Ocean Engineering 37 (2010) 1491–1502

Contents lists available at ScienceDirect

Ocean Engineering
journal homepage: www.elsevier.com/locate/oceaneng

Leader–follower formation control of underactuated autonomous

underwater vehicles
Rongxin Cui a,b, Shuzhi Sam Ge b,c,, Bernard Voon Ee How b,c, Yoo Sang Choo b
a
College of Marine Engineering, Northwestern Polytechnical University, Xi’an 710072, China
b
Centre for Offshore Research & Engineering, National University of Singapore, Singapore 117576, Singapore
c
Department of Electrical & Computer Engineering, National University of Singapore, Singapore 117576, Singapore

a r t i c l e in fo abstract

Article history: This paper is concerned with the leader–follower formation control of multiple underactuated
Received 6 May 2009 autonomous underwater vehicles (AUVs). In the proposed leader–follower control, the follower tracks a
Accepted 21 July 2010 reference trajectory based on the leader position and predetermined formation without the need for
Editor-in-Chief: A.I. Incecik
leader’s velocity and dynamics. This is desirable in marine robotics due to weak underwater
Available online 16 September 2010
communication and low bandwidth. A virtual vehicle is constructed such that its trajectory converges
Keywords: to the reference trajectory of the follower. Position tracking control is designed for the follower to track
Autonomous underwater vehicle the virtual vehicle using Lyapunov and backstepping synthesis. Approximation-based control technique
Formation control is employed to handle the model parametric uncertainties and unknown disturbances for the follower.
Leader–follower
The residual error between vehicles within the formation is proven to converge to a bounded compact
Tracking control
set and control performance is guaranteed by suitably choosing the design parameters. Extensive
Neural networks
simulations are provided to demonstrate the effectiveness of the approaches presented.
& 2010 Elsevier Ltd. All rights reserved.

1. Introduction cooperation is formation control, the structure in which AUVs

keep a desired formation configuration and at the same time
Autonomous underwater vehicles (AUVs) play an important complete the assigned tasks.
role in performing underwater tasks, in both civilian and military Various approaches have been proposed for formation control,
applications, such as deep sea inspections, long-distance, long- including behavioral (Balch and Arkin, 1998; Lawton, 2000),
duration surveys, oceanographic mapping and to detect, locate, virtual structure (Lewis and Tan, 1997; Ren and Beard, 2004; Ren
and neutralize undersea mines (Fossen, 1994; Evans and Nahon, and Sorensen, 2008), queues and artificial potential trenches
2004; Li and Lee, 2005; Woolsey and Techy, 2009; Lapierre and (Ge and Fua, 2004; Fua et al., 2007), and leader-following (Wang,
Soetanto, 2007). For specific applications, including ocean sam- 1991; Edwards et al., 2004). In the behavioral approach (Balch and
pling, mapping, minesweeping and ocean floor survey, it is Arkin, 1998; Lawton, 2000), the control action for each vehicle is
beneficial to carry out the missions through multiple AUVs derived by a weighted average of each desired behavior, such as
cooperation for reasons such as increasing efficiency and service formation keeping, goal seeking and obstacle avoidance. This
area, and providing redundancy in case of failure (Skjetne et al., approach allows decentralized implementation and it is natural to
2002; Stilwell and Bishop, 2000; Fiorelli et al., 2006). The deduce control strategies when the robots have multiple
fundamental idea is to use relatively inexpensive, simple, and competing objects. In the virtual structure approach (Lewis and
small AUVs instead of expensive specialized AUV to cooperatively Tan, 1997; Ren and Beard, 2004), the entire formation is treated as
solve difficult or complex underwater tasks. It is expected that the a single rigid body, and the motion of each agent is derived from
cooperative AUVs which forms a mobile network of sensors, can the trajectory of a corresponding point on the structure. In the
outperform a single large vehicle with multiple sensors, when queues and artificial potential trenches approach (Ge and Fua,
the goal is to climb the gradient of an environmental field 2004), the formation structures were presented by queues and
(Ogren et al., 2004). One fundamental problem in multi-AUVs artificial potential trenches, of which the explicit representation
of every single node is not required and the scalability of the
formation is improved when the team size changes. The original
 Corresponding author at: Department of Electrical & Computer Engineering,
scheme (Ge and Fua, 2004) was extended to improve the
National University of Singapore, Singapore 117576, Singapore.
performance of the scheme when only local communication is
Tel.: +65 6516 6821; fax: +65 6779 1103. present, and resulted in a weakly connected network (Fua et al.,
E-mail address: [email protected] (S. Sam Ge). 2007). In the leader–follower approach (Wang, 1991; Edwards

0029-8018/$ - see front matter & 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.oceaneng.2010.07.006
1492 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502

et al., 2004), the leader tracks a predefined path and the follower track the virtual vehicle trajectory (Cui et al., 2009). In addition,
maintains a desired geometric configuration with the leader. The from a practical perspective, we consider that the follower
follower can in turn be designated as a leader for another vehicle parametric uncertainties and unknown disturbances. In this case,
resulting a scalability of the formation. The advantage of this traditional model-based control may not be applicable since they
approach is that specifying a single quantity (the leader’s motion) are generally useful for systems that are linear in parameters,
directs the group behavior. Therefore, it is simple since a reference with exactly known regressors, and time-invariant (Narendra
trajectory is clearly defined by the leader and the internal and Annaswamy, 1989). To overcome the limitations of the
formation stability is induced by the control laws of individual model-based control, approximation-based control is employed
vehicles. to compensate for functional uncertainties and unknown dis-
In addition to the challenges in formation control of mobile turbances from the environment. The contributions of this paper
robots, aircrafts and satellites, formation control of the under- can be summarized as follows:
actuated AUVs has the following issues: (i) the dynamics of AUV
are nonlinear and the hydrodynamic coefficients are difficult to be
determined accurately (Evans and Nahon, 2004); (ii) the ocean (i) The design of a virtual vehicle with a trajectory which
disturbances cannot be ignored in control design for AUVs converges to the reference trajectory of the follower with the
(Fossen, 1994); (iii) information exchange between the AUVs is velocity is guaranteed to be bounded. Uniform semiglobal
difficult due to the weak underwater communication (Schoen- practical asymptotic stability of the closed-loop signals is
wald, 2000); and (iv) since most AUVs are underactuated, stability proved.
of uncontrolled dynamics needs to be considered in control (ii) Through the Property 2 presented in the paper, we change
design. the position tracking error system to a cascaded system, and
A leader–follower formation control algorithm for AUVs was design the tracking controller for the follower via Lyapunov
proposed in Edwards et al. (2004), in which the leader navigates analysis, backstepping and approximate-based control to
the mission waypoints and each follower maintains its place in enable tracking of the virtual vehicle by the follower.
formation using the position of the leader via an exogenous Semiglobal uniform boundedness of the closed-loop signals
system with knowledge on the internal positions of all AUVs. is guaranteed and
Formation control for underwater gliders was designed based on (iii) The stability proof is extended to n vehicles for leader–
virtual bodies and artificial potentials in Fiorelli et al. (2006). follower formation control. The residual error between
A Lagrange approach based formation control of marine surface vehicles within the formation converges to a bounded
vehicle was proposed in Ihle et al. (2006). The backstepping compact set and control performance is guaranteed by
technique was used to solve a geometric task and a dynamic task suitably choosing the design parameters.
for the formation of marine craft in Skjetne et al. (2002), where
the formation is viewed as a flexible system that maneuvers along The remainder of the paper is organized as follows: in Section 2,
a parameterized path. Each individual vehicle has a relative the problem formulation for leader–follower formation control of
position to a point called formation reference point (FRP), and the AUVs is presented. In Section 3, a virtual vehicle is constructed
control is designed according to the two tasks, respectively. using leader position information. Next, the position tracking
Since most AUVs are underactuated, research focusing on path control is designed for the follower in Section 4. The results are
following and formation keeping for underactuated vehicles are extended to multiple leader–follower pairs formation in Section 5.
as follows: nonlinear path following control for an underactuated Simulation results are illustrated in Section 6. Concluding remarks
AUV was proposed in Lapierre and Soetanto (2007), where the are given in Section 7.
explicit AUV model is needed for the control design. State and
output feedback control for an underactuated surface ship path
following at a constant forward speed in the presence of
environmental disturbances was developed in Do and Pan 2. Problem formulation and preliminaries
(2004), the parameters of the AUV model should be known as
well. The cross-tracking control using potential shaping was This section describes the kinematic and dynamic model of the
proposed in Woolsey and Techy (2009). Sliding-mode based AUVs, and provides the formulation of the leader–follower
formation control was investigated for underactuated surface formation control problem.
vessels subject to the ocean disturbance (Fahimi, 2007), where the
ocean disturbances are considered in the control design.
The control method in these works are all model-based, where
2.1. Vehicle modeling
it is assumed that the model parameters are all available for the
control design.
In this paper, we consider leader–follower formation of For the motion description of the AUV, we define a body-fixed
underactuated AUVs in the horizontal plane. The follower frame fBg and a global coordinate frame fEg as shown in Fig. 1. In
acquires the leader position measurement to track the reference the body-fixed frame, the dynamics of an AUV, with nomenclature
trajectory in accordance with the predesigned distance without indicated in Table 1, can be described as (Fossen, 1994)
information on leader velocity and dynamics. This is desirable in Mn_ þ CðnÞn þDðnÞn þ gðZÞ ¼ t þ w ð1Þ
marine robotics due to the weak underwater communication, low
bandwidth and low update rates. It is expected to use less where n and Z are the generalized velocity and position of the
information from other members in the formation compared to AUV, respectively, M, CðnÞ, and DðnÞ are the inertia matrix, the
that on land and in air (Schoenwald, 2000). We divide the matrix of coriolis’ and centripetal terms, and the damping matrix,
formation control into two steps: (i) we design a virtual vehicle respectively, gðZÞ is an unknown vector of restoring forces due to
which uses leader position information, and with a trajectory that buoyancy and gravitational forces and moments, t is the vector of
converges to the reference trajectory of the follower, and the input signals, and w is the vector of external disturbances.
(ii) employing Lyapunov and backstepping synthesis, the model- Without loss of generality, we consider the AUVs formation in
based position tracking control was derived for the follower to the horizonal plane for applications at a constant depth, such as
 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502 1493

with m11 ¼ mXu_ , m22 ¼ mYv_ , m33 ¼ Iz Nr_ , d11 ¼ Xu Xujuj juj,
d22 ¼ Yv Yvjvj jvj, and d33 ¼ Nr Nrjrj jrj (Aguiar and Pascoal,
2007; Lapierre and Soetanto, 2007).
It is obvious that the AUV is an underactuated vehicle as there
is no control input in the sway direction. Position of the AUV in
global coordinate frame fEg can be described as
Z_ ¼ RðcÞn ð2Þ
where the transformation matrix R can be described as
2 3
cosc sinc 0
6 7
RðcÞ ¼ 4 sinc cosc 0 5 ð3Þ
Fig. 1. The underactuated AUV model in horizontal plane. 0 0 1

As the ocean disturbances can be largely attributed to exogenous

Table 1 effects of the environment, which has finite energy, the following
Nomenclature.
assumption is made on the disturbances.
Symbol Description Unit
Assumption 1. For the time-dependent ocean disturbances
x Position in surge m wi ðtÞ, i ¼ 1,2,3, there exist constants w i A R þ , 8t 4 t0 such that
y Position in sway m Jwi ðtÞJ r w i .
c Yaw rad
u Linear velocity in surge m s1
v Linear velocity in sway m s1 2.2. Leader–follower formation
r Yaw velocity rad s  1
t1 Force in surge N In the leader–follower formation considered in this paper,
t2 Moments in yaw Nm
there is one global leader vehicle which does not follow any other
w1 Disturbance in surge N
w2 Disturbance in sway N vehicle, and stable control need to be designed for other vehicles
w3 Disturbance in yaw Nm to keep a predefined distance with respect to its leader. There are
m Weight of the AUV kg many pairs of leaders and followers within the formation and the
L Length of AUV m
team of n vehicles can be decomposed into n  1 decentralized
Iz The moments of inertia in yaw kg m2
Xu_ Added mass in surge kg
subsystems of two vehicles. In each subsystem, the follower AUV
Xu Linear damping coefficients in surge kg s  1 tries to maintain a desired distance and a desired angle relative to
Xujuj Quadratic damping coefficients in kg m  1 its leader. When all the vehicles are in the expected positions, the
surge desired formations are established. We formulate one pair leader–
Yv_ Added mass in sway kg
follower formation problem as follows (Kyrkjebo and Pettersen,
Yv Linear damping coefficients in sway kg s  1
Yvjvj Quadratic damping coefficients in kg m  1 2006): Given the position Z of a leader vehicle, the reference
sway trajectory for the follower is set such that its position is shifted by
Nr_ Added mass in yaw kg m2 a distance d and an angle y relative to the leader. The reference
Nr Linear damping coefficients in yaw kg m2 s  1 trajectory of the follower can be generated as the leader cruises. A
Nrjrj Quadratic damping coefficients in yaw kg m2 s  1
virtual vehicle is designed to track the reference trajectory as
Zm ¼ ½xm ,ym , cm T Generalized position of the leader [m, m, rad]
[m s  1, m s  1, shown in Fig. 2, where the notation ‘virtual/reference’ is used as
nm ¼ ½um ,vm ,rm T Generalized velocity of the leader
rad s  1] the virtual vehicle and reference vehicle is coincide ideally. Next,
Zr ¼ ½xr ,yr , cr T Generalized position of the reference [m, m, rad] we formulate the position tracking control for the follower to
vehicle track the virtual vehicle trajectory.
nr ¼ ½ur ,vr ,rr T Generalized velocity of the reference [m s  1, m s  1,
vehicle rad s  1]
Zv ¼ ½xv ,yv , cv T Generalized position of the virtual [m, m, rad] 2.3. Preliminaries
vehicle
nv ¼ ½uv ,vv ,rv  T Generalized velocity of the virtual [m s  1, m s  1,
vehicle rad s  1] In this subsection, we recall some definitions, lemmas, and
Zf ¼ ½xf ,yf , cf T Generalized position of the follower [m, m, rad] function properties, which are necessary in the design of the
nf ¼ ½uf ,vf ,rf T Generalized velocity of the follower [m s  1, m s  1, proposed control.
rad s  1]
sgnðÞ Signum function
Ai ith AUV in the team
fEg Global coordinate frame
fBg Body-fixed frame

ocean floor survey. Then, n ¼ ½u,v,rT , Z ¼ ½x,y, cT , t ¼ ½t1 ,0, t2 T ,

w ¼ ½w1 ,w2 ,w3 T , gðZÞ ¼ 0,
2 3 2 3
m11 0 0 0 0 m22 v
6 0 m22 7 6
0 5, CðnÞ ¼ 4 0 0 m11 u 7
M¼4 5,
0 0 m33 m22 v m11 u 0
2 3
d11 0 0
6 0 7
DðnÞ ¼ 4 0 d22 5
0 0 d33 Fig. 2. Leader–follower formation.
1494 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502

Definition 1 (Chaillet and Lorı́a, 2008). Uniform semiglobal of I. Then, we have

practical asymptotic stable (USPAS): Consider the following Z aþb Z 1
x ¼ a þ sb
parameterized nonlinear time-varying systems f ða þ bÞf ðaÞ ¼ f uðxÞ dx b f uða þsbÞ ds ð11Þ
a 0
x_ ¼ f ðt, x, yÞ ð4Þ
Define f ðxÞ ¼ sinðxÞ or f ðxÞ ¼ cosðxÞ, we can obtain (9) and (10),
n m
where x A R , t A R Z 0 , y A R is a constant parameter and f ðt, x, yÞ respectively. Moreover, from the fact that sinðxÞ r 1, and
R1 R1
is locally Lipschitz in x and piecewise continuous in t for all cosðxÞ r 1, we have jZs ða,bÞj ¼ j 0 sinða þ sbÞ dsjr 0 jsinða þ sbÞ
y A Rm . R1
j ds r1 and jZc ða,bÞj ¼ j 0 cosða þ sbÞ dsj r 1. This completes the
Let Y  Rm be a set of parameters. System (4) is said to be
proof. &
uniform semiglobal practical asymptotic stable (USPAS) on Y if,

given any D 4 d 4 0, some exists y ðd, DÞ A Y such that Bd is
3. Virtual vehicle design
uniform asymptotic stable (UAS) on BD for the system
x_ ¼ f ðt, x, y Þ, where Bd ¼ fx A Rn : jxj r dg, and BD ¼ fx A Rn :

3.1. Reference vehicle kinematics
jxj r Dg.
In the fEg frame, the position of the reference vehicle can be
Lemma 1 (Panteley and Loria, 1998). Global uniform asymptotic
denoted as
stabling (UGAS) of a cascaded system: Consider two interconnected
subsystems in a cascaded form which can be described as follows: Zr ¼ Zm þRðcm Þl ð12Þ

S1 : x_ 1 ¼ f1 ðt, x1 Þ þ gðt, xÞx2 ð5Þ where Zm is the generalized position of the leader, cm is the yaw
of the leader, l ¼ ½dcosy,dsiny,0T , d is the desired distance
S2 : x_ 2 ¼ f2 ðt, x2 ,uÞ between the leader and the follower, and y is the desired angle
between the xb-axis of the leader and the vector from the leader to
the reference vehicle as shown in Fig. 2. Using the form
where x ¼ ½x1 , x2 T are the states and u is the input, function f1 ðt, x1 Þ is Z_ m ¼ Rðcm Þnm , where nm ¼ ½um ,vm ,rm T is the generalized velocity
continuously differentiable in (t, x1 ) and f2 ðt, x2 ,uÞ, gðt, xÞ are of the leader, the dynamics of Zr can be described as
continuous in their arguments and locally Lipschitz.
x_ r ¼ ðum dy rm Þcoscm ðvm þdx rm Þsincm
The cascaded system S1 and S2 given by (5) is globally uniformly
y_ r ¼ ðum dy rm Þsincm þðvm þ dx rm Þcoscm
asymptotically stable if the following conditions are met: (i) the
system x_ 1 ¼ f1 ðt, x1 Þ is uniformly globally asymptotically stable, (ii) c_ ¼ c_
r m ð13Þ
there exists a control law that globally asymptotically stabilizes the
P
system 2 , and (iii) the function gðt, xÞ satisfies where dx ¼ dcosy and dy ¼ dsiny. Eq. (13) can be collected in a
compact form
Jgðt, xÞJ r W1 ðJx2 JÞ þ W2 ðJx2 JÞJx1 J ð6Þ
Z_ r ¼ Rðcm Þnr ð14Þ
Lemma 2 (Ge and Wang, 2004; Dawson et al., 1990). For bounded where nr ¼ ½um dy rm ,vm þ dx rm ,rm T .
initial conditions, if there exists a C1 continuous and positive definite
Lyapunov function VðxÞ satisfying k1 ðJxJÞ r VðxÞ r k2 ðJxJÞ, such Remark 1. According to (14), it is apparent that by cruising along
that V_ ðxÞ rmVðxÞ þc, where k1 , k2 : Rn -R are class K functions the reference vehicle trajectory with speed nr , which includes the
and c is a positive constant, then the solution x ¼ 0 is uniformly leader velocity um, vm, and rm, the follower will maintain a desired
bounded. distance and angle with respected to the leader. In this paper, we
consider that the leader position Zm is available for control design
In the following, we introduce two useful function properties. of the follower, but the velocity, dynamics, and the control input
of the leader are unavailable to the follower. A virtual vehicle is
Property 1. The function f ðxÞ ¼ logðcoshðxÞÞ has the following introduced in the following subsection to generate the desired
property: trajectory and velocity for the follower as we are precluded from
f_ ðxÞ ¼ x_ tanhðxÞ ð7Þ designing a model-based observer for the leader states.

Proof. Using the fact that ðd=dxÞcoshðxÞ ¼ sinhðxÞ and 3.2. Virtual vehicle design
tanhðxÞ ¼ sinhðxÞ=coshðxÞ, we have
In this section, a virtual vehicle that depends on leader position
1
f_ ðxÞ ¼ sinhðxÞx_ ¼ x_ tanhðxÞ ð8Þ measurement is derived. According to (2), the trajectory of the
coshðxÞ
virtual vehicle has the following form:
This completes the proof. & Z_ v ¼ Rðcv Þnv ð15Þ
Property 2. For all a,b A R, the following two equalities exist Based on the practical considerations, we assume that the velocity
Z 1 of the leader is bounded.
sinða þ bÞ ¼ sina þ b cosða þ sbÞ ds ¼ sina þbZc ða,bÞ ð9Þ
0 Assumption 2. The velocity of the reference vehicle is bounded
as
Z 1
cosða þ bÞ ¼ cosab sinða þsbÞ ds ¼ cosabZs ða,bÞ ð10Þ supJnr J ¼ V M ð16Þ
0 t

with properties jZc ða,bÞj r1 and jZs ða,bÞj r1. with a known bound V M 4 0.

Proof. Let f be a continuously differentiable real-valued function The virtual vehicle tracking error is defined as ev ¼ Zv Zr A R3 . To
defined on an open interval I, and let a as well as a+ b be points achieve the control objective, a filtered tracking error re is defined
 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502 1495

as (Purwar et al., 2008) 4. Position tracking of the follower

re ¼ ev þ f ð17Þ
Using the velocity information from virtual vehicle designed in
T
where f ¼ ½f1 , f2 , f3  , of which the derivative is defined as Section 3, we design the position tracking control for the follower.
The formation task is achieved when the follower track the virtual
f_ ¼ b1 ðfÞKre ð18Þ vehicle position. We transform the kinematic equations of virtual
vehicle to the velocity frame and obtain
where b1 ðfÞ ¼ ½l1 tanhðf1 =l1 Þ, l2 tanhðf2 =l2 Þ, l3 tanhðf3 =l3 ÞT ,
K ¼ diag½k1 ,k2 ,k3 , ki , li A R þ , are design parameters and x_ v ¼ Uv coscW , y_ v ¼ Uv sincW , c _ ¼ rv þ b_ ð26Þ
W v
fi ð0Þ ¼ 0, i ¼ 1,2,3. The proposed control for the virtual vehicle pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
2 2
where Uv ¼ uv þvv is the speed of the virtual vehicle,
is designed as
bv ¼ arctanðvv =uv Þ is the slideslips angle as shown in Fig. 2, and
nv ¼ R1 ðcv Þðb1 ðfÞ þ b2 ðfÞÞ ð19Þ cW is the slideslips angle added to yaw. The following equations
are tenable:
where b2 ðfÞ ¼ ½k1 tanhðf1 =k1 Þ, k2 tanhðf2 =k2 Þ, k3 tanhðf3 =k3 ÞT . qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Taking the time derivative of re, we obtain r ¼ ðxf xv Þ2 þ ðyf yv Þ2
xf xv ¼ rcosðj þ aÞ
r_ e ¼ Kr e Rðcm Þnr þ b2 ðfÞ ð20Þ
yf yv ¼ rsinðj þ aÞ
 
yv yf
j þ a ¼ arctan ð27Þ
xv xf
Theorem 1. Consider the system dynamics described by (18) and
(20) satisfying Assumption 2, with control (19). For any bounded
where r is the length of the vector from the follower to the virtual
initial conditions, all the closed-loop signals are USPAS.
vehicle and a is the angle from xb to the vector. Combined with
(26), the time derivatives of r and a are given by
Proof. Consider the following Lyapunov function candidate
r_ ¼ uf cosavf sina þ Uv cosw ð28Þ
T
Ve ðtÞ ¼ 12reT re þ fch fch ð21Þ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi sina cosa Uv sinw
where fch ¼ ½ k1 logðcoshðf1 =k1 ÞÞ, k2 logðcoshðf2 =k2 ÞÞ, a_ ¼ uf  vf rf  ð29Þ
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi r r r
k3 logðcoshðf3 =k3 ÞÞT . Taking the derivative of (21), using
Property 1, and combining with (19), we have where w is the angle between the resultant speed Uv and
the vector from the follower to the virtual vehicle as shown in
_ T b ðfÞ
V_ e ðtÞ ¼ reT r_ e þK 1 f Fig. 2.
2
The coordinates transformation is only valid for non-zero
_ T b ðfÞ
¼ reT ðKr e Rðcm Þnr þ b2 ðfÞÞ þ K 1 f values of r, as a_ is undefined when r ¼ 0. The above transforma-
2
¼ Kr Te re reT Rðcm Þnr K 1 b1 ðfÞT b2 ðfÞ ð22Þ tion is similar to that introduced in Aicardi et al. (1995), Aguiar
and Pascoal (2007), Hong et al. (2008), and Widyotriatmo et al.
(2010), where it was employed to handle steering and path
It follows that
! following problems for the unicycle-like vehicle (Aicardi et al.,
1 3V M 1995; Hong et al., 2008; Widyotriatmo et al., 2010) and overcome
_
V e ðtÞ r minfki g  Jre J2 K 1 b1 ðfÞT b2 ðfÞ ð23Þ
2 Jre J the stringent condition imposed by Brockett’s results (Aguiar and
Pascoal, 2007). Now, we present the position tracking problem for
Then, letting e be any given positive constant, we choose the gain the follower vehicle and design an appropriate control to satisfy
matrix satisfying the conditions: (i) a converges to zero while t-1; (ii) r
1 3V M converges to d, d 4 0 a predefined arbitrary small positive
minfki g  40 ð24Þ constant; and (iii) all the signals are bounded.
2 e
and obtain the following bound on V_ e :
Jre J2 Z e2 ) V_ e r Jre J2 K 1 b1 ðfÞT b2 ðfÞ ð25Þ 4.1. Kinematic control

Note that Ve is positive definite and radially unbounded. Due to

In this subsection, we design the desired control input uf and rf
the linear dependency of 1=e, based on the results in Kyrkjebo and
for the follower vehicle using the results for the cascaded system.
Pettersen (2006) and Chaillet and Lorı́a (2008), USPAS can be
Since the AUV is underactuated, the uncontrolled sway velocity vf
obtained for the closed-loop signals. &
needs to be considered in the control design.
Now, we design the control following the error dynamics
Remark 2. From (19), JR1 ðcv ÞJ r 1, b1 ðfÞ and b2 ðfÞ are satura-
defined in (28) and (29). From (29), a possible strategy for the
tion functions with design parameters ki and li , which can be
kinematic control design is to regulate rf to make a to 0, and
adjusted to limit the maximum velocity of the virtual vehicle. In
regulate uf to make the follower track the virtual vehicle position.
particular, the maximum speedffi in horizonal plane of the virtual
pffiffiffiffiffiffiffiffiffiffiffiffiffiffi To make a-0, we choose
vehicle is Uv,max ¼ ð u2v þ v2v Þmax ¼ maxi ¼ 1,2 fki g þmaxi ¼ 1,2 fli g,
and the maximum angular velocity is found to be jrv jmax ¼ k3 þ l3 . sina cosa Uv sinw
rf ¼ uf  vf  þ k4 a ð30Þ
r r r
Remark 3. Notice that a quadratic term of fch in the Lyapunov where k4 is a positive constant.
function candidate (21) is introduced to compensate for the term
b2 ðfÞ in the re -dynamics (20) as derived in (22), due to Property 1. Lemma 3. Consider the system (29) with kinematic control (30). For
As such, we obtain that V_ e is negative definite like (25), which any bounded initial conditions, a converges to zero asymptotically
guarantees USPAS of the closed-loop signals. The choice of this and the closed-loop system is globally uniformly asymptotically
quadratic term depends on b2 ðfÞ in the re -dynamics (20). stable.
1496 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502

Proof. Select the Lyapunov function candidate V1 ¼ a2 =2, with Using the inequality ab r la2 =2þ b2 =2l, we obtain 8t A ½t0 ,T with
derivative is given by finite T,
! !
V_ 1 ¼ aa_ ¼ k4 a2 : ð31Þ _ h 2T 1 3k5 2 h 2T 3 h 3T 2
V1r þ þ r þ h 1T þ þ þ vf
2 2 2 2 2 2
Then, a is bounded and limt-1 a ¼ 0. &
2
k5 2 h 3T
The system described by (28) and (29) can be viewed as a þ d þ þ2Uv þw2 u2 =2 r kV1 þc ð40Þ
cascaded system combined with r and a. We design uf based on 2 2
the technique used for cascaded system. Consider the nominal
system of (28) where h 1T :¼ sup0 r T jh1 ðÞj, h 2T :¼ sup0 r T jh2 ðÞj, h 3T :¼ sup0 r T j
r_ ¼ uf þ Uv cosw ð32Þ h3 ðÞj, k :¼ maxfðh 2T þ1 þ 3k5 Þ=2, ð2h 1T þ h 2T þ 3 þh 3T Þ=2g. Since
the speed of the virtual vehicle Uv and the ocean disturbance
and select 2 2
jw2 j are bounded, then, c :¼ k5 d =2 þh 3T þ 2Uv,max þ w2,max u2 =2 is a
uf ¼ k5 ðrdÞ þ Uv cosw ð33Þ finite constant. Thus, for all T Z t0 , r and vf are bounded for all
t A ½t0 ,T. To prove that the same is true for all t Z T, we use the fact
where that a-0 as T-1 and choose T to be sufficient large that
d22 Uv,max
0 o k5 o  and dARþ ð34Þ h 21 sinaðTÞ m1
m11 d k5 cosaðTÞ 4 þ þ z þ a1
2 2 2 1
From (30), rf can be described as h 21 m2 sinaðTÞ
h 11 4 þ h 31 þ þa2 ð41Þ
sina cosa Uv sinw 2 2 2
rf ¼ ½k5 ðrdÞ þUv cosw vf  þk4 a ð35Þ
r r r
where m1 , m2 , a1, a2 are small enough positive constants, h 21 :¼
The closed-loop dynamics of vf can be described as
  supt r T jh2 ðtÞj, h 11 ¼ inf t 4 T jh1 ðtÞj, and z 1 ¼ supt 4 T jk5 rdcosaðTÞ
d22 m11 sina cosa þ Uv ð1cosaðTÞÞcoswðTÞj are bounded positive constants. Using
v_ f ¼  vf  ðk2 r þ Uv cosrÞ ðk5 r þ Uv coswÞ þ k4 a
m22 m22 r r Young’s inequality for (39), we have
d22 m11 ðk5 r þ Uv coswÞ2 sina m11 ðk5 r þ Uv cosfÞcosa z h 31
¼ vf  þ vf V_ 1 r 2minfa1 ,a2 gV1 þ 1 þ , 8t 4 T ð42Þ
m22 m22 r m22 r 2m1 2m2
m11 Uv sinwðk5 r þ Uv coswÞ m11 ðk5 r þ Uv coswÞk4 w It is clear that r and vf are uniformly bounded. &
þ  aþ 2
m22 r m22 m22
Using Lemma 1, we prove that r-d, i.e. r -0 when t-1.
ð36Þ
Using Property 2, we have sina ¼ aZc ð0, aÞ and cosa ¼ 1aZs ð0, aÞ.
It follows that the dynamics of r and a are
where r ¼ rd. Eq. (36) can be described in a compact form as
follows: r_ ¼ ðk5 r þUv coswÞcosavf sina þUv cosw
v_ f ¼ h1 ðÞvf þ h2 ðÞr þ h3 ðÞ þw2u ð37Þ k5 r þ Uv coswð1cosaÞvf sina
¼ k5 r þ½Zs ð0, aÞUv coswZc ð0, aÞvf a ð43Þ
where
   
d22 m11 d Uv
h1 ðÞ ¼  k5 1 cosa þ coswcosa a_ ¼ k4 a ð44Þ
m22 m22 r r
m11 2
h2 ðÞ ¼  ðk sina þ k4 k5 aÞ
m22 5
("   #
m11 d2 d U 2 cos2 w Remark 4. The use of Property 2 allows us to the formulation of a
h3 ðÞ ¼  2k25 d þk25 þ 2k5 Uv 1 cosw þ v sina
m22 r r r conjunction item in (43), which shows the boundedness of the
  conjunction item directly.
d U 2 sinwcosw
k5 Ur sinw 1 k4 ðk5 dUv coswÞa v
r r Theorem 2. Consider the system dynamics described by (28) and
(29) with control (33) and (35). For all bounded initial conditions, r
and a converge to zero asymptotically, and the closed-loop system is
w2u ¼ w2 =m22 UGAS.

Proof. With the control (33) and (35), (28) and (29) have been
Lemma 4. Consider the system dynamics described by (28) and (29) changed to (43) and (44). From Lemma 3, we find that (44) is
with the kinematic control (33) and (35). For any bounded initial uniformly asymptotically stable. According to Lemma 1, we need
conditions, r and vf are uniformly bounded. to prove the conjunction item gðÞ ¼ ½Zs ð0, aÞUv cosfZc ð0, aÞvf 
satisfies (6). Since Uv is bounded and vf is bounded according to
Proof. Consider the Lyapunov function candidate Lemma 4, then Jgðt, aÞJ r W, where W is a positive constant. Then,
according to Lemma 1, the system described by (28) and (29) is
V1 ðr,vf Þ ¼ 12 r2 þ 12v2f ð38Þ
globally uniformly asymptotically stable. &
The derivative of V1 is given by
4.2. Dynamic control
V_ 1 ðr,vf Þ ¼ rf½k5 ðrdÞ þ Uv coswcosa
In this section, the kinematic control is extended to the
vf sina þ Uv coswg þ vf ½h1 ðÞvf þ h2 ðÞr þ h3 ðÞ þw2 u
dynamic case, and we address a challenging problem where
¼ k5 cosar2 h1 ðÞv2f þ ½h2 ðÞsinarvf parametric uncertainties exist in the dynamics of the follower
þh3 ðÞvf þ k5 rdcosa þ Uv ð1cosaÞcosw þw2 uvf ð39Þ subjected to unknown ocean disturbances. In this case, traditional
 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502 1497

model-based control is not suitable as parametric certainty is Since the parameters m11 ,m22 ,m33 ,d11 ,d33 ,w 1 , and w 3 are un-
required. The combination of approximation-based and domina- known, the model-based control (50) is not feasible. Function
tion design techniques allows us to handle time-varying dis- approximators are used to handle the unknown entities in the
turbances. Using the backstepping and Lyapunov synthesis, a control as follows:
stable tracking control is first designed. Subsequently, function  T
sina ^ T SðZÞ
approximators are employed to compensate the unknown items t ¼ diag½k6 ,k7 z þ  ,a þ W
r
in the control.
_^ ^
W i ¼ Gi ðSi ðZÞzi þ si W i Þ ð52Þ
4.2.1. Function approximators
There are many types of function approximators, neural ^ T ,W
^ :¼ blockdiag½W
where t ¼ ½t1 , t2 T , z ¼ ½z1 ,z2 T , W ^ T  contains
1 2
networks (NN), fuzzy systems, and wavelet networks (Ge et al., T T T
the approximation parameters, SðZÞ ¼ ½S1 ðZÞ,S2 ðZÞ are the basis
2002). In this paper, linearly parametrized NN is used to functions, and si are positive constants. The neural network
approximate the continuous function f ðZÞ : Rq -R, W^ T SðZÞ approximates W*TS(Z) defined by
" #
f ðZÞ ¼ W T SðZÞ þ eðZÞ ð45Þ m22 vf rf þd11 uf þm11 a_ u sgnðz1 Þw 1
W T SðZÞ ¼ eðZÞ
T q
where the input vector Z ¼ ½z1 ,z2 , . . . ,zq  A R , weight vector ðm11 m22 Þuf vf þd33 rf þ m33 a_ r sgnðz2 Þw 3
W A Rl , the NN node number l 41, and SðZÞ A Rl . Universal ð53Þ
approximation results indicate that, if l is chosen sufficiently 2
where eðZÞ A R is the approximation error, Z ¼ ½uf ,vf ,rf , au ,
large, WTS(Z) can approximate any continuous function, f(Z), to
any desired accuracy over a compact set Z A OZ to arbitrary any
ar , a_ u , a_ r T are the input variables. Consider the augmented
Lyapunov function candidate
degree of accuracy as
f ðZÞ ¼ W T SðZÞ þ e ðZÞ, 8Z A Oz  Rq 1 2 1 2 1 1 1X2
~ T G1 W
~ i
ð46Þ V3 ðtÞ ¼ r þ a þ m11 z21 þ m33 z22 þ W ð54Þ
2 2 2 2 2i¼1 i i
where W* is the ideal constant weight vector, and eðZÞ is the
approximation error which is bounded over the compact set, i.e., where W ~ i :¼ W
^ i W  . Its derivative is given by
i
jeðZÞj r e , 8Z A OZ with e 40 is an unknown constant. The ideal     X2
W 2 W ~ TW
weight vector W* is an artificial quantity required for analytical V_ 3 ðtÞ r  k5  r  k4  a2 k6 z21 k7 z22 þzT e si W i
^ i
2 2
purposes. W* is defined as the value of W that minimizes jej for all i¼1

Z A OZ  Rq , i.e., ð55Þ
( )
Using Young’s inequality, we have
W  :¼ arg min sup jf ðZÞW T SðZÞj ð47Þ
W A Rl Z A OZ si W ~ i r  si JW ~ i J2 þ si JW  J2 ,
i
2 2
2 2
JzJ Je J
4.2.2. Control design zT e r þ ð56Þ
2 2
Following the backstepping methodology, let uf and rf in (33)
and (35) be the virtual control inputs, respectively, and au and ar
where e is an unknown constant.
be the corresponding virtual control. Define the error variables as
Accordingly, we have
z1 ¼ uf au , z2 ¼ rf ar ð48Þ        
W 2 W 1 1
V_ 3 ðtÞ r  k5  r  k4  a2  k6  z21  k7  z22
Consider the following Lyapunov function candidate: 2 2 2 2
V2 ðtÞ ¼ 12 r 2 þ 12 a2 þ 12 m11 z21 þ 12m33 z22 si ~ 2 si Je J2
 JW i J þ JWi J2 þ ð57Þ
2 2 2
The derivative of V2(t) can be described as
sina
V_ 2 ¼ k5 r 2 þ gðÞrak4 a2 þ az1 az2 V_ 3 r mV3 þ C ð58Þ
r
_ a_ u Þ þ z2 m33 ðr_ a_ r Þ
þ z1 m11 ðu (       !)
W W 2 k6  12 2 k7  12 si
¼ k5 r 2 þgðÞrak4 a2 m :¼ min 2 k5  ,2 k4  ,
2 2 m11
,
m33
, min
i ¼ 1,2 lmax ðG1 Þ
  i
sina
þ z1 m22 vf rf d11 uf þ t1 þw1 m11 a_ u þ a ð59Þ
r
þ z2 ½ðm11 m22 Þuf vf d33 rf þ t2 þ w3 am33 a_ r  ð49Þ si 1
C :¼ JWi J2 þ Je J2 ð60Þ
2 2
where gðÞ ¼ ½Zs ð0, aÞUv coswZc ð0, aÞvf .
We design the control law for the follower as follows:
Theorem 3. Consider the follower AUV with dynamics (1) and
sina
t1 ¼ m22 vf rf þ d11 uf þ m11 a_ u  ak6 z1 sgnðz1 Þw 1 satisfying Assumptions 1 and 2, under the action of the control law
r (52). For each compact set O0 , where ðZf ð0Þ, nf
t2 ¼ ðm11 m22 Þuf vf þd33 rf þ a þ m33 a_ r k7 z2 sgnðz2 Þw 3 ð50Þ ^ 1 ð0Þ, W
ð0Þ, W ^ 2 ð0ÞÞ A O0 , the trajectories of the position tracking
closed-loop system are semiglobally uniformly bounded. Thep tracking
ffiffiffiffiffiffiffiffiffiffiffiffi
where k6 and k7 A R þ , and obtain error ½r , a converges to a compact set Ozs :¼ fJ½r , aJ r 2C=mg,
    where m and C are defined in (59) and (60), respectively.
W 2 W
V_ 2 r k5  r  k4  a2 k6 z21 k7 z22
2 2 Proof. From (58) and Lemma 2, it is clear that the signals
z1 sgnðz1 Þw 1 z2 sgnðz2 Þw 3 ~ 1 , and W
r , a,z1 ,z2 , W ~ 2 are semiglobally uniformly bounded. With
   
W 2 W W*i as a constant, we know that W ^ i is also bounded, for i ¼ 1,2.
r  k5  r  k4  a2 k6 z21 k7 z22 ð51Þ
2 2 Therefore, all signals are bounded. It is straightforward to prove
1498 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502

that ½r , a converges to Ozs using a similar approach as that found n1 

X 
si,j  2 1
in Ge and Wang (2004). & C :¼ JWi,j J þ Je J2 ð65Þ
j¼1
2 2 j

Similar to the proof of Theorem 3, we conclude that all signals are

5. Extension to n vehicles formation bounded and the formation is stable. The tracking error
P
between all leader–follower pairs ½r , a  with r 2 ¼ nj¼ 2 r 2j and
In this section, we extend our results to leader–follower P
formation of n vehicles. Let A ¼ fAi ji A I g be a set of n AUVs, where a 2 ¼ n1
j ¼ 1 aj converge to the compact set O zs :¼ fJ½r , a J r
2  
qffiffiffiffiffiffiffiffiffiffiffiffi
I ¼ f0,2, . . . ,n1g. A directed graph G is used to model the
2C =m g. &
interaction topology among the AUVs. In G, the ith node
represents the ith AUV Ai and a directed edge from Ai to Aj
denoted as (Ai,Aj) represents a directional information exchange Remark 5. From Theorem 4, we know that as n increases, the size
link from Ai to Aj, i.e., Ai is the leader of Aj. Fig. 3 shows two typical of the compact set O zs increases and remains bounded as long as n
leader–follower formation structures where A0 is denoted as the is finite. To reduce the residual error, the control parameters
global leader. In Fig. 3(a), A1 and A2 are the followers of A0, and A3 k4,j ,k5,j ,k6,j ,k7,j , si,j and Gi,j can be regulated such that the size of
and A4 are the followers of A1 and A2, respectively, etc. In Fig. 3(b), O zs is reduced. An alternative is to increase the density of
Ai is the follower of Ai  1. For the leader–follower formation approximators that span the approximation space, thus reducing
discussed in this paper, the formation graph can be viewed as a the approximation error ej that makes up part of the C .
tree with root A0.
Remark 6. In order to prevent collisions between the AUVs: (i)
Theorem 4. Consider leader–follower formation of n AUVs, each for each leader–follower pair, the distance between them need to
AUV Aj , j ¼ 1, . . . ,n1 with dynamics (1) using the control (52) to
regulate its position relative to its leader, for each compact set O0,j ,
where ðZj ð0Þ, nj ð0Þ, W ^ 1,j ð0Þ, W^ 2:j ð0ÞÞ A O0,j , the trajectories of the
closed-loop system are semiglobally uniformly bounded. The tracking
error of the formation is bounded and converges to a compact set 250
qffiffiffiffiffiffiffiffiffiffiffiffi
O zs :¼ fJ½r , a J r 2C =m g,where m and C are defined in (64) and
(65), respectively. 200

Proof. Select the Lyaponov function candidate

150
)
y [m]

X
n1 X
n
1 2 1 2 1 1 1X 2
VF ðtÞ ¼ V3,j ðtÞ ¼ r þ a þ m z2 þ m z2 þ ~ T G1 W
W ~ i,j
2 j 2 j 2 11,j 1,j 2 33,j 2,j 2 i ¼ 1 i,j i,j
100
j¼1 j¼2

ð61Þ
where subscript j denote the jth AUV. Derivation of V(t) is given by 50
X
n1      
Wj 2 W 1
V_ F ðtÞ r  k5,j  r j  k4,j  j a2j  k6,j  z21,j
j¼1
2 2 2 0
  )
1 2 si,j ~ 2 si,j  2 Je j J2
 k7,j  z  JW i,j J þ JWi,j J þ ð62Þ
2 2,j 2 2 2 −150 −100 −50 0 50 100 150 200
x [m]
Then we have Fig. 4. Trajectories of the AUVs.

V_ F r m VF þ C ð63Þ
   
Wj Wj 12
m :¼ min 2 k5,j  ,2 k4,j  ,
j 2 2 2
  !)
2 k6,j  12 2 k7,j  12 si,j 10
, , min ð64Þ
Tracking error of each AUV [m]

m11,j m33,j i ¼ 1,2 lmax ðG1 Þ

i,j 1.5
8
1
6
100 200 300 400 500
Level 4

Level 2

0
Level
0 100 200 300 400 500
time [s]
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
Fig. 3. Leader–follower formation structure. Fig. 5. Trajectory tracking errors of the AUVs: ri ¼ ðxf ,i xv,i Þ2 þ ðyf ,i yv,i Þ2 .
 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502 1499

pffiffiffiffiffiffiffiffiffiffiffiffi
satisfy d 4 2C=m þ L, where L is the length of the AUV and (ii) for also serves as the leader of A5. A4 follows A2 and serves as the
AUVs at the same level in formation, for example in Fig. 3(a), A0, leader of A6 as well. The desired distance between each pair of
A1 and A2 form a triangular formation where
pffiffiffiffiffiffiffiffiffiffiffi
ffi A0 is the leader, A1 the leader–follower is 14 m. For the leader–follower pair in the
and A2 need to satisfy jZ1 Z2 j 4 2 2C=m þL. The methodologies left, such as (A0,A1), (A1,A3), the desired relative angle is 2p=3. For
to address this condition as part of the trajectory planning will be the leader–follower pair in the right, such as (A0,A2), (A2,A4),
addressed in the future work. the desired relative angle is 4p=3. Initial conditions of the global
leader A0 are Z0 ð0Þ ¼ 0 and n0 ð0Þ ¼ 0. Initial conditions of other
6. Simulation results AUVs are Z1 ð0Þ ¼ ½13,5,0T , Z2 ð0Þ ¼ ½15,10,0T , Z3 ð0Þ ¼ ½21,
14,0T , Z4 ð0Þ ¼ ½18,27,0T , Z5 ð0Þ ¼ ½32,34,0T , and Z6 ð0Þ ¼
To illustrate the performance of the control scheme proposed ½35,34,0T . The initial velocities of all the AUVs are ni ¼ 0,
for leader–follower formation control of underactuated AUVs, the i ¼ 1, . . . ,6.
dynamics of underactuated vehicles are adapted from Dong and Control parameters and gains are selected as: l1 ¼ 2, l2 ¼ 1,
Guo (2005): m11 ¼ 200 kg, m22 ¼ 250 kg, m33 ¼ 80 kg, d11 ¼ ð70 þ l3 ¼ 0:5 k1 ¼ 2, k2 ¼ 1, k3 ¼ 0:1, k4 ¼ 0:1, k5 ¼ 0:1, k6 ¼ 300,
100jujÞ kg=s, d22 ¼ ð100 þ 200jvjÞ kg=s, and d33 ¼ ð50 þ100jrjÞ kg=s. k7 ¼ 150 and d ¼ 0:5 m. The simulation time range is 500 s. In
We use 7 AUVs to form a wedge formation, which is similar to the formation, the global leader A0 cruises in a straight trajectory
Fig. 3(a), in the simulation, where A0 is the global leader, A1 and A2 during the first 10 s with surge speed 1.0 m/s, and then rotates at
are its left and right follower, respectively. A3 followers A1 and it the same speed with the yaw speed 0.3 rad/s as shown in Fig. 4.

0.5
Norm of NN weights for AUV 1

Norm of NN weights for AUV 2

0.6
0.4
0.5
0.3 0.4

0.2 0.3
0.2
0.1
0.1
0 0
0 100 200 300 400 500 0 100 200 300 400 500
time [s] time [s]

0.4
Norm of NN weights for AUV 3

Norm of NN weights for AUV 4

0.35 0.5
0.3
0.4
0.25
0.2 0.3
0.15 0.2
0.1
0.1
0.05
0 0
0 100 200 300 400 500 0 100 200 300 400 500
time [s] time [s]

0.45 0.8
Norm of NN weights for AUV 5

Norm of NN weights for AUV 6

0.4 0.7
0.35 0.6
0.3
0.5
0.25
0.4
0.2
0.3
0.15
0.1 0.2
0.05 0.1
0 0
0 100 200 300 400 500 0 100 200 300 400 500
time [s] time [s]
^ 1 J and JW
Fig. 6. Norm of neural weights of each AUV, JW ^ 2 J. (a) Norm of neural weights of AUV 1. (b) Norm of neural weights of AUV 2. (c) Norm of neural weights of
AUV 3. (d) Norm of neural weights of AUV 4. (e) Norm of neural weights of AUV 5. (f) Norm of neural weights of AUV 6.
1500 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502

Without loss of generality, we define the disturbance as time- where $j is one of the three functions from the jth set
varying forces/moment in frame fEg as 2
2 3 f1=ð1 þ ea1j ðZj þ b1j Þ Þ,ea2j jZj b2j j ,1=ð1 þ eða3j ðZj b3j ÞÞ Þg
10 þ 1:8sinð0:7tÞ þ 1:2sinð0:05tÞ þ 1:2sinð0:9tÞ
6 7 such that each is composed of a unique combination of the
fc ðtÞ ¼ 4 5þ 0:4sinð0:1tÞ þ0:2cosð0:6tÞ 5
0 functions from the 7 sets. The constant parameters akj and bkj,
k ¼ 1, 2, 3 are user-defined. We choose Gi ¼ 10, si ¼ 0:1, i ¼ 1,2
Then, in the body frame fBg, the disturbance acting on the AUV is with control signals saturated whenever jt1 j Z350 N,
jt2 j Z 80 Nm.
wðtÞ ¼ RT ðcÞfc ðtÞ
Simulation results are shown in Figs. 4–8. The trajectories of all
where wðtÞ ¼ ½w1 ðtÞ,w2 ðtÞ,w3 ðtÞT . In the simulation, linearly the AUVs are shown in Fig. 4 and the tracking errors of the AUVs
^ :¼ are shown in Fig. 5. It can be observed that the formation tracking
parameterized function approximators are used, in which W
performance of the AUVs is satisfactory, despite the time-varying
^ T ,W
blockdiag½W ^ T  are adaptable parameters, and Z ¼ ½uf ,vf ,rf , au , disturbance. Fig. 5 verifies that the initial tracking errors of all
1 2
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
ar , a_ u , a_ r  A R7 are the inputs to the approximator. Motivated by AUVs, ri ¼ ðxf ,i xv,i Þ2 þ ðyf ,i yv:i Þ2 , are sufficiently reduced to
Tee and Ge (2006), the basis functions are defined by
small values and are bounded.
m Y
7 The norms of the NN weights are bounded with slight
i ðZÞ
Si ðZÞ :¼ P
37
, mi :¼ $j , j ¼ 1, . . . ,37 ð66Þ oscillations as shown in Fig. 6. The boundedness of control inputs
m
j ¼ 1 j ðZÞ j¼1 are shown in Figs. 7 and 8.

400 350
350 300
300 250
250
200
τ1,1

τ1,2

200
150
150
100 100

50 50
0 0
0 100 200 300 400 500 0 100 200 300 400 500
time [s] time [s]

300 400
350
250
300
200
250
τ1,3

τ1,4

150 200
150
100
100
50
50
0 0
0 100 200 300 400 500 0 100 200 300 400 500
time [s] time [s]

500 500
450
400
400
300 350
τ1,5

τ1,6

200 300
250
100
200
0 150
0 100 200 300 400 500 0 100 200 300 400 500
time [s] time [s]

Fig. 7. Control input t1 of each AUV. (a) Control input t1 of AUV 1. (b) Control input t1 of AUV 2. (c) Control input t1 of AUV 3. (d) Control input t1 of AUV 4. (e) Control
input t1 of AUV 5. (f) Control input t1 of AUV 6.
 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502 1501

40 40

30 30

20 20

τ2,1

τ2,2
10 10

0 0

−10 −10

−20 −20
0 100 200 300 400 500 0 100 200 300 400 500
time [s] time [s]

40 80

30 60
20
40
τ2,3

τ2,4
10
20
0

−10 0

−20 −20
0 100 200 300 400 500 0 100 200 300 400 500
time [s] time [s]

20 20

10 10

0 0
τ2,5

τ2,6

−10 −10

−20 −20

−30 −30
0 100 200 300 400 500 0 100 200 300 400 500
time [s] time [s]

Fig. 8. Control input t2 of each AUV. (a) Control input t2 of AUV 1. (b) Control input t2 of AUV 2. (c) Control input t2 of AUV 3. (d) Control input t2 of AUV 4. (e) Control
input t2 of AUV 5. (f) Control input t2 of AUV 6.

7. Conclusion References

In this paper, leader–follower formation control has been Aguiar, P., Pascoal, A.M., 2007. Dynamic positioning and way-point tracking of
proposed for underactuated AUVs while the velocity of the leader underactuated AUVs in the presence of ocean currents. International Journal of
is not available for control design. A virtual vehicle with a Control 80 (7), 1092–1108.
Aicardi, M., Casalino, G., Bicchi, A., Balestrino, A., 1995. Closed loop steering of
trajectory that converges to the reference trajectory of the unicycle like vehicles via Lyapunov techniques. IEEE Robotics Automation
follower has been constructed. Next, the position tracking control Magazine 2 (1), 27–35.
has been designed for the follower using backstepping methodol- Balch, T., Arkin, R.C., 1998. Behavior-based formation control for multirobot teams.
ogy and Lyapunov analysis. An approximation-based control has IEEE Transactions on Robotics and Automation 14 (6), 926–939.
Chaillet, A., Lorı́a, A., 2008. Uniform semiglobal practical asymptotic stability
been employed to handle the parametric uncertainties and the for non-autonomous cascaded systems and applications. Automatica 44 (2),
unknown disturbances of the AUV. Simulation results have 337–347.
demonstrated the effectiveness of the leader–follower formation Cui, R., Ge, S.S., How, B.V.E., Choo, Y.S., 2009. Leader–follower formation control of
control of the underactuated underwater vehicles. One possible underactuated AUVs with leader position measurement. In: Proceedings of the
2009 IEEE International Conference on Robotics and Automation, Kobe, Japan,
future research direction could be the leader-follower formation
pp. 2441–2446.
with random communication time delays, such as the delays Dawson, D.M., Qu, Z., Lewis, F.L., Dorsey, J.F., 1990. Robust control for the tracking
modeled by Markov chains (Shi and Yu, 2009). of robot motion. International Journal of Control 52 (3), 581–595.
1502 R. Cui et al. / Ocean Engineering 37 (2010) 1491–1502

Do, K.D., Pan, J., 2004. State- and output-feedback robust path-following Lewis, M.A., Tan, K.H., 1997. High precision formation control of mobile robots
controllers for underactuated ships using Serret-Frenet frame. Ocean En- using virtual structures. Autonomous Robots 4 (4), 387–403.
gineering 31 (5–6), 587–613. Li, J., Lee, P., 2005. Design of an adaptive nonlinear controller for depth control of
Dong, W., Guo, Y., 2005. Global time-varying stabilization of underactuated surface an autonomous underwater vehicle. Ocean Engineering 32, 2165–2181.
vessel. IEEE Transactions on Automatic Control 50 (6), 859–864. Narendra, K., Annaswamy, A., 1989. Stable Adaptive Systems. Prentice-Hall,
Edwards, D., Bean, T., Odell, D., Anderson, M., 2004. A leader–follower algorithm Englewood Cliffs, NJ.
for multiple AUV formations. In: Proceedings of IEEE/OES Autonomous Ogren, P., Fiorelli, E., Leonard, N., 2004. Cooperative control of mobile sensor
Underwater Vehicles, vol. 1, Sebasco, ME, pp. 40–46. networks: adaptive gradient climbing in a distributed environment. IEEE
Evans, J., Nahon, M., 2004. Dynamics modeling and performance evaluation Transactions on Automatic Control 49 (8), 1292–1302.
of an autonomous underwater vehicle. Ocean Engineering 31 (14–15), Panteley, E., Loria, A., 1998. On global uniform asymptotic stability of nonlinear
1835–1858. time-varying systems in cascade. Systems and Control Letters 33 (2),
Fahimi, F., 2007. Sliding-mode formation control for underactuated surface 131–138.
vessels. IEEE Transactions on Robotics 23 (3), 617–622. Purwar, S., Kar, I., Jha, A., 2008. Adaptive output feedback tracking control of robot
Fiorelli, E., Leonard, N., Bhatta, P., Paley, D., Bachmayer, R., Fratantoni, D., 2006. manipulators using position measurements only. Expert Systems with
Multi-AUV control and adaptive sampling in Monterey Bay. IEEE Journal of Applications 34 (4), 2789–2798.
Oceanic Engineering 31 (4), 935–948. Ren, W., Beard, R.W., 2004. Formation feedback control for multiple spacecraft via
Fossen, T., 1994. Guidance and Control of Ocean Vehicles. Wiley Interscience, virtual structures. IEE Proceedings—Control Theory and Applications 151 (3),
New York. 357–368.
Fua, C.H., Ge, S.S., Do, K.D., Lim, K.W., 2007. Multirobot formations based on the Ren, W., Sorensen, N., 2008. Distributed coordination architecture for multi-robot
Queue-formation scheme with limited communication. IEEE Transactions on formation control. Robotics and Autonomous Systems 56 (4), 324–333.
Robotics 23 (6), 1160–1169. Schoenwald, D., 2000. AUVs: in space, air, water, and on the ground. IEEE Control
Ge, S.S., Fua, C.H., 2004. Queues and artificial potential trenches for multirobot Systems Magazine 20 (6), 15–18.
formations. IEEE Transactions on Robotics 21 (4), 646–656. Shi, Y., Yu, B., 2009. Output feedback stabilization of networked control systems
Ge, S.S., Hang, C.C., Lee, T.T., Zhang, T., 2002. Stable Adaptive Neural Network with random delays modeled by Markov chains. IEEE Transactions on
Control. Kluwer Academic, Boston, MA. Automatic Control 54 (7), 1668–1674.
Ge, S.S., Wang, C., 2004. Adaptive neural network control of uncertain MIMO non- Skjetne, R., Moi, S., Fossen, T., 2002. Nonlinear formation control of marine craft.
linear systems. IEEE Transactions on Neural Networks 15 (3), 674–692. In: Proceedings of IEEE Conference on Decision and Control, vol. 2, Las Vegas,
Hong, K.-S., Agustinus, T.T., Song, J.-B., 2008. Mobile robot control architecture NV, pp. 1699–1704.
for reflexive avoidance of moving obstacles. Advanced Robotics 22 (13–14), Stilwell, D.J., Bishop, B.E., 2000. Platoons of underwater vehicles. IEEE Control
1397–1420. Systems Magazine 20 (6), 45–52.
Ihle, I.A.F., Jouffroy, J., Fossen, T.I., 2006. Formation control of marine surface craft: Tee, K.P., Ge, S.S., 2006. Control of fully actuated ocean surface vessels using a class
a Lagrangian approach. IEEE Journal of Oceanic Engineering 31 (4), 922–934. of feedforward approximators. IEEE Transactions on Control System Technol-
Kyrkjebo, E., Pettersen, K., 2006. A virtual vehicle approach to output synchroniza- ogy 14 (4), 750–756.
tion control. In: Proceeding of 45th IEEE Conference on Decision and Control, Wang, P.K.C., 1991. Navigation strategies for multiple autonomous mobile robots
vol. 1, San Diego, CA, pp. 6016–6021. moving in formation. Journal of Robotic Systems 8 (2), 177–195.
Lapierre, L., Soetanto, D., 2007. Nonlinear path-following control of an AUV. Ocean Widyotriatmo, A., Hong, K.S., Prayudhi, L.H., 2010. Robust stabilization of a
Engineering 34 (11–12), 1734–1744. wheeled vehicle: Hybrid feedback control design and experimental validation.
Lawton, J.R.T., 2000. A behavior-based approach to multiple spacecraft formation Journal of Mechanical Science and Technology 24 (2), 513–520.
flying. Ph.D. Thesis, Department of Electrical and Computer Engineering, Woolsey, C., Techy, L., 2009. Cross-track control of a slender, underactuated AUV
Brigham Young University. using potential shaping. Ocean Engineering 36 (1), 82–91.