American Institute of Aeronautics and Astronautics

1
Spacecraft Formation Flying Control using Sliding Mode
and Neural Networks Controller
J onghee Bae
1
, Youdan Kim
2
, and Changook Park
3

Seoul National University, Seoul, 151-742, Korea
This paper proposes a spacecraft formation flying control using sliding mode and neural
networks controller. To deal with the spacecraft formation flying problem, the leader-
follower approach is adopted and the sliding mode controller is designed to efficiently
perform the spacecraft formation flying. The main benefit of sliding mode control is the
robust stability of the closed-loop system. To improve the performance of sliding mode
control, an adaptive controller based on neural networks is used to compensate for the effect
of the modeling error, unknown disturbance, and nonlinearities. The stability analysis of the
closed-loop system is performed using Lyapunov stability theorem. Numerical simulation
results are presented to show the effectiveness of the proposed controller.
Nomenclature
m = mass
J = moment of inertia
x = x-coordinate position
y = y-coordinate position
z = z-coordinate position
q = quaternion
f = control force vector
= control torque vector
= gravitational coefficient
= angular velocity vector
2
J
a = J 2 acceleration
GT
T = gravity gradient torque

I. Introduction
N recent years, formation flight of multiple spacecrafts has become an increasingly attractive area of research.
Instead of a single large spacecraft, the formation flight of multiple small spacecrafts has many advantages; it
provides the improved high resolution, financial benefits, reconfigurability, and robustness compared to a large
single spacecraft. For example, multiple spacecrafts can cover a wider area for the observation of desired objective
such as Earth, planet, asteroid, and so on. When a large single spacecraft has serious problems including hardware
failures and/or software fault, a mission cannot be performed; however, even through one of the multiple spacecrafts
in formation flying has a fault, the mission can be accomplished without a service stop.
1,2

Generally, formation flying controllers are categorized into three approaches: leader-follower approach, virtual
structure approach, and behavioral approach. In the leader-follower approach, a leader spacecraft works as a
manager to offer a reference trajectory, while others follow the leader. Therefore, leader-follower approach is easy
for implementation and stability analysis. Virtual structure approach is similar to the leader-follower approach.
However, the leader is not a real satellite, but a virtual, and the entire formation is assumed as a rigid body. This
approach is easily able to establish a coordinate behavior of the group, but it is difficult to change the formation

1
Graduate Student, School of Mechanical and Aerospace Engineering, jhbae23@snu.ac.kr
2
Professor, School of Mechanical and Aerospace Engineering, ydkim@snu.ac.kr, Senior member AIAA
3
Professor, School of Mechanical and Aerospace Engineering, chanpark@snu.ac.kr, Member AIAA
I
AIAA Guidance, Navigation, and Control Conference
10 - 13 August 2009, Chicago, Illinois
AIAA 2009-5671
Copyright 2009 by the American Institute of Aeronautics and Astronautics, Inc. All rights reserved.

American Institute of Aeronautics and Astronautics


2
pattern for performing specific missions. In the behavioral approach, each spacecraft has some desired objectives
such as formation keeping and goal convergence, and therefore it is easy to change the formation pattern. However,
the stability analysis of the formation controller is more difficult than other approaches.
1-3
In this study, the leader-
follower approach is adopted because it is easy to analyze the stability of the formation controller. This method is
very efficient for spacecraft formation keeping, because the entire system is controlled by the reference information
provided by a leader. Followers only need the command signal from a leader. This makes the whole system simple
to implement the algorithm.
Most spacecraft formation flying controllers have been derived using the simplified relative dynamic equations
such as Hills equation.
4
Hills equation is a linear approximate version of the nonlinear dynamics, and therefore it
does not contain nonlinear high order terms such as J 2 perturbation, atmospheric drag, solar radiation pressure, and
gravity gradient torque. To obtain high-precision spacecraft dynamics, nonlinear dynamic equation should be
considered. In this study, nonlinear spacecraft dynamic equation including J 2 perturbation and gravity gradient
torque is used to design a formation flying controller. Recently, to improve the performance of the spacecraft
formation flight in the presence of uncertainties, many research have been performed to develop formation flying
algorithm utilizing various control theories: LQR,
5
adaptive output feedback,
6
robust control,
7
etc. Also, a sliding
mode controller has been studied by Yeh et al.
8
The sliding mode controller is a systematic controller guaranteeing
the stability of the closed-loop system and provides satisfactory performance in the presence of modeling error and
uncertainties. The main benefit of the sliding mode controller is the robust stability of the closed-loop system with
model uncertainties, disturbance, and nonlinearities. However, sliding mode controller has drawbacks: large control
input requirement, and chattering problem. The performance of sliding mode controller can be improved by
combining adaptive controller such as neural networks.
9

On the other hand, neural networks as adaptive controller have been studied extensively over the decades.
9-13
Neural networks can represent any continuous function through the modification of their synaptic weights by
applying a set of training data which consist of input signals and corresponding output responses. Furthermore,
neural networks have a capability of adaptive synaptic weights with respect to the environment changes in real
time.
14
Therefore, neural networks are widely used to compensate for unknown dynamics and/or unknown
disturbance of the nonlinear systems.
The main contribution of this paper is the design of spacecraft formation flying controller using sliding mode
controller and neural networks in the presence of modeling error and unknown disturbance. The stability of the
proposed control system is analyzed using Lyapunov stability theorem. Note that the sliding mode controller is
designed for the systemwithout modeling error and external disturbance. To compensate for the effects of modeling
error and unknown disturbance, multilayer neural networks are used in this study.
This paper is organized as follows. The spacecraft dynamics and output feedback linearization are briefly
introduced in Section II. The sliding mode controller is described with the assumption that modeling error and
unknown disturbance do not exist in Section III, and the adaptive controller based on neural networks is derived in
the presence of modeling error and unknown disturbance in Section IV. Numerical simulation and analysis to verify
the performance of the proposed controller are described in Section V. Finally, conclusions are presented in Section
VI.

II. Spacecraft Dynamics
In this study, six degree-of-freedom (DOF) spacecraft dynamics model is considered which includes the models
of J 2, gravity gradient torque, and bounded uncertainties.
A. Relative Position Dynamics
A rotating local-vertical-local horizontal (LVLH) frame is used for spacecraft formation flying as shown in Fig.
1, and the 6 DOF spacecraft dynamics model is considered. In Fig. 1, the x-axis points in the radial direction, the z-
axis is perpendicular to the orbital plane and points in the direction of the angular momentum vector, and the y-axis
points the along-track direction. Inertial coordinate system {X, Y, Z} is attached to the center of the Earth, body-
fixed coordinate frame {x, y, z} is attached to the leader spacecraft, and R denotes the position vector from the origin
of the inertial coordinate to the leader spacecraft. The relative dynamics for an eccentric reference orbit can be
represented as

American Institute of Aeronautics and Astronautics


3

2
2
2
2
2 2 2 3/ 2 2
2
2 2 2 3/ 2
2 2 2 3/ 2
( )
2
{( ) }
2
{( ) }
{( ) }
x
J x
y
J y
z
J z
R x f
x y y x a
R x y z R m
f
y
y x x y a
R x y z m
z f
z a
R x y z m

+
= + + + + +
+ + +
= + + +
+ + +
= + +
+ + +

(1)
where
T
r x y z =

denotes position vector of the follower spacecraft, is the latitude angle of the reference
spacecraft, is a gravitational coefficient, and
3
T
x y z
f f f f R =

is the control input vector. If the reference
orbit is circular,

becomes constant.
The J 2 gravity model is given as
15


2
2
2 2
2 2 2 2 2 2 2 2 2 3 2 2 2 5/2
3 15
3
2{( ) } 2{( ) } {( ) }
J E E
z z
a J R x J R z
R x y z R x y z R x y z


=

+ + + + + + + + +

(2)
where x is a unit vector in the radial direction of the follower spacecraft, and z is a unit vector in the z direction of
the earth centered frame (ECF).

























B. Attitude Dynamics in Quaternion Form
Consider the attitude dynamics of a spacecraft as follows

GT
J J T

= + + (3)

0
1

2
T
q q = (4)

Figure 1. Local Vertical Local Horizontal frame

American Institute of Aeronautics and Astronautics


4

1
[ ]
2
q T =

(5)

3 2
3 1
2 1
0
0
0

=



(6)
where
3 3
( ) J J m R

= denotes the moment of inertia,
3 1
R

is a satellite body angular velocity,
3 1
R

is a
control torque,
GT
T is a gravity gradient torque, and
0 3 3
[ ] T q I q

= + . In quaternion kinematics,
4 1
0

T
T
q q q R

=


and 1
T
q q = , where
3 1
q R

is the vector part, and
0
q is the scalar part of the quaternion. Assume that the center of
the mass is the same as the center of the geometry,
GT
T produces the coupling effect of the position and attitude
dynamics as follows
15


2 2 2 3/ 2

3
{( ) }
F F
GT
R J R
T
R x y z

=
+ + +
(7)
where

F
R is the unit vector of the follower position.
Equations (1) and (3)-(5) construct a coupled position and attitude dynamic model of the spacecraft including J 2
and gravity gradient torque effects. Let us define the states and the control inputs as follows:
6 1
1

T
T T
x r q R



,
6 1
2

T
T T
x v q R

, and
6 1
T
T T
u f R



. With the choices of
1 2
, x x , and u, the spacecraft dynamics can be
rearranged as

1 2
2 11 1 2 12 1 2
( , , ) ( , , )
x x
x F x x t F x x t Gu
=
= + +

(8)
where

( )
3 3
11 12 1
1
1
0 0
, ,
1 1
1 [ ] [ ]
0 [ ]
2 2
2
r
GT
F I
m
F F G
T T J J T
T J

= = =

+

(9)
and

2
2
2
2
2 2 2 3/ 2 2
2
2 2 2 3/ 2
2 2 2 3/ 2
( )
2
{( ) }
2
{( ) }
{( ) }
J x
r J y
J z
R x
y y x a
R x y z R
y
F x x y a
R x y z
z
a
R x y z

+
+ + + +

+ + +


= + +

+ + +


+
+ + +

Equation (8) will be used for the controller design and stability analysis in the subsequent section. Note that the
12
F
and G are affected by the modeling error of mass and moment of inertia.



American Institute of Aeronautics and Astronautics


5
C. Output Feedback Linearization
The spacecraft dynamic model considered in the previous section is a special case of general Multi-Input-Multi-
Output (MIMO) nonlinear systems. Output feedback linearization uses full state feedback to linearize the nonlinear
dynamics with chosen output variables.
16
Output vector is differentiated a sufficient number of times, m, until a
control input appears in the resulting equation. If the nonlinear system is output feedback linearizable, then there
exists a linearizability index
i
r . The relative degree of the nonlinear system is defined as

1
m
i
i
r r
=
=

(10)
In this study, let us define an output vector of the system as follows.

1

r
Y x
q

= =


(11)
Differentiating Eq. (11) with respect to time gives

1
[ ]
2
v
r
Y
T q



= =

(12)
Differentiating Eq. (12) with respect to time again yields

( )
3 3
1
1
11 12 1
1 1
[ ] [ ]
2 2
1
0 0
1 1
1 [ ] [ ]
0 [ ]
2 2
2
r
GT
v
Y
T T
F I
m
u
T T J J T
T J
F F Gu

=

+






= + +

+





+ +

(13)
Note that the above systemis output feedback linearizable if
1
det 0 G .
17
This condition can be simplified as
0
0 q .
Let us define the map
14 1 14 1
: R R

to be a diffeomorphism as

1
3
1
2
5
1
4

( ) , ( )
1
[ ]
2
2[ ]
r
q
v
X X
T
T
q

= = =









(14)
Using Eq. (14), the systemcan be transformed to

American Institute of Aeronautics and Astronautics


6

( )
1 3
2 4
3
1 1 1
4 4
1
5 2 4
1
1 1
[ ][ ] [ ] [ ]
2 2

[ ]
r
GT
F f
m
T T T J J T T J
T

=
=
= +
= + + +
=

(15)
Note that the relative degree of the above system is greater than the number of variables, and it has internal
dynamics which is not controlled by the control input using output feedback linearization. Therefore, it is necessary
to verify the stability of the internal dynamics. When
1 4
0 = = = , the system has zero dynamics,
5
0 =

. Let us
consider the Lyapunov function candidate as follows.

2
5
1
2
z
V = (16)
Differentiating Eq. (16) with respect to time gives

5 5
0
z
V = =

(17)
Note from Eq. (17) that the zero dynamics is neutrally stable, and therefore the total systemcan be controlled by the
control input using output feedback linearization.

III. Sliding Mode Controller Design
The control design objective is to select a control input vector u that makes the position r and quaternion q to
track some desired commanded values in the presence of parameter uncertainties.
Let us define a sliding surface as follows.

1 1 1
0
( )
t
d
s e d
dt


= +

(18)
where
1 1 1
( )
d
e t x x = denotes an error vector,
1
d
x is a desired vector of
1
x , and
1
is a positive constant that
determines the bandwidth of the error dynamics.
Differentiating Eq. (18) with respect to time gives

1 2 1 1 1
( )
d
s x x e t = + (19)
Let us consider the following Lyapunov function candidate.

1 1 1
1
2
T
V s s = (20)
Differentiating Eq. (20) with respect to time and substituting Eq. (19) into the resulting equation gives

{ }
1 1 2 1 1 1
T d
V s x x e = +

(21)
Note that
2
x can be considered as a virtual control. The stabilizing controller
2
d
x for the
1
x -subsystem with
2
x as a
virtual input vector is taken as

American Institute of Aeronautics and Astronautics


7

2 1 1 1 1 1
d d
x x e k s = (22)
where
1
k is a positive constant that defines the desired reaching time to the sliding surface. Substituting Eq. (22)
into Eq. (21) gives

1 1 1
0 V k s =

(23)
Let us define the sliding surface for
2
x -subsystem as follows.

2 2 2
0
( )
t
d
s e d
dt


= +

(24)
where
2 2 2
( )
d
e t x x = denotes an error vector,
2
d
x is a desired vector of
2
x , and
2
is a positive constant that
determines the bandwidth of the error dynamics. Differentiating Eq. (24) with respect to time gives

2 11 12 2 2 2
( )
d
s F F Gu x e t = + + + (25)
Now, consider the following Lyapunov function candidate.

2 1 2 2
1
2
T
V V s s = + (26)
Differentiating Eq. (26) with respect to time yields

{ }
2 1 1 2 11 12 2 2 2
( )
T d
V k s s F F Gu x e t = + + + +

(27)
Now, let us propose a control input as follows.

{ }
1
11 12 2 2 2 2 2
( )
d
u G F F x e t k s

= + (28)
where
2
k is a positive constant that defines the desired reaching time to the sliding surface. Substituting Eq. (28)
into Eq. (27) yields

2 1 1 2 2
0 V k s k s =

(29)
Finally, it can be stated that the closed-loop system is globally stable.

IV. Neural Networks Adaptive Controller Design
In the preceding section, the sliding mode controller is designed with the assumption that modeling error and
unknown disturbance do not exist. Generally, sliding mode controller makes the system stay in selected manifold in
state space in the presence of parameter uncertainties. However, the controller has drawbacks such as high gains and
control chattering problem. Also, large parameter uncertainty degrades the performance of sliding mode controller.
To deal with this problem, pure sliding mode controller is combined with an adaptive controller. In this section, the
adaptive controller based on neural networks is used to compensate for the effect of the modeling error and
unknown disturbance.



American Institute of Aeronautics and Astronautics


8
A. Effect of Modeling Error
In this study, the modeling errors in spacecraft dynamics are considered. The estimated values of
12
F and G in
Eq. (8) are defined as
12

F and

G, respectively. Then, the control input can be expressed as follows:

{ }
1
11 12 2 2 2 2 2

( )
d
u G F F x e t k s

= + (30)
Substituting Eq. (30) into Eq. (27) gives

{ }
{ }
2 1 1 2 11 12 2 2 2
1 1 2 2 2
1 1 2 2 2 1
( )

T d
T
T
V k s s F F x e t Gu Gu Gu
k s k s s Gu Gu
k s k s s
= + + + + +
= +

(31)
where
1
Gu Gu represents the termcaused by the modeling error.

B. Neural Networks Structure
Neural networks with one or more hidden layers are referred to as multi-layer perceptron neural networks. Given
an input
nn
x , the three-layer neural networks as shown in Fig. 2 has an output
nn
y as

2 1
0 0 3
1 1
1,2, ,
i k
N N
nn ij jk nn j i
j k
y w v x v w i N
= =

= + + =



(32)
where
jk
v is a weight connecting the first layer to the second layer,
ij
w is a weight connecting the second layer to
the third layer,
0 j
v and
0 i
w are biases, and
i
N is the number of neurons in the i-th layer. In this study, the following
sigmoid activation function ( ) is adapted.

( )
1
( )
1
z
z
e

=
+
(33)



















1
1
1
nn
x
2
nn
x
1
1
N
nn
x

1
N
nn
x
1

2
1 N

2
N

1
nn
y
2
nn
y
3
N
nn
y



Figure 2. Three-layer Neural Network structure

American Institute of Aeronautics and Astronautics


9
Neural networks input-output mapping, Eq. (32), can be expressed in matrix form as follows.

( )
T T
nn nn
y W V x = (34)
where
2 3
( 1) N N
W R
+
,
1 2
( 1) N N
V R
+
,
1
1 N
nn
x R
+
, and
3
N
nn
y R , respectively. Also,
2 2
1
:
N N
R R
+
are defined as
follows:

3 3
3 3
3 3
2 2 2 3 2 2 2 2 3 2
10 20 30 0 10 20 30 0
11 21 31 1 11 21 31 1
12 22 32 2 12 22 32 2
1 2 3 1 2 3
N N
N N
N N
N N N N N N N N N N
w w w w v v v v
w w w w v v v v
W V w w w w v v v v
w w w w v v v v




= =












1 2
1
1
N
T
nn nn nn nn
x x x x

=



2
1 2
( ) 1 ( ) ( ) ( )
T
N
z z z z =



1 2
3
N
T
nn nn nn nn
y y y y

=


Neural networks can approximate any continuous functions
3 1
:
N N
R R to any desired accuracy over a
compact set. This is known as the universal approximation capability.
18
That is, for any given
N
, there exist hidden
layer neurons
2
, , N W and V such that

( )
( ), ( ) in some input space
T T
nn nn nn N nn
W V x x x x = + (35)
The ideal weight matrices W and V satisfying Eq. (35) cannot be determined because no information on the error
term is given. Instead, the estimated weight matrices are used in the controller, and they are adjusted by the adaptive
laws. Let

W and

V be the estimated values of the ideal weight matrices W and V, respectively, and let the weight
deviation or weight estimation errors be defined as

W W W =

and

V V V =

.
Neural networks estimation error can be expressed as follows.

( ) ( ) ( )

' '
T T T T T T T T
nn nn nn nn
W V x W V x W V x W V x = + +

(36)

( ) ( )

' '
T T T T T T
nn nn nn
W V x W V x W V x = (37)
where

' ( )/
T
nn
z V x
d z dz
=
= , and

( )
T
nn
V x = . Furthermore, satisfies the following inequality for some
positive constants , 1,2,3
i
c i = :
13


1 2 3

nn nn
F F
c c x W c x V + + (38)
In this study, it is assumed that the ideal weight matrices are bounded as follows:

American Institute of Aeronautics and Astronautics


10
,
M M
F F
W W V V (39)
where
M
W and
M
V are known positive constants, and
F
denotes the Frobenious norm of matrix.
Differentiating
2
s with respect to time and substituting Eq. (8) and modeling errors into the resulting equation,
the following equation is obtained.

2 11 12 2 2 2

d
s F F Gu x e = + + + + (40)
where the error is defined as

12 12

F F Gu Gu + (41)
The error term can be estimated using a neural network with one hidden layer as follows:

( )
( ), ( )
T T
nn nn nn N
W V x x x = + (42)
Substituting Eq. (42) into Eq. (40) yields

( )
2 11 12 2 2 2

( )
d T T
nn nn
s F F Gu x e W V x x = + + + + + (43)
Now, let us take the stabilizing control input as

( ) { }
1
11 12 2 2 2 2 2

d T T
nn
u G F F x e W V x k s

= + (44)
The estimated weight matrices for the neural networks are used in Eq. (44), because the ideal weight matrices are
unknown. Using Eqs. (36) and (44), the time derivative of
2
s can be obtained as

( ) ( )
( )
2 2 2
2 2

( )

' ' ( )
T T T T
nn nn nn
T T T
nn nn nn
s k s W V x W V x x
k s W V x W V x x


= + +
= + + +

(45)
Now, a neural network weight-tuning algorithm is proposed to guarantee the tracking stability of the closed-loop
system by using the Lyapunov stability theory. Let us consider the following Lyapunov function candidate.

3 1 1 2 2
1 1 1 1
2 2 2 2
T T T T
w v
V s s s s tr W W tr V V


= + + +


(46)
where
w
and
v
are some positive design parameters. Differentiating Eq. (46) with respect to time and substituting
Eq. (45) into the resulting equation gives

( ) { } 3 1 1 2 2 2
1 1

' ' ( )
T T T T T T
nn nn nn
w v
V k s s k s W V x W V x x tr W W tr V V


= + + + + + +




(47)
Let us choose the weight adaptation law as follows:

{ } 2

'
T T
w nn w w
W W V x s k W = = +

(48)

American Institute of Aeronautics and Astronautics


11

2

'
T T
v nn v v
V V x s W k V = = +

(49)
where
w
k and
v
k are some positive design parameters. Using Eqs. (48), (49), and (38), the following inequality
equation is obtained.

3 1 1 2 2 2 1 2 2 2 3 2

T T
N nn nn w v
F F
V k s k s s c s c x W s c x V s k tr W W k tr V V

+ + + + + +


(50)
Since
2 2
2ab a b + ,

W W W =

, and

V V V =

, the following inequalities can be obtained.

2
2
2 2
4
a
a s s + (51)

2
2
2 2
2
2 2 2

4
nn nn
F F
c
c x W s x W s + (52)

2
2
2 2
3
3 2 2

4
nn nn
F F
c
c x V s x V s + (53)

{ }
2

T
w w M
F F
k tr W W k W W W

(54)

{ }
2

T
v w M
F F
k tr V V k V V V

(55)
Substituting Eqs. (51)-(55) into Eq. (50) gives

{ } { }
{ }
( ) ( )
2 2 2 2
2 2
2 2 2 2 2 2
3 1 2
3 1 1 2 2 2 2 2 2
2 2
2 2 2 2
2 2 2
1 1 2 2 1
2 2
2 2
1 1 2

4 4 4 4

2
2 2
2 2 2 2
2
N
nn nn
F F
w M v M
F F F F
w v
nn nn
F F F F
w v w v
M M M M
F F
nn
c c c
V k s k s s s x W s x V s
k W W W k V V V
k k
k s k x W x V s C W V
k k k k
W W V V W V
k s k x

+ + + + + + + +
+ +

= + + +


+ +
+

{ }
2 2 2 2
2 2 2
2 2

2 2
w v
nn
F F F F
k k
W x V s W V C

+ +



(56)
where
2 2 2 2
1 2 3
1
4
N
c c c
C
+ + +
= , and
2 2
2 1
2 2
w v
M M
k k
C C W V = + + .
The second term in Eq. (56) can be made equal to
2
*
2 2
k s by choosing the following gain.

2 2
2 2
*
2 2

2
nn nn
F F
k x W x V k = + + + (57)
where
*
2
0 k > . Substituting Eq. (57) into Eq. (56) gives

2 2
*
3 1 1 2 2 2
2 2
w v
F F
k k
V k s k s W V C +

(58)

American Institute of Aeronautics and Astronautics


12
Let us choose the constant that satisfies the following condition.

*
1 2
0 min , , ,
2 2
w w v v
k k
k k

< <


(59)
Then, the following inequality can be obtained.

3 3 2
2 V V C +

(60)
The above equation implies that
3
0 V

, if
3 2
/ 2 V C > . Therefore, the sliding surfaces
1
s and
2
s and the
weight estimation errors W

and V

are bounded and converge exponentially to the residual set D.

( )
2
2 2
2 2
2
1 2 1 2
1
1
, , , , 1,2
max( , )
j j j j
F F
j w v
C
D s s W V s s W V j

=


= + + + =

(61)
Note that the size of the set D can be made arbitrarily small by adjusting the design parameters
1
k and
2
k , because
1 2 3 1
, , , , ,
M M
c c c C W V and
N
are independent of
1
k and
2
k .
The universal approximation theoremis only guaranteed when the ideal weight matrices and the ideal number of
hidden layer neurons exist. In this study, the weight matrices of neural networks are adjusted by the adaptive laws,
and the size of hidden layer neuron is fixed. The size of hidden layer neuron has influence on the performance of
neural networks: if the size of hidden layer neuron is too small, the neural networks may not compensate for the
effects of modeling error and unknown disturbances correctly. Therefore, the size of hidden layer neuron should be
chosen carefully to consider the complexity and the effects of the modeling errors and unknown disturbance.

V. Numerical Simulation and Performance Analysis
In this section, numerical simulation is performed to evaluate the performance of the proposed sliding mode
controller with neural networks. For spacecraft formation flying, it is assumed that the leader is in circular orbit with
a 42,240 km. The mass of the follower spacecraft is 10.9 kg, and the size is 0.4 0.4 0.6 m m m . Multiple follower
spacecrafts should keep their relative positions with respect to the leader spacecraft, and track their desired target
points with the proposed formation flying controller. In this study, formation flying with one follower spacecraft is
performed to verify the effectiveness of the proposed controller.
Initial position of the follower spacecraft and the reference trajectory of the follower spacecraft are chosen as
follow.

( ) ( )
{ }
{ }
0 0 0
3
3
10 90 50 [ ]
100sin(4 ) 1 exp( 0.05 ) [ ]
100cos(4 ) 1 exp( 0.05 ) [ ]
0 [ ]
d
d
d
x y z m
x t t m
y t t m
z m

=
=
=
=

Figure 3 shows the reference trajectory of the follower. The leader spacecraft is located at ( ) ( ) , , 0,0,0 x y z = . In this
study, a sine wave type bounded disturbance with amplitude of 0.0001 and a frequency of 0.05 rad/s is considered.
This disturbance acts in the x, y, and z axes of the follower spacecraft.







American Institute of Aeronautics and Astronautics


13
















A. Formation Flying with Sliding Mode Controller
Let us consider the ideal case in which modeling error and unknown disturbance are not included in the
simulation. The simulation results of the nominal model are shown in Figs. 4-7. Figure 4 shows the follower
spacecraft trajectory relative to the leader, and Figs. 5-6 show the relative position error and quaternion error. Figure
7 illustrates the history of control input. The total simulation time is 36,000 seconds. The time histories from 0 to
2,000 seconds are described in Figs. 5-7, because responses of the follower do not vary a great deal after 2,000
seconds. Moreover, the follower tracks and approaches its reference trajectory within 1,000 seconds. It is observed
that the trajectory converges to the reference trajectory shortly. In this nominal model case, the controller shows fast
convergence without overshoot and steady-state error.
Now, to evaluate the robustness of designed pure sliding mode controller, let us consider the condition that the
modeling error exists. It is assumed that the m and J contain 10% uncertainties of the nominal value which is
noticeable in the follower spacecraft. Figures 8-11 show the simulation results in the presence of parameter
uncertainty and unknown disturbance as defined earlier. Figures 8-10 illustrate the robust performance of sliding
mode controller. To achieve robustness in the presence of parameter uncertainty and unknown disturbance, the
larger gain and control chattering are required. However, the control gain,
1
k and
2
k , cannot be need an arbitrary
large because of control input limitation and undesired control chattering. In addition, to reduce the control
chattering, the width of the boundary layer has to be increased; however, more expanded width of the boundary
layer causes the large steady-state errors. Therefore, in this case, pure sliding mode controller compromise between
the control chattering and the steady-state error.




















-200
-100
0
100
200
-150
-100
-50
0
50
100
-200
-150
-100
-50
0
50
reference trajectory

Figure 3. Reference trajectory of the follower spacecraft
-200
-100
0
100
200
-150
-100
-50
0
50
100
-200
-150
-100
-50
0
50

trajectory of follower satellite

follower trajectory
initial follower position
leader

Figure 4. Trajectory of the follower spacecraft (nominal)

American Institute of Aeronautics and Astronautics


14
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-6
-4
-2
0
2
relative position error
time, s
x
e
,

m
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-2
0
2
4
6
time, s
y
e
,

m
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-20
0
20
time, s
z
e
,

m
Figure 5. Relative position error (nominal)
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
0
1
attitude error
time, s
q
1
e
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
0
1
time, s
q
2
e
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
0
1
time, s
q
3
e
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
-0.5
0
time, s
q
4
e
Figure 6. Quaternion error (nominal)
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.5
0
0.5
control input
time, s
f1
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.5
0
0.5
time, s
f2
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
-0.5
0
0.5
1
time, s
f3
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.01
-0.005
0
0.005
0.01
time, s
f4
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.01
-0.005
0
0.005
0.01
time, s
f5
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.01
-0.005
0
0.005
0.01
time, s
f6
Figure 7. Control input (nominal)
-200
-100
0
100
200
-150
-100
-50
0
50
100
-200
-150
-100
-50
0
50

trajectory of follower satellite

follower trajectory
initial follower position
leader

Figure 8. Trajectory of the follower spacecraft (SMC)

American Institute of Aeronautics and Astronautics


15

B. Formation Flying with Sliding Mode Controller based on Neural Networks
To deal with the weakness of the pure sliding mode controller, adaptive sliding mode controller based on neural
networks is proposed in Section IV. Simulation results of the proposed controller are shown in Figs. 12-15. Figures
13 and 14 show the history of the relative position error and quaternion error, and Fig. 15 illustrates the control input
history from0 to 2,000 seconds, respectively. The control history shows that the adaptive controller based on neural
networks works well even when modeling error and unknown disturbance exist. As can be observed from the results,
the control input of adaptive controller is smaller and smoother than the pure sliding mode controller, especially near
1,000 seconds when the follower almost approaches its reference trajectory. It can be concluded that the sliding
mode controller based on neural networks perform much better than the pure sliding mode controller because of
compensating the uncertainties.
Table 1 summarizes the performance of the formation controllers. To demonstrate the performance of the
proposed adaptive control law, it is compared with the control law designed by the pure sliding mode control with
two different cases. As shown in Table 1, the performance of adaptive controller based on neural networks is
significantly improved for both cases: (i) the case that only modeling error exists, and (ii) the case that modeling
error and unknown disturbance exist. It is seen that the total relative position error in the adaptive controller is much
smaller than the pure sliding mode controller. In addition, the adaptive controller requires smaller control effort;
therefore, the adaptive controller results in fuel savings as compared with sliding mode controller. This result
confirms that the adaptive controller based on neural networks performs well in the presence of modeling error and
unknown disturbance such as solar radiation pressure, fourth-body gravitation, and atmospheric drag.
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-6
-4
-2
0
2
relative position error
time, s
x
e
,

m
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-2
0
2
4
6
time, s
y
e
,

m
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-20
0
20
time, s
z
e
,

m
Figure 9. Relative position error (SMC)
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
0
1
attitude error
time, s
q
1
e
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
0
1
time, s
q
2
e
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
0
1
time, s
q
3
e
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
-0.5
0
time, s
q
4
e
Figure 10. Quaternion error (SMC)
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.5
0
0.5
control input
time, s
f1
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.5
0
0.5
time, s
f2
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
-0.5
0
0.5
1
time, s
f3
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.01
-0.005
0
0.005
0.01
time, s
f4
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.01
-0.005
0
0.005
0.01
time, s
f5
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.01
-0.005
0
0.005
0.01
time, s
f6

Figure 11. Control input (SMC)

American Institute of Aeronautics and Astronautics


16
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.5
0
0.5
control input
time, s
f1
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.5
0
0.5
time, s
f2
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
-0.5
0
0.5
1
time, s
f3
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.01
-0.005
0
0.005
0.01
time, s
f4
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.01
-0.005
0
0.005
0.01
time, s
f5
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-0.01
-0.005
0
0.005
0.01
time, s
f6
Figure 15. Control input (NN)
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-6
-4
-2
0
2
relative position error
time, s
x
e
,

m
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-2
0
2
4
6
time, s
y
e
,

m
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-20
0
20
time, s
z
e
,

m
Figure 13. Relative position error (NN)
-200
-100
0
100
200
-150
-100
-50
0
50
100
-200
-150
-100
-50
0
50

trajectory of follower satellite

follower trajectory
initial follower position
leader

Figure 12. Trajectory of the follower spacecraft (NN)

0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
0
1
attitude error
time, s
q
1
e
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
0
1
time, s
q
2
e
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
0
1
time, s
q
3
e
0 200 400 600 800 1000 1200 1400 1600 1800 2000
-1
-0.5
0
time, s
q
4
e
Figure 14. Quaternion error (NN)

American Institute of Aeronautics and Astronautics


17



VI. Conclusion
In this study, the pure sliding mode controller and the adaptive sliding mode controller based on neural networks
are proposed for multiple spacecraft formation flying. The leader-follower approach and output feedback
linearization are applied for the coupled nonlinear relative position and attitude dynamics model. Sliding mode
controller has robustness in the presence of uncertainties and unknown disturbance. However, this controller has
drawbacks such as control chattering and large control input requirement. To deal with this problem, adaptive
controller based on neural networks is proposed, which compensates for the effects of the modeling errors and
unknown disturbances. Numerical simulation results show that the proposed adaptive controller has efficient
performance for multiple spacecraft formation flying. The proposed formation flying technique can be applied not
only for the satellite formation flying but also for the multiple vehicle formation operation such as ground vehicles,
mobile robots, and aerial robots.

Acknowledgments
This research has been supported by KARI under the KOMPSAT-3 Development Program that is funded by the
MEST (Ministry of Education, Science, and Technology) of the Republic of Korea.

References
1
Lawton, J . R., Young, B. J ., and Beard R. W., A Decentralized Approach to Elementary Formation Maneuvers, IEEE
International Conference on Robotics and Autonomation, San Francisco, CA, April 2002.
2
Beard, R. W., Lawton, J . R., and Hadaegh, F. Y., A Coordination Architecture for Spacecraft Formation Control, IEEE
Transactions on Control System Technology, Vol. 9, No. 6, 2001, pp. 777-790.
3
Ren, W., and Beard, R. W., Decentralized Scheme for Spacecraft Formation Flying via the Virtual Structure Approach,
Journal of Guidance, Control, and Dynamics, Vol. 27, No. 1, 2004, pp. 73-82.
4
Schaub, H., and J unkins, J . L., Analytical Mechanics of Space Systems, AIAA Education Series, American Institute of
Aeronautics and Astronautics, Reston, VA, 2003.
5
Sparks, A., Satellite Formation Keeping Control in the Presence of Gravity Perturbation, American Control Conference,
Chicago, IL, J une 2000.
6
Wong, H., Kapila, V., and Sparks, A., Adaptive Output Feedback Tracking Control of Multiple Spacecraft, American
Control Conference, Arlington, VA, J une 2001.
Table 1. The performance comparison of proposed controllers

Controller
Total relative position error
(m)
Total control input
Sliding mode controller
(nominal case)
200.3903 14.7919
Sliding mode controller
(modeling error case)
394.2034 18.8329
Adaptive controller based on NN
(modeling error case)
225.0455 13.1095
Sliding mode controller
(modeling error and uncertainties case)
421.4861 19.0225
Adaptive controller based on NN
(modeling error and uncertainties case)
263.2571 12.3296


American Institute of Aeronautics and Astronautics


18
7
Wu, C., and Chen, B., Adaptive Attitude Control of Spacecraft Mixed H2/H Approach, Journal of Guidance, Control,
and Dynamics, Vol. 24, No. 4, 2001, pp. 755-767.
8
Yeh, H., Nelson, E., and Sparks, A., Nonlinear Tracking Control for Satellite Formations, Journal of Guidance, Control,
and Dynamics, Vol. 25, No. 2, 2002, pp. 376-386.
9
Gurfil, P., Idan, M., and Kasdin, N. J ., Adaptive Neural Control of Deep-space Formation Flying, Journal of Guidance,
Control, and Dynamics, Vol. 26, No. 3, 2003, pp. 491-501.
10
Lewis, F. L., Yesildirek, A., and Liu, K., Multilayer Neural-net Robot Controller with Guaranteed Tracking Performance,
IEEE Transactions on Neural Networks, Vol. 7, No. 2, 1996, pp. 388-399.
11
Kim, B. S., and Calise, A. J ., Nonlinear Flight Control using Neural Networks, Journal of Guidance, Control, and
Dynamics, Vol. 20, No. 1, 1997, pp. 26-33.
12
Farrel, J . A., Stability and Approximator Convergence in Nonparametric Nonlinear Adaptive Control, IEEE Transactions
on Neural Networks, Vol. 9, No. 5, 1998, pp. 1008-1020.
13
Lee, T., and Kim, Y., Nonlinear Adaptive Flight Control using Backstepping and Neural Networks Controller, J ournal of
Guidance, Control, and Dynamics, Vol. 24, No. 4, 2001, pp. 675-682.
14
Haykin, S., Neural Networks: A Comprehensive Foundation, Prentice Hall, Upper Saddle River, NJ , 1999, Chap. 1.
15
Xu, Y., Tatsch, A., and Fitz-Coy, N. G., Chattering Free Sliding Mode Control for a 6DOF Formation Flying Mission,
AIAA Guidance, Navigation, and Control Conference, San Francisco, CA, Aug. 2005.
16
Slotine, J . E., and Li, W., Applied Nonlinear Control, Prentice Hall, Englewood Cliff, NJ , 1991.
17
Khalil, H. K., Nonlinear System, 3
rd
edition, Prentice Hall, Upper Saddle River, NJ , 2002.
18
Hornik, K., Stinchcombe, M., and White, H., Multilayer Feedforward Networks are Universal Approximators, Neural
Networks, Vol. 2, No. 5, 1989, pp.359-366.