−10
0 50 100 150
time (sec)

Figure 8.3: Simulation of a controlled network of pendulum/cart systems with dissipation designed
for asymptotic stability of a relative equilibrium (ASSRE). The pendulum angle, cart position and
cart velocity are plotted as a function of time for each of two pendulum/cart systems in the network.

control is completely connected. It is of interest in future work to determine whether this latter
condition can be relaxed.
In Theorem 6.1 we prove asymptotic stabilization of a synchronized motion on the constant
momentum surface; however, we cannot select the value of the momentum – it is determined by the
initial conditions. In Remark 6.2 we propose a control law to simultaneously drive the momentum
to a desired value. This control law appears to work in simulation; however, the stability analysis is
more subtle. It raises a number of interesting questions. For example, suppose we have a dynamical
system depending upon a parameter λ, i.e., the Lagrangian is given by a function L(q, q̇, λ) where
q is the state variable. Assume that for each λ ∈ [0, ], the (controlled) system is Lyapunov stable.
If we now let λ evolve in time such that it “slowly” goes to a value ¯ ∈ (0, ), can we still conclude
that the system is Lyapunov stable in the infinite time domain ? See [12] for results in the case
when the unperturbed system has a uniformly asymptotically stable equilibrium. We plan to build
on these tools to study our parameter dependency problem in future work.
Another future direction is the inclusion of collision avoidance in our framework. For instance,
in our example, the carts move on parallel tracks and hence collision avoidance is not an issue.
However, it is interesting to consider the case in which all of the carts are on the same track and
the pendulum/cart systems should be controlled without collisions for stable synchronization.

22
References
[1] A. M. Bloch, D.E. Chang, N. E. Leonard, and J. E. Marsden. Controlled Lagrangians and the stabiliza-
tion of mechanical systems II: Potential shaping. IEEE Trans. Automatic Control, 46(10):1556–1571,
2001.
[2] A. M. Bloch, N. E. Leonard, and J. E. Marsden. Controlled Lagrangians and the stabilization of
mechanical systems I: The first matching theorem. IEEE Trans. Automatic Control, 45(12):2253–2270,
2000.
[3] F. Bullo. Stabilization of relative equilibria for underactuated systems on Riemannian manifolds. Au-
tomatica, 36(12):1819–1834, 2000.
[4] J.P. Desai, J.P. Ostrowski, and V. Kumar. Modeling and control of formations of nonholonomic mobile
robots. IEEE Trans. Robotics and Automation, 17(6):905–908, 2001.
[5] E Frazzoli. Robust Hybrid Control for Autonomous Vehicle Motion Planning. PhD thesis, MIT, 2001.
[6] E Frazzoli. Maneuver-based motion planning and coordination for multiple uavs. In AIAA/IEEE Digital
Avionics Systems Conference, volume 2, pages 1–12, 2002.
[7] N. H. Getz. Dynamic Inversion of Nonlinear Maps with Applications to Nonlinear Control and Robotics.
PhD thesis, UC Berkeley, 1996.
[8] H. Hanssmann, N. E. Leonard, and T. R. Smith. Symmetry and reduction for coordinated rigid bodies.
European Journal of Control, 12(2):176–194, 2006.
[9] R. A. Horn and C. R. Johnson. Matrix Analysis. Cambridge University Press, 1985.
[10] A. Jadbabaie, J. Lin, and A.S. Morse. Coordination of groups of mobile autonomous agents using
nearest neighbor rules. IEEE Trans. Automatic Control, 48(6):988–1001, 2003.
[11] H. Khalil. Nonlinear Systems. Prentice Hall, Upper Saddle River, New Jersey, 2nd edition, 1996.
[12] H.K. Khalil and P. V. Kokotovic. On stability properties of nonlinear systems with slowly varying
inputs. IEEE Trans. Automatic Control, 36(2):229–229, February 1991.
[13] J. E. Marsden. Lectures on Mechanics. Cambridge University Press, New York, New York, 1992.
[14] S. Nair and N. E. Leonard. Stabilization of a coordinated network of rotating rigid bodies. In Proc.
IEEE Conf. Decision and Control, pages 4690–4695, 2004.
[15] S. Nair, N. E. Leonard., and L. Moreau. Coordinated control of networked mechanical systems with
unstable dynamics. In Proc. IEEE Conf. Decision and Control, pages 550–555, 2003.
[16] P. Ögren, E. Fiorelli, and N. E. Leonard. Cooperative control of mobile sensor networks: Adaptive
gradient climbing in a distributed environment. IEEE Trans. on Automatic Control, 49(8):1292–1302,
2004.
[17] R. Olfati-Saber and R. M. Murray. Graph rigidity and distributed formation stabilization of multi-
vehicle systems. In Proc. IEEE Conf. Decision and Control, pages 2965–2971, 2002.
[18] R. Ortega, M. W. Spong, F. Gómez-Estern, and G. Blankenstein. Stabilization of underactuated
mechanical systems via interconnection and damping assignment. IEEE Trans. Automatic Control,
47(8):1281–1233, 2002.
[19] E. D. Sontag and Y. Wang. On characterizations of the input-to-state stability properties. Systems and
Control Letters, 24(5):351–359, 1995.
[20] M. W. Spong. The control of underactuated mechanical systems. In Proc. 1st International Conf.
Mechatronics, pages 26–29, Mexico City, Mexico, 1994.
[21] H.G. Tanner, G.J. Pappas, and V. Kumar. Leader-to-formation stability. IEEE Trans. Robotics and
Automation, 20(3):443–455, 2004.

23