Systems & Control Letters: Peng Lin, Kaiyu Qin, Zhongkui Li, Wei Ren
Systems & Control Letters: Peng Lin, Kaiyu Qin, Zhongkui Li, Wei Ren
Systems & Control Letters: Peng Lin, Kaiyu Qin, Zhongkui Li, Wei Ren
Corresponding author. Tel.: +1 435 797 2831; fax: +1 435 797 3054.
E-mail address: wei.ren@usu.edu (W. Ren).
finally move on a unit circle. Also, in [20], collective circular
motions were addressed by introducing a cyclic pursuit policy to
make one vehicle pursuit another vehicle along the line of sight
rotated by a common offset angle. Subsequently, the results of [20]
were extended in [21] by introducing a rotation matrix to an
existing second-order consensus protocol. However, in [1821],
the radii of the circles and the desired formation shape cannot be
arbitrarily set. While the circular motions with the same radius or
a priori unknown radii might be appropriate for some applications,
there exist other applications such as persistent surveillance,
where it is desirable to have different desired radii of circular
orbits for different agents. To specify a desired circle center, a
possible approach is to introduce a virtual leader. However, in
real applications and a distributed control context, it might not
be realistic to assume that the virtual leaders state is known by
all agents. Using the algorithms in [18,19], it is not clear how to
introduce a virtual leader whose state is known by only a subset
of the agents to define a desired circle center. Moreover, in [20,21]
a certain control parameter, namely, the rotation angle, must be
exactly equal to a certain value to generate circular motions. As a
result, the algorithms are not robust in this case. Motivated by the
works of [1821], a collective rotating formation control problem
was investigated in [22] for second-order multi-agent systems in
2D. Control protocols were proposed to make all agents surround
a common point with a desired formation structure. However, the
approach adopted in [22] is based on the complex system theory
and cannot be directly applied to the 3D situation.
In this paper, we extend the work of [22] to address the
collective rotating formation control problem for second-order
0167-6911/$ see front matter 2011 Elsevier B.V. All rights reserved.
doi:10.1016/j.sysconle.2011.03.002
366 P. Lin et al. / Systems & Control Letters 60 (2011) 365372
multi-agent systems in a plane in 3D. We are interested in
collection motions where all agents finally converge to a desired
plane in 3D and move on desired circular orbits. This kind of
collective motion is relevant for many practical applications. Due
to the existence of the rotating mode, the desired relative position
between every two agents is time varying and thus the approach
used in the common formation control problems by introducing
desired relative separation vectors to the corresponding consensus
problems is invalid for the rotating formation control problem.
To solve the rotating formation control problem, we propose
two protocols and employ a Lyapunov-based approach to give
sufficient conditions under which all agents finally move around
a common point with a specific formation structure. When there
exists a virtual leader that specifies the desired center and radii
of the orbits, it is only required that one agent knows the
desired circle center and radius. Compared with [1821], the
protocols designed in this paper guarantees that the desired
rotating formation can be arbitrarily set except for some singular
points. Also, the protocols are robust to control parameter changes.
Therefore, the results in this paper complements some of the
existing results [1822].
The following notations will be used throughout this paper.
R
m
and C
m
denote the set of all m dimensional real and complex
column vectors; I
m
denotes the m dimensional unit matrix;
denotes the Kronecker product; 1 represents [1, . . . , 1]
T
with
compatible dimensions (sometimes, we use 1
n
to denote 1 with
a dimension n); 0 denotes the zero vector or zero matrix with
appropriate dimensions; I = {1, 2, . . . , n}; arg() denotes the
argument of a vector; i
j
denotes the imaginary unit; denotes
the 2-norm; denotes the conjugate transpose; diag{ } denotes
a diagonal matrix; denotes the direct sum.
2. Preliminaries
2.1. Graph theory
Let G(V, E, A) be an undirected graph of order n, where V =
{s
1
, . . . , s
n
} is the set of nodes, E V V is the set of edges,
and A = [a
ij
] is the weighted adjacency matrix. The node indices
belong to a finite index set I. An edge of G is denoted by e
ij
=
(s
i
, s
j
). The weighted adjacency matrix is defined as a
ii
= 0 and
a
ij
0, where a
ij
= a
ji
> 0 if and only if e
ij
E. Since the graph
considered is undirected, it means that once e
ij
E, then e
ji
E.
Thus, Ais a symmetric nonnegative matrix. The set of neighbors of
node s
i
is denoted by N
i
= {s
j
V : (s
i
, s
j
) E}. The in-degree
and out-degree of node s
i
are defined as d
in
(s
i
) =
n
j=1
a
ji
and
d
o
(s
i
) =
n
j=1
a
ij
, respectively. Then, the Laplacian corresponding
to the undirected graph G is defined as L = [l
ij
], where l
ii
= d
o
(s
i
)
and l
ij
= a
ij
, i = j. Obviously, the Laplacian of any undirected
graph is symmetric. A path is a sequence of ordered edges of the
form (s
i
1
, s
i
2
), (s
i
2
, s
i
3
), . . . , where i
j
I and s
i
j
V. If there
is a path from every node to every other node, the graph is said
to be connected. If the undirected graph G is connected, then its
Laplacian L has a zero eigenvalue with an associated eigenvector 1
and all its other n 1 eigenvalues are all positive [23].
2.2. Transformation of coordinates
In this subsection, we introduce some concepts and results of
transformation of coordinates (referring to [24]). Consider two
rectangular coordinate systems with a common origin, denoted
by S
a
(x, y, z) and S
b
(, , ). Let i
x
, i
y
, i
z
R
3
and i
, i
, i
R
3
be two sets of orthogonal unit vectors parallel to their respective
coordinate axes. Consider an arbitrary vector p expressed in terms
of the components along the x, y, z axes and the , , axes. Then
we have two equivalent representations: p = xi
x
+ yi
y
+ zi
z
=
i
+ i
+ i
R
3
.
Note that the agents may finally move inthe clockwise direction
or counterclockwise direction. Without loss of generality, we
assume that all agents finally move in the counterclockwise
direction. Moreover, for convenience of discussion, we introduce
a new rectangular coordinate system S
n
such that the third
coordinate axis of S
n
is parallel to the unit vector i
and S
n
and
S
o
share the common origin. The rotation matrix from S
o
to S
n
is
denoted as R
no
R
33
. Here, we do not define the other two axes
of S
n
explicitly and thus the rotation matrix R
no
can take different
values, which does not affect the analysis and results in this paper.
Definition 3.1. A rotating formation in R
3
is a constant vector
F(h(, ), i
)
= [h
T
1
, . . . , h
T
n
, i
T
]
T
= [(
1
cos
1
,
1
sin
1
), . . . , (
n
cos
n
,
n
sin
n
), i
T
]
T
R
2n+3
,
where h
i
= (
i
cos
i
,
i
sin
i
)
T
,
i
[0, 2) and
i
R is
a positive constant for any i I. The multi-agent system (2)
converges to the formation F(h, , i
) if
P. Lin et al. / Systems & Control Letters 60 (2011) 365372 367
Fig. 1. One example of a formation structure with four agents.
lim
t+
i
T
v
i
(t) = 0, (3)
lim
t+
_
v
i
(t) R
T
no
R
l
0
_
2
_
R
no
v
i
(t)
_
= 0, (4)
lim
t+
__
r
i
(t) +
1
R
T
no
R
l
0
_
2
_
R
no
v
i
(t)
_
_
r
k
(t) +
1
R
T
no
R
l
0
_
2
_
R
no
v
k
(t)
__
= 0, (5)
lim
t+
_
1
i
R
T
no
R
l
0
(
i
)R
no
v
i
(t)
1
k
R
T
no
R
l
0
(
k
)R
no
v
k
(t)
_
= 0, (6)
lim
t+
__
_
_
_
v
i
(t)
_
_
_
_
h
i
_
= 0, (7)
for any i, k I and a positive constant , where R
l
0
(
i
) =
_
cos
i
sin
i
0
sin
i
cos
i
0
0 0 1
_
is the rotation matrix of a rotation about the
axis l
0
= [0, 0, 1]
T
with an angle
i
.
1
In particular, if only the
conditions (3)(6) are satisfied, it is said that a quasi-rotating
formation is achieved.
In Definition 3.1, condition (3) means that each agent finally
travels on a plane perpendicular to the unit vector i
2
_
R
no
v
i
(t) and R
T
no
R(
i
)R
no
v
i
(t) are the coordinate
representations of v
i
rotated about the axis i
with an angle =
2
and =
i
, respectively, in S
o
. Then, condition (4) means
that the acceleration of each agent tends to contain only the
component of the centripetal acceleration R
T
no
R
_
2
_
R
no
v
i
(t). By
mechanical knowledge, it can be obtained that conditions (3) and
(4) guarantee that each agent finally moves on a circle with the
angular velocity on a plane perpendicular to the vector i
. Also,
r
i
(t) +
1
R
T
no
R
_
2
_
R
no
v
i
(t) denotes the circle center surrounded
by agent s
i
at time t. Then, condition (5) means that the circle
centers surrounded by all agents tend to be the same as t
+. Condition (6) means that lim
t+
__
_
_
v
i
i
_
_
_
_
_
_
v
k
k
_
_
_
_
= 0 and
lim
t+
[(arg(v
i
) arg(v
k
)) (
i
k
)] = 0 for any i, k I,
which makes all agents finally form a formation that has the same
shape but perhaps a different size from the desired formation h.
Condition (7) guarantees that the final formation has the same size
as the desired formation h. In summary, under conditions (3)(7),
all agents finally surrounda commonpoint inthe counterclockwise
direction with an angular velocity on a plane perpendicular
to the vector i
. Fig. 3
shows one example of two agents, agents s
i
and s
k
, in a desired
rotating formation, where the vectors from the circle center
to the positions of agents s
i
and s
k
are
1
R
T
no
R
_
2
_
R
no
v
i
(t)
and
1
R
T
no
R
_
2
_
R
no
v
k
(t), which correspond to h
i
and h
k
,
respectively.
4. Rotating formation control of second-order multi-agent
systems in 3D
Due to the existence of the rotating mode, the desired relative
position between every two agents is time varying and thus the
approach used in the common formation control problems by in-
troducing desired relative separation vectors to the corresponding
consensus problems is invalid for the rotating formation control
problem.
4.1. Quasi-rotating formation
In this subsection, we study quasi-rotating formation control
to make all agents finally surround a common point on a plane
perpendicular to the vector i
2
_
R
no
v
i
, = [v
T
1
, . . . , v
T
n
, c
T
1
, . . . , c
T
n
]
T
and
=
_
I
n
(R
T
no
R
R
no
) 0
3n3n
0
3n3n
0
3n3n
_
_
_
F[(H
1
LH
1
) I
3
]F
T
L I
3
F
_
(H
1
LH
1
)
_
R
T
no
R
_
2
_
R
no
__
F
T
L
_
R
T
no
R
_
2
_
R
no
_
_
_
R
6n
where H = diag{
1
,
2
, . . . ,
n
}, F = diag{R
T
no
R(
1
)R
no
, . . . , R
T
no
R(
n
)R
no
} and L is the Laplacian of the graph G.
Box I.
where u
i1
= R
T
no
R
R
no
v
i
and u
i2
=
s
k
N
i
a
ij
[
2
i
v
i
1
i
1
k
R
T
no
R(
i
k
)R
no
v
k
]
s
k
N
i
a
ij
[(r
i
r
k
) +R
T
no
R(
2
)R
no
(v
i
v
k
)] for any i I, where a
ij
is the (i, j)th entry of the
adjacency matrix A, N
i
is the neighbor set of agent s
i
and R
=
diag
__
0 1
1 0
_
, 1
_
. In protocol (8), u
i1
is a local velocity feedback
term that plays a role in making each agent surround a point on a
plane perpendicular to the vector i
, while u
i2
is a distributed state
information feedback term that plays a role in making all agents
converge to a specific structure.
We define the equations as in Box I, Here, it should be noted
that R(
i
) = R(
i
)
T
= R(
i
)
1
, R(
i
k
) = R(
i
)R(
k
) and
R(
i
)R(
k
) = R(
k
)R(
i
). Then using protocol (8) for (2), the closed-
loop system can be written in a vector form as
= . (9)
Let
= (I
n
R
no
) = [ v
T
1
, . . . , v
T
n
, c
T
1
, . . . , c
T
n
]
T
R
6n
, where
v
i
= [ v
T
i1
, v
T
i2
, v
T
i3
]
T
R
3
and c
i
= [ c
T
i1
, c
T
i2
, c
T
i3
]
T
R
3
for all i I.
Then system (9) can be written as
, (10)
where
=
_
I
n
R
F[(H
1
LH
1
) I
3
]
F
T
L I
3
F
_
(H
1
LH
1
) R
_
2
__
F
T
L R
_
2
_
_
and
F = diag{R(
1
), R(
2
), . . . , R(
n
)}.
Further, let
1
= [ v
11
, v
12
, . . . , v
n1
, v
n2
, c
11
, c
12
, . . . , c
n1
, c
n2
]
T
R
4n
and
2
= [ v
13
, v
23
, . . . , v
n3
, c
13
, c
23
, . . . , c
n3
]
T
R
2n
.
Also, let
R(
i
) =
_
cos
i
sin
i
sin
i
cos
i
_
and
F = diag{
R(
1
), . . . ,
R(
n
)}.
Here, it should be also noted that
R(
i
) also satisfies that
R(
i
) =
R(
i
)
T
=
R(
i
)
1
,
R(
i
k
) =
R(
i
)
R(
k
) and
R(
i
)
R(
k
) =
R(
k
)
R(
i
). Then system (9) is equivalent to the following two
systems:
1
=
1
1
, (11)
and
2
=
2
2
, (12)
where
1
=
_
_
I
n
R
_
2
_
F[(H
1
LH
1
) I
2
]
F
T
L I
2
F
_
(H
1
LH
1
)
R
_
2
__
F
T
L
R
_
2
_
_
_
and
2
=
_
I
n
H
1
LH
1
L
H
1
LH
1
L
_
.
Remark 4.1. In fact, system (10) or (11)(12) is the representa-
tion of system (9) in the rectangular coordinate system S
n
. Specif-
ically, v
i
, i = 1, 2, . . . , n, are the coordinate representations
of velocities of all agents in S
n
whereas c
i
, i = 1, 2, . . . , n,
are the coordinate representations of the circle centers of all
agents in S
n
. Then it is easy to see that system (9) satisfies con-
ditions (3)(6) if and only if system (10) or system (11)(12) sat-
isfies that lim
t+
_
1
R(
i
)[ v
i1
, v
i2
]
T
R(
k
)[ v
k1
, v
k2
]
T
_
=
0, lim
t+
v
i3
= 0, lim
t+
( c
i
c
k
) = 0andlim
t+
([
v
i1
,
v
i2
]
T
R(
2
)[ v
i1
, v
i2
]
T
) = 0 for any i, k I.
Remark 4.2. It should be noted that the approach used in [22]
cannot be applied directly to analyze the stability of the system
(11) because the multi-agent system in [22] is described in a
complex number field, which has a lower dimension and is thus
much easier to analyze than system (11).
Define
sgn(x) =
_
1, if x = 0
1, if x =
0, otherwise
and
M = diag{1, sgn(|
2
1
|), sgn(|
3
1
|), . . . , sgn(|
n
1
|)}.
Assumption 1. 1
T
n
HM1
n
= 0 when |
i
k
| = 0 or |
i
k
| =
for all i, k I.
Remark 4.3. When 1
T
n
HM1
n
= 0 and |
i
k
| = 0 or |
i
k
| =
hold for all i, k I, then it is easy to see that
n
i=1
h
i
= 0
and all agents in the desired formation are in a line. This might
make the matrix
1
have other imaginary eigenvalues that are not
equal to i
j
and 0. For example, consider a multi-agent system
consisting of two agents with [
1
,
1
,
2
,
2
] = [1, 0, 1, ] and
L =
_
1 1
1 1
_
. By simple calculations, it can be obtained that
1
has imaginary eigenvalues at 2i
j
except 0 and i
j
, which renders
the desired quasi-rotating formation to be unreachable. Therefore,
Assumption 1 is made to eliminate these unreachable formation
structures.
Lemma 4.1. Suppose that the graph G is connected. Under Assump-
tion 1, the following statements hold.
(a)
1
has a zero eigenvalue of multiplicity 2 with associated
eigenvectors [0
T
2n
, 1
T
n
[1, 0]]
T
and [0
T
2n
, 1
T
n
[0, 1]]
T
.
(b)
1
has two simple eigenvalues at i
j
and i
j
with associated
eigenvectors [
T
1
, 0
T
2n
]
T
and [
T
2
, 0
T
2n
]
T
, where
1
=
F[(1
T
n
H)
[i
j
, 1]]
T
and
2
=
F[(1
T
n
H) [i
j
, 1]]
T
.
(c) All its other 4n 4 eigenvalues have negative real parts.
Proof. Let
a
be an eigenvalue of
1
and [z
T
1
, z
T
2
]
T
be its corre-
sponding eigenvector, where z
1
, z
2
C
2n
. Then we have that
_
I
n
R
_
2
__
z
1
F[(H
1
LH
1
) I
2
]
F
T
z
1
(L I
2
)z
2
=
a
z
1
, (13)
F
_
(H
1
LH
1
)
R
_
2
__
F
T
z
1
_
L
R
_
2
__
z
2
=
a
z
2
. (14)
P. Lin et al. / Systems & Control Letters 60 (2011) 365372 369
For statement (a), it can be checked that
1
has a zero
eigenvalue with two associated linearly independent eigenvectors
[z
T
1
, z
T
2
] = [0
T
2n
, (1
T
n
([1, 0]
T
))
T
]
T
and [z
T
1
, z
T
2
] = [0
T
2n
, (1
T
n
([0, 1]
T
))
T
]
T
. To prove that the multiplicity of the zero eigenvalue
is 2, we only need to prove that
1
has no generalized eigenvec-
tor of a grade higher than 1 associated with the zero eigenvalue.
Suppose that there is a grade 2 generalized eigenvector associated
with the zero eigenvalue. That is, there is a nonzero 4n 1 col-
umn vector, denoted by [ z
T
1
, z
T
2
]
T
where z
1
, z
2
R
2n
, such that
(
1
0)[ z
T
1
, z
T
2
]
T
= [z
T
1
, z
T
2
]
T
. After simple calculations, it can be
obtained that z
1
= z
2
and thus (I
n
R(
2
))z
2
F[(H
1
LH
1
)
I
2
]
F
T
z
2
(L I
2
) z
2
= 0, because z
1
= 0
n
and z
2
= 1
n
([1, 0]
T
)
or 1
n
([0, 1]
T
). On the other hand, since
R(
2
) +
R(
2
)
T
= 0, then
F[(H
1
LH
1
)
R(
2
)]
F
T
+[
F[(H
1
LH
1
)
R(
2
)]
F
T
]
T
= 0. There-
fore, z
2
(I
n
R(
2
))[(I
n
R(
2
))z
2
F[(H
1
LH
1
) I
2
]
F
T
z
2
(L I
2
) z
2
] + [z
2
(I
n
R(
2
))[(I
n
R(
2
))z
2
F[(H
1
LH
1
)
I
2
]
F
T
z
2
(L I
2
) z
2
]]
= 2n + z
F[(H
1
LH
1
)
R(
2
)]
F
T
z
2
+
[z
F[(H
1
LH
1
)
R(
2
)]
F
T
z
2
]
= 2n = 0, which is a contradic-
tion. Therefore, the multiplicity of the zero eigenvalue is 2. State-
ment (a) is proved.
For statements (b) and (c), pre-multiplying both sides of (14) by
I
n
R
_
2
_
yields that
_
I
n
R
_
2
__
F
_
(H
1
LH
1
)
R
_
2
__
F
T
z
1
_
I
n
R
_
2
__ _
L
R
_
2
__
z
2
=
F[(H
1
LH
1
) I
2
]
F
T
z
1
+ (L I
2
)z
2
=
a
_
I
n
R
_
2
__
z
2
. (15)
Substituting (15) into (13), we have that
a
_
I
n
R
_
2
__
z
2
=
_
I
n
R
_
2
_
a
I
2n
_
z
1
. (16)
Combining (15) and (16), we have that
K
1
z
1
K
2
z
1
K
3
z
1
= 0, (17)
where K
1
= I
n
R
_
2
_
a
I
2n
, K
2
=
F[(H
1
LH
1
) I
2
]
F
T
and
K
3
=
1
a
(L I
2
)
_
I
n
R
_
2
_
1
_
_
I
n
R
_
2
_
a
I
2n
_
. Consider a
system given by
x = (K
1
K
2
K
3
)x, (18)
where x C
n
. Construct a Lyapunov function for system (18) as
V = x
x. Calculating
V, we have that
V = x
_
K
1
+ K
1
2K
2
+ (L I
2
)
_
I
n
R
_
2
_
1
_
+ (L I
2
)
_
I
n
R
_
2
_
1
_
a
+
2
(L I
2
)
_
x
= x
_
(
a
+
a
)I
2n
2K
2
a
+
2
(L I
2
)
_
x. (19)
Because K
2
and LI
2
are both positive semi-definite, if the real part
of
a
is positive, then
V 0. This implies that all the eigenvalues
of the matrix K
1
K
2
K
3
have negative real parts. Eq. (17) implies
that z
1
should be a zero vector and hence z
2
should also be a zero
vector from (16). Therefore, the eigenvector of
1
associated with
the eigenvalue
a
is a zero vector, which is a contradiction. Thus,
the real part of
a
is nonpositive.
Suppose that
a
= bi
j
(b = 0). Then the derivative of V
becomes
V = 2x
K
2
x 0. Note that Rank(K
2
) = 2n 2 and
K
2
1
= K
2
2
= 0, where
1
=
F[(1
T
n
H) [1, 0]]
T
and
2
=
F[(1
T
n
H) [0, 1]]
T
. Then by LaSalles Invariant Principle, it follows
that the solution of (18) will converge to the kernel space of the
matrix K
2
that is spannedby the vectors
1
and
2
. Also, note that z
1
is an eigenvector of the matrix K
1
K
2
K
3
associated with the zero
eigenvalue from (17). Therefore, the vector z
1
falls into the kernel
space of the matrix K
2
, i.e., K
2
z
1
= 0. Then z
1
can be represented by
a linear combination of
1
and
2
. That is, there exist two numbers
b
1
, b
2
C such that b
1
1
+ b
2
2
= [
1
,
2
][b
1
, b
2
]
T
= z
1
.
We now prove that
1
has two simple eigenvalues at i
j
and
i
j
with associated eigenvectors [
T
1
, 0
T
2n
]
T
and [
T
2
, 0
T
2n
]
T
. From
(16), we have z
2
=
1
a
z
1
_
I
n
R
_
2
__
z
1
. Observing the form
of
1
and
2
, we easily see that
F
_
(H
1
LH
1
)
R
_
2
__
F
T
z
1
=
0 and z
2
can be represented linearly by
1
and
2
since z
1
and
_
I
n
R
_
2
__
z
1
can both be represented linearly by
1
and
2
. It
follows that
_
L
R
_
2
__
z
2
=
a
z
2
from (14). Then as in the
subsequent proof of statement (c), it can be proved that z
2
= 0
when
a
= bi
j
= 0 by a contradiction approach. Thus by simple
calculations, it can be checked that
1
has two eigenvalues at i
j
with associated eigenvectors [
T
1
, 0
T
2n
]
T
and [
T
2
, 0
T
2n
]
T
. We then
prove that the multiplicities of the eigenvalues i
j
are both 1.
Suppose that there is a grade 2 generalized eigenvector associated
with the eigenvalue i
j
. That is, there is a nonzero 4n 1 column
vector, denoted by [ z
T
1
, z
T
2
]
T
, such that
_
I
n
R
_
2
__
z
1
F[(H
1
LH
1
) I
2
]
F
T
z
1
i
j
z
1
(L I
2
) z
2
=
1
, (20)
F
_
(H
1
LH
1
)
R
_
2
__
F
T
z
1
_
L
R
_
2
__
z
2
i
j
z
2
= 0
2n
, (21)
where z
1
, z
2
C
2n
. Pre-multiplying both sides of (20) with I
n
R
_
2
_
, we have
z
1
F
_
(H
1
LH
1
)
R
_
2
__
F
T
z
1
i
j
_
I
n
R
_
2
__
z
1
_
L
R
_
2
__
z
2
=
_
I
n
R
_
2
__
1
.
(22)
Rewriting
1
, we have
1
= [
1
i
j
e
i
j
1
,
1
e
i
j
1
, . . . ,
n
i
j
e
i
j
n
,
n
e
i
j
n
]
T
. Then combining (21) and (22) yields
z
2
= I
n
_
i
j
1
1 i
j
_
z
1
+
1
. (23)
Pre-multiplying both sides of (20) with
1
, we have
1
(L I
2
) z
2
=
1
,
i.e.,
1
(L I
2
)
_
I
n
_
i
j
1
1 i
j
_
z
1
+
1
_
=
1
(L I
2
)
1
=
1
.
Since L is symmetric and positive semi-definite, then
1
(L
I
2
)
1
0. Note that
1
> 0. Then we have
1
(L I
2
)
1
=
1
. It is a contradiction. This proves that the multiplicity of the
eigenvalue i
j
is 1. Similarly, it can be proved that the multiplicity of
the eigenvalue i
j
is 1. Statement (b) is proved.
Next, we prove statement (c). It is easy to see that (K
1
K
3
)z
1
=
_
I
n
R
_
2
_
a
I
2n
_
_
I
2n
+
1
a
_
L
R
_
2
__
_
z
1
= 0. Since
370 P. Lin et al. / Systems & Control Letters 60 (2011) 365372
I
n
R
_
2
_
a
I
2n
is nonsingular for any
a
= i
j
, then it follows
that
_
I
2n
+
1
a
_
L
R
_
2
__
_
z
1
= 0, i.e.,
_
L
R
_
2
__
z
1
=
a
z
1
. It
thus follows that z
1
is also an eigenvector of L
R
_
2
_
associated
with a nonzero eigenvalue and has the form of e ([i
j
, 1]
T
) or
e ([i
j
, 1]
T
), where e = [e
1
, . . . , e
n
]
T
R
n
is an eigenvector of
L associated with a nonzero eigenvalue. If z
1
= e ([i
j
, 1]
T
), then
z
1
= [
1
,
2
][b
1
, b
2
]
T
can be decomposed into
m
R(
m
)[b
1
, 0]
T
+
R(
m
)[0, b
2
]
T
=
m
R(
m
)[b
1
, b
2
]
T
= e
m
[i
j
, 1]
T
for all m I.
Clearly, all the included angles of each pair of
m
[b
1
, b
2
]
T
and
|e
m
|[i
j
, 1]
T
are equal for all m I. Then it follows that
k
=
m
or |
k
m
| = for all k, m I. Moreover, if z
1
= e
([i
j
, 1]
T
), it can be similarly obtained that all
k
should satisfy
that
k
=
m
or |
k
m
| = for all k, m I. Then,
1
and
2
can be written as
1
= [(1
T
n
HM) ([1, 0]
R(
1
)
T
)]
T
and
2
= [(1
T
n
HM) ([0, 1]
R(
1
)
T
)]
T
. Note that L1
n
= 0 and L is a
symmetric matrix. Then 1
T
n
e = 0 and hence 1
T
n
HM1
n
= 0, which
contradicts Assumption 1. Statement (c) is proved.
Lemma 4.2. If the graph G is connected, then lim
t+
v
i3
(t) = 0
and lim
t+
[ c
i3
(t) c
k3
(t)] = 0 for any i, k I. That is, all agents
finally converge to a plane perpendicular to the vector i
.
Proof. Consider system (12). Let
b
be an eigenvalue of
2
and
[y
T
1
, y
T
2
]
T
be its corresponding eigenvector, where y
1
, y
2
R
n
. Then
we have
y
1
H
1
LH
1
y
1
Ly
2
=
b
y
1
, (24)
H
1
LH
1
y
1
Ly
2
=
b
y
2
. (25)
Similar to the proof of Lemma 4.1, it can be proved that
2
has
a simple zero eigenvalue with an associated eigenvector [0
T
n
, 1
T
n
]
T
and all its other 2n 1 eigenvalues have negative real parts.
Then it is easy to see that the solution of system (12) converges
to the eigenvector space, span{[0
T
n
, 1
T
n
]
T
}, of
2
associated with
the zero eigenvalue. This implies that lim
t+
v
i3
(t) = 0 and
lim
t+
[ c
i3
(t) c
k3
(t)] = 0 for any i, k I.
Remark 4.4. As
2
has no imaginary eigenvalues except 0,
Assumption 1 is not required in Lemma 4.2, which is different from
Lemma 4.1.
Theorem 4.5. Consider a network of second-order agents with a
fixed topology. If the graph G is connected, the multi-agent sys-
tem (2) with protocol (8) achieves a desired quasi-rotating formation
F(h(, ), i
) under Assumption 1.
Proof. From Lemma 4.1,
1
has a zero eigenvalue of multiplicity
2 with associated eigenvectors [0
T
2n
, 1
T
n
[1, 0]]
T
and [0
T
2n
, 1
T
n
[0, 1]]
T
, two simple eigenvalues at i
j
and i
j
with associated
eigenvectors [
T
1
, 0
T
2n
]
T
and [
T
2
, 0
T
2n
]
T
and all its other 4n 4
eigenvalues have negative real parts. Also, the initial condition
1
(0) can be decomposed into
1
(0) =
1
[0
T
2n
, 1
T
n
[1, 0]]
T
+
2
[0
T
2n
, 1
T
n
[0, 1]]
T
+
3
[
T
1
, 0
T
2n
]
T
+
3
[
T
2
, 0
T
2n
]
T
+
10
, where
1
,
2
R,
10
R
3n
and
3
,
3
C are conjugate complex
numbers. Then, it is easy to see that
lim
t+
[
1
(t)
1
[0
T
2n
, 1
T
n
[1, 0]]
T
2
[0
T
2n
, 1
T
n
[0, 1]]
T
3
e
i
j
t
[
T
1
, 0
T
2n
]
T
3
e
i
j
t
[
T
2
, 0
T
2n
]
T
] = 0,
i.e.,
lim
t+
[
1
(t)
1
[0
T
2n
, 1
T
n
[1, 0]]
T
2
[0
T
2n
, 1
T
n
[0, 1]]
T
[
F[(1
T
n
H)
T
12
]
T
, 0
T
2n
]
T
] = 0,
where
T
12
= [i
j
(
3
3
)cost (
3
+
3
) sin t, (
3
+
3
) cos t +i
j
(
3
3
)sint] R
2
. This implies that lim
t+
( c
i1
c
k1
) = 0, lim
t+
( c
i2
c
k2
) = 0 and lim
t+
(
1
R(
i
)
[ v
i1
, v
i2
]
T
R(
k
)[ v
k1
, v
k2
]
T
) = 0 for any i, k I. From
Lemma 4.2, we have lim
t+
v
i3
= 0 and lim
t+
( c
i3
c
k3
) =
0 for any i, k I. Moreover, calculating [
v
i1
,
v
i2
]
T
, we have
lim
t+
_
[
v
i1
,
v
i2
]
T
R
_
2
_
[ v
i1
, v
i2
]
T
_
= 0. Therefore, it follows
from Remark 4.1 that the multi-agent system (2) with protocol (8)
achieves a desired quasi-rotating formation F(h(, ), i
).
Remark 4.6. By analyzing the eigenvalueeigenvector solution of
the system as in [2, Theorem 3], the final motions of systems
(11) and (12) can be accurately given in terms of the left
and right eigenvectors of
1
and
2
. Then through coordinate
transformation, the final point, i.e., the common center of the
circles, surroundedby the agents andthe corresponding radii of the
circles can also be calculated. We have omitted these expressions
here due to their complexity.
4.2. Rotating formation
In this subsection, we will study the rotating formation control
problem based on protocol (8) to make all agents finally surround
a common point on a plane perpendicular to the vector i
while
maintaining the desired formation structure h. To achieve this, we
introduce a virtual leader approach.
Suppose that one of the agents, denoted by s
i
0
, has access to
the desired velocity v
0
= R
T
no
[
i
0
cos t,
i
0
sin t, 0]
T
R
3
and the
desired circle center c
0
R
3
. The protocol of agent s
i
0
is given as
u
i
0
= u
i
0
1
+ u
i
0
2
, (26)
where u
i
0
1
= R
T
no
R
R
no
v
i
0
and
u
i
0
2
=
s
k
N
i
0
a
i
0
j
[
2
i
0
v
i
0
1
i
0
1
k
R
T
no
R(
i
0
k
)R
no
v
k
]
s
k
N
i
0
a
i
0
j
_
(r
i
0
r
k
) + R
T
no
R
_
2
_
R
no
(v
i
0
v
k
)
_
(v
i
0
v
0
)
_
r
i
0
+ R
T
no
R
_
2
_
R
no
v
i
0
c
0
_
.
The protocols of all other agents are given as
u
i
= u
i1
+ u
i2
, i I \ {i
0
}, (27)
where u
i1
= R
T
no
R
R
no
v
i
and
u
i2
=
s
k
N
i
a
ij
[
2
i
v
i
1
i
1
k
R
T
no
R(
i
k
)R
no
v
k
]
s
k
N
i
a
ij
_
(r
i
r
k
) + R
T
no
R
_
2
_
R
no
(v
i
v
k
)
_
.
In (26), the term
_
r
i
0
+ R
T
no
R
_
2
_
R
no
v
i
0
c
0
_
plays a role in
making the final point surrounded by all agents converge to c
0
while the term (v
i
0
v
0
) plays a role in guaranteeing that the
final formation has the desired size.
We define the equations as in Box II. It can be easily obtained
that
v
0
= R
T
no
R
[
i
0
cos t,
i
0
sin t, 0]
T
= R
T
no
R
R
no
R
T
no
[
i
0
cos t,
i
0
sin t, 0]
T
= R
T
no
R
R
no
v
0
.
Thus,
v
i
= v
i
i
i
0
R
T
no
R(
i
i
0
)R
no
v
0
= v
i
i
i
0
R
T
no
R(
i
i
0
)R
no
R
T
no
R
R
no
v
0
= v
i
R
T
no
R
R
no
i
i
0
R
T
no
R(
i
i
0
)R
no
v
0
.
P. Lin et al. / Systems & Control Letters 60 (2011) 365372 371
v
i
= v
i
i
i
0
R
T
no
R(
i
i
0
)R
no
v
0
, c
i
= r
i
+ R
T
no
R
_
2
_
R
no
v
i
c
0
,
= [ v
1
, . . . , v
n
, c
1
, . . . , c
n
]
T
,
=
_
I
n
(R
T
no
R
R
no
) 0
3n3n
0
3n3n
0
3n3n
_
_
F[(H
1
LH
1
+ E
1
) I
3
]F
T
(L + E
1
) I
3
F
_
(H
1
LH
1
+ E
1
)
_
R
T
no
R
_
2
_
R
no
__
F
T
(L + E
1
)
_
R
T
no
R
_
2
_
R
no
_
_
R
6n
where E
1
is a diagonal matrix whose i
0
th diagonal entry is 1 and all other entries are 0.
Box II.
Moreover, it is easy to see that
2
i
v
i
1
i
1
k
R
T
no
R(
i
k
)R
no
v
k
=
_
2
i
v
i
1
i
0
R
T
no
R(
i
i
0
)R
no
v
0
_
1
i
1
k
R
T
no
R(
i
k
)R
no
v
k
1
i
0
R
T
no
R(
i
i
0
)R
no
v
0
_
=
2
i
v
i
1
i
1
k
R
T
no
R(
i
k
)R
no
v
k
and
(r
i
r
k
) + R
T
no
R
_
2
_
R
no
(v
i
v
k
)
=
_
r
i
+ R
T
no
R
_
2
_
R
no
v
i
c
0
_
_
r
k
+ R
T
no
R
_
2
_
R
no
v
k
c
0
_
= c
i
c
k
.
Then by simple calculations, using protocol (26)(27) for (2), the
closed-loop system can be written in a vector form as
, (28)
where L is the Laplacian of the graph G. It should be noted that
v
0
(t) = R
T
n
0
R
R
no
v
0
. Clearly, if the graph G is connected, L +E
1
and
L +
2
i
0
E
1
are both symmetric positive definite matrices according
to [5, Lemma 3] and hence H
1
LH
1
+ E
1
is also a symmetric
positive definite matrix. Then by a similar argument to that of the
proofs of Lemma 4.1 and Theorem 4.5, the following lemma and
theorem can be obtained.
Lemma 4.3. If the graph G is connected, then all the eigenvalues of
). Moreover, lim
t+
_
v
i
i
i
0
R
T
no
R(
i
i
0
)R
no
v
0
_
= 0 and lim
t+
_
r
i
+ R
T
no
R
_
2
_
R
no
v
i
c
0
_
= 0 for all i I.
Remark 4.8. Theorem 4.7 shows that protocol (26)(27)makes all
agents finally surround a pre-specified point c
0
with a desired
formation structure h. It should be noted that if we eliminate the
term
_
r
i
0
+ R
T
no
R
_
2
_
R
no
v
i
0
c
0
_
, then the multi-agent system
(2) can also reach the desired rotating formation h but the common
point surrounded by the agents might not be c
0
.
Remark 4.9. In [21], to achieve circular motions, one control gain,
namely, the rotation angle, must be exactly equal to a certain value
to have the eigenvalues on the imaginary axis, which is not robust.
Inthis paper, we do not impose suchanassumptionand the control
parameters
i
,
i
, i = 1, . . . , n, can be arbitrarily chosen in the real
number field except for the points satisfying that 1
T
n
HM1
n
= 0
when |
i
k
| = 0 or |
i
k
| = for all i, k I.
Remark 4.10. In protocols (8) and (26)(27), each agent need
access its neighbors position and velocity information, which
might bring in communication time-delays. Further research could
be directed towards considering the communication time-delays.
1 2
4 3
Ga
Fig. 4. The communication topology for the multi-agent system (2).
Fig. 5. Position trajectories of the multi-agent system using protocol (8).
5. Simulations
Numerical simulations will be given to illustrate the theoreti-
cal results obtained in the previous sections. The graph G
a
in Fig. 4
is the communication topology for the multi-agent system (2),
where the weight of each edge is 1. The initial condition of the
multi-agent system is taken as [r
T
1
, v
T
1
, r
T
2
, v
T
2
, r
T
3
, v
T
3
, r
T
4
, v
T
4
]
T
=
[1, 2, 4, 3, 2, 0, 1, 2, 3, 0, 2, 1, 1, 3, 4, 0, 3, 1, 1,
1, 3, 3, 0, 2]
T
and i
is taken as
_
1
3
,
1
3
,
1
3
_
T
.
We first present simulation results to illustrate Theorem4.5, we
take [
1
,
2
,
3
,
4
] = [0, /3, /3, 0] and [
1
,
2
,
3
,
4
] =
[0.5, 1, 1, 2]. Fig. 5 shows the position trajectories of all agents
with protocol (8). It is obvious that all agents finally surround a
common point with a formation structure that has the same shape
but a different size fromthe desired formation structure on a plane
perpendicular to the vector i
6
1
2
1
6
0
1
3
1
6
1
2
1
3
_
_
_
cos t
sin t
0
_
and the desired circle center
c
0
= [1, 1, 1]
T
. Fig. 6 shows the position trajectories of all
agents with protocol (26)(27). It is clear that all agents finally
372 P. Lin et al. / Systems & Control Letters 60 (2011) 365372
Fig. 6. Position trajectories of the multi-agent system using protocol (26)(27).
surround a common point with the desired formation structure
on a plane perpendicular to the vector i