Automatica 77 (2017) 353–359

Contents lists available at ScienceDirect

Automatica
journal homepage: www.elsevier.com/locate/automatica

Brief paper

Neural network-based output feedback control for reference tracking

of underactuated surface vessels✩
Bong Seok Park a , Ji-Wook Kwon b , Hongkeun Kim c
a
Division of Electrical, Electronic and Control Engineering, Kongju National University, Cheonan 31080, Republic of Korea
b
Yujin Robot Co., Ltd, Seoul 08589, Republic of Korea
c
School of Mechatronics Engineering, Korea University of Technology and Education, Cheonan 31253, Republic of Korea

article info abstract

Article history: This paper proposes an adaptive output feedback control for trajectory tracking of underactuated surface
Received 2 January 2015 vessels (USVs). For the realistic dynamical model of USVs, we consider the USV model, where the mass
Received in revised form and damping matrices are not diagonal. Moreover, except the mass matrix, the system parameters and
7 October 2016
nonlinearities of the USV are all assumed to be unknown. Despite this uncertain circumstance, we develop
Accepted 19 October 2016
an adaptive observer based on the neural networks to estimate the velocity data of USVs. Then, an output
feedback control law is designed by simultaneously considering the input saturation and underactuated
Keywords:
problems.
Underactuated surface vessel © 2016 Elsevier Ltd. All rights reserved.
Adaptive observer
Neural network
Input saturation

1. Introduction controllers under persistent excitation condition, yielding that

the yaw velocity should not be zero and thus, a straight line
Autonomous underwater vehicles (AUVs) and underactuated cannot be tracked. A solution to this problem was firstly proposed
surface vessels (USVs) have received increasing attention from in Do, Jiang, and Pan (2002a). After this relaxed result, the
control engineers for developing marine resources. However, the controllers for stabilization and tracking of USVs were designed
design of controllers for USVs is complicated because the sway in Do, Jiang, and Pan (2002b) and Ghommam, Mnif, and Derbel
direction cannot be directly controlled. In addition, a USV cannot (2010), the full-state stabilization scheme was presented in Xie
be transformed into a driftless chained system (Reyhanoglu, 1997). and Ma (2015), and the sliding mode controllers were proposed
Because of these problems, some of the control techniques such in Perera and Soares (2012) and Yu, Zhu, Xia, and Liu (2012) to
as for mobile robots cannot be applied to the control of USVs show the robustness against the uncertainties such as unknown
(Chwa, 2011). Therefore, extensive research has concentrated on hydrodynamic damping coefficients and external disturbances.
attempting to overcome these limitations. These papers assumed that the mass and damping matrices of USVs
A tracking controller combined with adaptive technique and are diagonal because in the absence of this assumption, the cascade
backstepping technique was proposed in Godhavn, Fossen, and structure is broken which leads to many problems in the design
of the controller. However, since the mass and damping matrices
Berge (1998). Two constructive tracking solutions based on the
in real USVs are not diagonal, this assumption is not tenable.
Lyapunov direct method were developed in Jiang (2002) to achieve
A solution to this problem was proposed in Do and Pan (2005)
global asymptotic tracking. However, these studies designed
considering non-diagonal matrices. All these papers designed the
controllers based on full-state feedback.
In contrast to the state feedback results, output feedback
✩ This work was supported by Basic Science Research Program through the controllers using only position information were proposed in
National Research Foundation of Korea (NRF) funded by the Ministry of Education, Antonelli, Caccavale, Cjiaverini, and Villani (2000), Wondergem,
Science and Technology (NRF-2012R1A1A1041216). The material in this paper was Lefeber, Pettersen, and Nijmeijer (2011) and Zhang, Jia, and Qi
not presented at any conference. This paper was recommended for publication
(2011). Unfortunately, these controllers are for fully actuated
in revised form by Associate Editor Huaguang Zhangunder under the direction of
Editor Toshiharu Sugie. ships. Therefore, an output feedback controller for underactuated
E-mail addresses: [email protected] (B.S. Park), [email protected] ships was developed in Do, Jiang, Pan, and Nijmeijer (2004), but
(J.-W. Kwon), [email protected] (H. Kim). the mass matrices of USVs should be diagonal. For USVs having
http://dx.doi.org/10.1016/j.automatica.2016.11.024
0005-1098/© 2016 Elsevier Ltd. All rights reserved.
354 B.S. Park et al. / Automatica 77 (2017) 353–359

non-diagonal mass matrices, Do et al. proposed a solution in Y|v|r , Y|r |r , Nv , N|v|v , N|r |v , Nr , N|v|r , and N|r |r are linear and quadratic
Do and Pan (2006). However, they assumed that the nonlinear drag coefficients. In this paper, we assume that η is measurable,
damping matrix is diagonal and hydrodynamic derivatives are while ν is not. Moreover, M is only known.1
constants, and did not consider the input saturation problem.
Ignoring the input saturation problem in the design of the Remark 1. We emphasize that the off-diagonal term m23 of the
controller can degrade the performance of the real physical system mass matrix M is not zero in general because the shape of the
(Chen, Ge, How, & Choo, 2013). bow is different from that of the stern. This means that the sway
Motivated by these observations and in order to reflect more dynamics is also influenced by the yaw moment τr and thus,
realistic situation, we consider the USV model that neither the reflects a complicated but realistic structure of the USVs.
mass matrix nor the damping matrix is diagonal. The nonlinear
In order to deal with the situation that τr acts directly on
damping terms in the model are further assumed to be unknown.
the sway dynamics by the non-diagonal nature of M, we use the
To this setup, a neural network-based adaptive observer is first
following state transformations (Do & Pan, 2005): x̄ = x + ε cos ψ ,
developed to estimate the velocity data of the USV despite the
ȳ = y + ε sin ψ , and v̄ = v + ε r, where ε = m23 /m22 . Then, the
uncertainties and external disturbances. In combination with the
USV (1) can be rewritten as
observer, an output feedback controller is then proposed for
the reference tracking of the USV. In particular, we develop the x̄˙ = u cos ψ − v̄ sin ψ, u̇ = ϕu + d1 + ((τu − ϖu )/m11 ) ,
additional controllers that are able to deal with the input saturation
ȳ˙ = u sin ψ + v̄ cos ψ, v̄˙ = ϕv + d2 , (2)
and underactuated problems simultaneously. Finally, the reference
trajectory can be any one, including a straight line, due to the ψ̇ = r , ṙ = ϕr + d3 + (m22 (τr − ϖr )/∆) ,
proposed approach angle that uses only position information, and
d11 (u) d22 (v,r )
where ϕu = vr + ϕv = − m v−
m22 m23 2 11
the proposed controller guarantees the ultimate boundedness of m11 m11
r − m11
u, m22
ur − m22
the tracking errors. d23 (v,r )
m22
ϕr = ∆ {(m11 m22 −
r, 1
) v + (m11 m23 − m23 m22 )ur −
m222 u
(d33 (v, r )r + d32 (v, r )v)m22 + (d23 (v, r )r + d22 (v, r )v)m23 }, d1 =
2. Underactuated surface vessel model τd,u /m11 , d2 = τd,v /m22 , d3 = (−m23 τd,v + m22 τd,r )/∆, and ∆ =
m22 m33 − m223 . In addition, ϖu = τu − σu and ϖr = τr − σr . Note
The kinematics and dynamics of USVs in the horizontal plane that, since M is only known, ϕi and the bounded di are uncertain.
are described as follows Skjetne, Fossen, and Kokotovic (2005): The control objective is to design a dynamic output feedback
M ν̇ = −C (ν)ν − D(ν)ν + τd + σ (τ ), controller so that, under the controller, the USV (1) with the output
η tracks the trajectory given below.
η̇ = J (ψ)ν, (1)

where η = [x, y, ψ] denotes position (x, y) and yaw angle ψ of

T Assumption 1. The reference trajectory is the one generated by
USV in the earth-fixed frame; ν = [u, v, r ]T denotes surge, sway, η̇d = J (ψd )νd , i.e., ψ̇d = rd and
and yaw velocities of USV in the body-fixed frame, respectively; ẋd = ud cos ψd − vd sin ψd , ẏd = ud sin ψd + vd cos ψd ,
τd = [τd,u , τd,v , τd,r ]T is the bounded disturbances induced by
waves, wind, and ocean currents; and τ = [τu , 0, τr ]T is the where ηd = [xd , yd , ψd ]T and νd = [ud , vd , rd ]T are the vectors to
control vector of the surge force τu and the yaw moment τr . The be tracked. Moreover, ψd and νd are continuous and bounded with
saturated control σ (τ ) = [σu (τu ), 0, σr (τr )]T is defined as follows: their first derivatives again bounded.
σi = τi,min if τi ≤ τi,min , σi = τi if τi,min < τi < τi,max , With the assumption, we define the tracking errors as
σi = τi,max if τi ≥ τi,max , where i = u, r, and the bounds τi,min and
τi,max are known. We define the mismatch between the controls xe = x̄ − x̄d , ye = ȳ − ȳd , ψe = ψ − ψa , (3)
with saturation and without saturation as the dead-zone function
where x̄d = xd +ε cos ψd , ȳd = yd +ε sin ψd , and ψa is an approach
ϖ = [ϖu , 0, ϖr ]T = τ − σ (τ ). Then, the saturated control in (1)
angle to be defined later.
is given by σ (τ ) = τ − ϖ .
The matrices J (ψ), D(ν), C (ν), and M are given as
3. Main results
cos ψ − sin ψ 0
 
J (ψ) = sin ψ cos ψ 0 , In this section, we design the controller using an adaptive
0 0 1 observer, which is capable of estimating the velocity data of USV

m11 0 0
 even in the presence of uncertainties.
M = 0 m22 m23 ,
0 m23 m33 3.1. Neural network
0 0 −m22 v − m23 r
 
C (ν) = 0 0 m11 u , Neural networks (NNs) can be used to approximate the un-
m22 v + m23 r −m11 u 0 known nonlinearities according to the universal approximation
property (Oussar, Rivals, & Dreyfus, 1998). NN can approximate
d11 (u) 0 0 any continuous function F (χj ) over the compact set Ωχj ∈ RNi to
 
D(ν) = 0 d22 (v, r ) d23 (v, r ) ,
arbitrary any accuracy as Chen and Ge (2013) F (χj ) = Wj∗T Θ (χj )+
0 d32 (v, r ) d33 (v, r )
κj∗ where j = 1, 2, 3, Wj∗ ∈ RNh is the constant optimal
where m11 = m − Xu̇ , m22 = m − Yv̇ , m23 = mxg − Yṙ , m33 = Iz − Nṙ , weights, κj∗ is the bounded reconstruction error, and Θ (χj ) =
d11 (u) = −(Xu + Xu|u| |u|), d22 (v, r ) = −(Yv + Y|v|v |v| + Y|r |v |r |),
[Θj,1 , . . . , Θj,Nh ]T is composed of the Gaussian activation functions
d23 (v, r ) = −(Yr + Y|v|r |v|+ Y|r |r |r |), d32 (v, r ) = −(Nv + N|v|v |v|+
N|r |v |r |), and d33 (v, r ) = −(Nr + N|v|r |v| + N|r |r |r |). Here, m is the
mass of USV; Xu̇ , Yv̇ , Yṙ , and Nṙ are the added masses; xg is the Xb -
coordinate of USV center of gravity in the body-fixed frame; Iz is the 1 In fact, M can be calculated using semi-empirical methods or hydrodynamic
inertia with respect to the vertical axis; Xu , Xu|u| , Yv , Y|v|v , Y|r |v , Yr , computation programs as in Skjetne et al. (2005).
 B.S. Park et al. / Automatica 77 (2017) 353–359 355

−χjT χj /2
with expected value 0 and variance 1 (i.e., Θj,n = √1 e , where γ2 > 0, and uv , vv , and rv are the virtual controls of û, v̂ , and
√ 2π
r̂ to be designed later, respectively. The filtered virtual controls uf ,
n = 1, . . . , Nh , and thus, we have ∥Θ (χj )∥ ≤ Nh /2π ). Ni and
Nh denote the number of the inputs and the hidden layer nodes,
vf , and rf are given by
respectively. ζ1 u̇f = −uf + uv , ζ2 v̇f = −vf + vv , ζ3 ṙf = −rf + rv (8)
with ζ1 , ζ2 , ζ3 > 0, and the additional controls α1 , α2 , and α3 in
∗
Assumption 2. The optimal weight vector is bounded, i.e., ∥Wj ∥
≤ WMj , where WMj is a positive constant. (7a) are given by

α̇1 = cosh2 α1 {−Tu α1 − ϖu /m11 } /γ2 ,

3.2. Adaptive observer design
α̇2 = cosh2 α2 {ỹ + W
2T Θ2 (
χ2 ) − k3 s1 + k3 s2 − k3 s3
Let x̂, ŷ, ψ̂ , û, v̂ , and r̂ be the estimates of x̄, ȳ, ψ , u, v̄ , and r, − v̇f − 2xe sin ψ + 2ye cos ψ}/γ2 ,
respectively. Then, the adaptive observer is proposed as follows:
α̇3 = cosh2 α2 {−Tr α3 − m22 ϖr /∆} /γ2 , (9)
x̂˙ = û cos ψ − v̂ sin ψ + l1 x̃ cos ψ − l1 ỹ sin ψ,
where k3 , Tu , Tr > 0. It will turn out shortly that since the α1 - and
ŷ˙ = û sin ψ + v̂ cos ψ + l1 x̃ sin ψ + l1 ỹ cos ψ, α3 -dynamics are driven by the mismatch ϖ = τ − σ (τ ), they
effectively compensate the effect of the saturation. On the other
ψ̂˙ = r̂ + l1 ψ̃, û = z1 + l2 x̃, v̂ = z2 + l2 ỹ, r̂ = z3 + l2 ψ̃, hand, the control α2 is to deal with the problem of underactuation,
1T Θ1 (
ż1 = l1 l2 x̃ + W χ1 ) + ξ1 + (σu /m11 ), i.e., number of inputs fewer than the degrees of freedom.
2T Θ2 ( Step 2: In order to track any reference trajectory, we introduce
ż2 = l1 l2 ỹ + W χ2 ) + ξ 2 , an approach angle ψa as follows:
ż3 = l1 l2 ψ̃ + W
3T Θ3 (
χ3 ) + ξ3 + (σr m22 /∆), (4)
ψa = β tanh(D2 /γ1 ) + ψd (1 − tanh(D2 /γ1 )), (10)
where l1 and l2 are positive constants, x̃ = (x̄ − x̂) cos ψ + (ȳ −
where β = tan−1 ( −xe ), D = x2e + y2e , and γ1 is a positive
−y

ŷ) sin ψ , ỹ = −(x̄ − x̂) sin ψ + (ȳ − ŷ) cos ψ , ψ̃ = ψ − ψ̂ , e
constant.
ξ1 = x̃ + xe cos ψ + ye sin ψ , ξ2 = ỹ − xe sin ψ + ye cos ψ , ξ3 =
ψ̃ + ψe , W j are the estimates of W ∗ , j = 1, 2, 3, 
j χ1 = [û, v̂, r̂ , ỹ]T , Remark 2. Some previous studies used approach angles (Burger,
χ2 = [û, v̂, r̂ , x̃]T , and 
 χ3 = [û, v̂, r̂ ]T . Using (2) and (4), we can Pavlov, Borhaug, & Pettersen, 2009; Lapierre & Soetanto, 2007).
obtain the following observer error dynamics: Most of them are only useful for straight-line path and require
velocity data of USVs. However, our approach angle (10) does not
x̃˙ = ũ − l1 x̃ + r ỹ, ỹ˙ = ṽ − l1 ỹ − r x̃, ψ̃˙ = r̃ − l1 ψ̃, require the USV velocity data. In addition, it does not become zero
ũ˙ = −l2 ũ + W1∗T Θ1 (χ1 ) − W
1T Θ1 (
χ1 ) − ξ 1 + d1 , even when the USV is on the path, i.e., xe = ye = 0, which leads to
the ability of tracking all kinds of paths that satisfy Assumption 1.
ṽ˙ = −l2 ṽ + W2∗T Θ2 (χ2 ) − W
2T Θ2 (
χ2 ) − ξ2 + d2 ,
Differentiating both sides of (3) along the solutions of (2), (7a),
r̃˙ = −l2 r̃ + W3 Θ3 (χ3 ) − W
∗ T 3 Θ3 (
T
χ3 ) − ξ3 + d3 , (5)
and (7b) yields
where ũ = u − û, ṽ = v̄ − v̂ , r̃ = r − r̂, W1∗T Θ1 (χ1 ) = ẋe = (ũ + s1 + e1 + uv + γ2 tanh α1 ) cos ψ − ud cos ψd
ϕu − l2 r ỹ, W2∗T Θ2 (χ2 ) = ϕv + l2 r x̃, and W3∗T Θ3 (χ3 ) = ϕr . Note
− (ṽ + s2 + e2 + vv + γ2 tanh α2 ) sin ψ + v̄d sin ψd ,
that the saturated control σ (τ ) and the dead-zone function ϖ do
not appear in (5). For the system (5), we consider the Lyapunov ẏe = (ũ + s1 + e1 + uv + γ2 tanh α1 ) sin ψ − ud sin ψd
function candidate Vo = (x̃2 + ỹ2 + ψ̃ 2 + ũ2 + ṽ 2 + r̃ 2 + W
TW
1
1 + + (ṽ + s2 + e2 + vv + γ2 tanh α2 ) cos ψ − v̄d cos ψd ,
W2 W2 + W3 W3 )/2 and the NN tuning laws
 T   T 
ψ̇e = (r̃ + s3 + e3 + rv + γ2 tanh α3 ) − ψ̇a , (11)
1 (t ) = W
W 1 (t0 ) + x̃(t )Θ1 (t ) − x̃(t0 )Θ1 (t0 )
where v̄d = vd + ε rd . We now design the virtual controls uv , vv ,
 t  t and rv to stabilize the tracking errors as follows:
− (Θ̇1 x̃ − Θ1 l1 x̃)dt − ϑ1 1 dt ,
W
t0 t0 uv = −k1 xe cos ψ − k1 ye sin ψ + ud cos(ψ − ψd )
2 (t ) = W
W 2 (t0 ) + ỹ(t )Θ2 (t ) − ỹ(t0 )Θ2 (t0 ) + v̄d sin(ψ − ψd ),
vv = k1 xe sin ψ − k1 ye cos ψ − ud sin(ψ − ψd )
 t  t
− (Θ̇2 ỹ − Θ2 l1 ỹ)dt − ϑ2 2 dt ,
W
t0 t0
+ v̄d cos(ψ − ψd ),
3 (t ) = W
W 3 (t0 ) + ψ̃(t )Θ3 (t ) rv = −k2 ψe , (12)

where k1 and k2 are positive constants to be designed.

 t  t
+ ψ̃(l1 Θ3 − Θ̇3 )dt − ϑ3 3 dt ,
W (6) Step 3: Since s1 = u − ũ − uf − γ2 tanh α1 , s2 = v̄ − ṽ − vf −
t0 t0
γ2 tanh α2 , and s3 = r − r̃ − rf − γ2 tanh α3 , their derivatives along
where ϑ1 , ϑ2 , and ϑ3 are positive constants to be designed and the solutions of (2), (5), and (9) are
j = W ∗ − W
W j for j = 1, 2, 3.
j
1T Θ1 ( τu
ṡ1 = W χ1 ) + ξ1 + l2 (r ỹ + ũ) − u̇f + Tu α1 + ,
m11
3.3. Controller design
ṡ2 = l2 (ṽ − r x̃) + k3 s1 − k3 s2 + k3 s3 + xe sin ψ − ye cos ψ, (13)
Step 1: Define the errors as m22 τr
3T Θ3 (
ṡ3 = l2 r̃ + W χ3 ) + ξ3 − ṙf + Tr α3 + .
s1 = û − uf − γ2 tanh α1 , s2 = v̂ − vf − γ2 tanh α2 , ∆
In the derivation (13), one can see that the effect of the dead-
s3 = r̂ − rf − γ2 tanh α3 , (7a)
zone function ϖ in (2) is compensated by (9), giving rise to the s1 -
e1 = uf − uv , e2 = vf − vv , e3 = rf − rv , (7b) and s3 -dynamics with τu and τr unsaturated. Therefore, we are able
356 B.S. Park et al. / Automatica 77 (2017) 353–359

to derive τu and τr , and apply them to the s1 - and s3 -dynamics with + r̃ (W3∗T Θ3 (χ3 ) − W
3T Θ3 (
χ3 ) − ψe ) − W
1T W
̇ 1
no saturation presented.
2T W
−W ̇ 2 − W ̇ 3 + ũd1 + ṽ d2 + r̃d3 .
3T W (17)
Finally, the control inputs τu and τr are designed to stabilize the
error dynamics (13) as follows:
Adding and subtracting Wj∗T Θj (
χj ) where j = 1, 2, 3, we have
τu = m11 (−ξ1 − W
1T Θ1 (
χ1 ) − Tu α1 − k3 s1 − k3 s2 + u̇f
Wj∗T Θj (χj ) − W
jT Θj (
χj ) + Wj∗T Θj (
χj ) − Wj∗T Θj (
χj )
− xe cos ψ − ye sin ψ), (14)
jT Θj (
=W χj ) + Wj∗T (Θj (χj ) − Θj (
χj )).
τr = ∆(−ξ3 − W3T Θ3 (
χ3 ) − Tr α3 − k3 s2 − k3 s3
∂Θ
+ ṙf − ψe )/m22 . On the other hand, we have Θj (χj ) = Θj ( χj ) + ∂j
χ + Gj where
χj j

Let us consider the Lyapunov function candidate χj = χj −
 χj and Gj stands for the high-order term. Therefore, by the
property of the radial basis function (Ge, Hang, Lee, & Zhang, 2001),
VT = Vo + Vc , (15) we have Wj∗T (Θj (χj ) − Θj (χj )) ≤ ϵj Wj∗T Θ̄j where ϵj > 0 and Θ̄j
1 are bounded vector functions. Since ϵj Wj∗T Θ̄j and dj are bounded,
Vc = (x2e + y2e + ψe2 + s21 + s22 + s23 + e21 + e22 + e23 ).
2 there exist positive constants δj such that ϵj ∥Wj∗ ∥ ∥Θ̄j ∥ + |dj | ≤ δj .
We now provide our main result as follows. Substituting (6), (16), and (17) into the time derivative of VT in (15),
we have
Theorem 1. Suppose that Assumptions 1–2 hold and the control
V̇T ≤ −l1 x̃2 − l1 ỹ2 − l1 ψ̃ 2 − l2 ũ2 − l2 ṽ 2 − l2 r̃ 2 − k1 x2e
input (14) with (4), (6), (8), (9), and (12) is applied to the USV (1).
Let µmax be any positive constant satisfying µmax > (WM2 1 + WM2 2 + − k1 y2e − k2 ψe2 − k3 s21 − k3 s22 − k3 s23 − e21 /ζ1
WM2 3 )/2. Then, there are gains l1 , l2 , k1 , k2 , k3 , ζ1 , ζ2 , ζ3 , ϑ1 , ϑ2 , and − e22 /ζ2 − e23 /ζ3 + ϑ1 W1T W 1 + ϑ2 W 2T W
2
ϑ3 (all positive) such that the error signals of the closed-loop system  √
initialized in the set {VT ≤ µmax } are ultimately bounded. + ϑ3 W 3T W
3 + rmax Nh ∥W 1 ∥|ỹ|/ 2π
 √
+ rmax Nh ∥W 2 ∥|x̃|/ 2π + (|xe | + |ye |)(|e1 |
Remark 3. We note that even in the case µmax ≤ (WM2 1 + WM2 2 +
+ |e2 | + 2γ2 ) + l2 (|s1 ||ũ| + |s2 ||ṽ| + |s3 ||r̃ |)
WM2 3 )/2, the result still holds. Indeed, if we design the control gains
for µ̄max > (WM2 1 + WM2 2 + WM2 3 )/2 instead of µmax , then Theorem 1 + |ψe |(|e3 | + γ2 ) + |ψe ||ψ̇a | + l2 rmax (|s1 ||ỹ|
implies the ultimate boundedness of the error signals initialized in + |s2 ||x̃|) + |ũ|δ1 + |ṽ|δ2
the set {VT ≤ µ̄max }. With the same gains, this in turn leads to + |r̃ |δ3 + |e1 ||u̇v | + |e2 ||v̇v | + |e3 ||ṙv |. (18)
the ultimate boundedness of the closed-loop solutions with their
initial conditions in {VT ≤ µmax } since the level set {VT ≤ µmax } is From Young’s inequality and Assumption 2, (18) can be written as
contained in {VT ≤ µ̄max }.  2
V̇T ≤ − l1 − l2 rmax /2 − Nh rmax
2
/2π (x̃2 + ỹ2 ) − l1 ψ̃ 2


Proof. We begin by estimating the bounds of some quantities in − (l2 /2 − 1/2) (ũ2 + ṽ 2 + r̃ 2 )
terms of the control gains. From Assumption 1, the definitions of
− (k1 − 3/2) (x2e + y2e ) − k2 − 1 − p21 /2 ψe2
 
ψa , e1 , and e2 , and (7a)–(11), one can find a positive constant p1
dependent on k1 such that |ψ̇a | ≤ p1 on the compact level set − (k3 − l2 ) (s21 + s22 ) − (k3 − l2 /2) s23
{VT ≤ µmax }. On the other hand, from the definitions of r̃ and
− 1/ζ1 − 1/2 − p22 /2 e21 − 1/ζ2 − 1/2 − p23 /2 e22
   
e3 , and from (12) and (7a), r has a maximum rmax , which depends
on the gain k2 , on the set {VT ≤ µmax }. Similarly, one can show − 1/ζ3 − 1/2 − p24 /2 e23 − (ϑ1 /2 − 1/4) ∥W 1 ∥2
 
that there are positive constants p2 , p3 , and p4 such that |u̇v | ≤ p2 ,
|v̇v | ≤ p3 , and |ṙv | ≤ p4 hold on the set {VT ≤ µmax }. We note that − (ϑ2 /2 − 1/4) ∥W 2 ∥2 − ϑ3 ∥W 3 ∥2 /2 + c1 , (19)
p2 , p3 , and p4 depend on both k1 and k2 .
where c1 = (4 + 9γ + δ + δ + δ )/2 + ϑ
2
2
2
1
2
2 3
2 2
1 WM1 /2 + ϑ 2
2 WM2 /2 +
Let us consider the function Vc in (15). Noting that ė1 =
−e1 /ζ1 − u̇v , ė2 = −e2 /ζ2 − v̇v , and ė3 = −e3 /ζ3 − ṙv , and using ϑ3 WM2 3 /2. If we choose l1 = 2
l2 rmax 2 / + Nh rmax
2
/2π + l∗1 , l2 = 1 + 2l∗2 ,
(7a), (11), (12), (13), and (14), one can obtain the time derivative of k1 = 3/2 + k∗1 , k2 = 1 + / + k∗2 , k3 = l2 + k∗3 , 1/ζ1 =
2
p1 2
Vc as follows: 1/2 + p22 /2 + ζ1∗ , 1/ζ2 = 1/ + / + ζ2∗ , 1/ζ3 = 1/2 + p24 /2 + ζ3∗ ,
2 p23 2
V̇c = −k1 x2e − k1 y2e − k2 ψe2 − k3 s21 − k3 s22 − k3 s23
ϑ1 = 1/2 + 2ϑ1∗ , ϑ2 = 1/2 + 2ϑ2∗ , and ϑ3 = 2ϑ3∗ , where l∗1 , l∗2 , k∗1 ,
k∗2 , k∗3 , ζ1∗ , ζ2∗ , ζ3∗ , ϑ1∗ , ϑ2∗ , and ϑ3∗ are all positive, then from (19), we
+ xe {(ũ + e1 + γ2 tanh α1 ) cos ψ have the following inequality:
− (ṽ + e2 + γ2 tanh α2 ) sin ψ}
V̇T ≤ −2c0 VT + c1 , (20)
+ ye {(ũ + e1 + γ2 tanh α1 ) sin ψ
where c0 = min{l∗1 , l∗2 , k∗1 , k∗2 , k∗3 , ζ1∗ , ζ2∗ , ζ3∗ , ϑ1∗ , ϑ2∗ , ϑ3∗ }.
+ (ṽ + e2 + γ2 tanh α2 ) cos ψ} We now claim that for any c0∗ > 0, there always exist the
+ ψe (r̃ + e3 + γ2 tanh α3 − ψ̇a ) well-defined control gains li , ki , ζi , and ϑi that guarantee c0 = c0∗ .
+ s1 (l2 r ỹ + l2 ũ) + s2 (l2 ṽ − l2 r x̃) + l2 s3 r̃ This can be shown by observing the dependency of the constants
rmax (k2 ), p1 (k1 ), p2 (k1 , k2 ), p3 (k1 , k2 ), and p4 (k1 , k2 ) on the gains.
− e21 /ζ1 − e22 /ζ2 − e23 /ζ3 − e1 u̇v − e2 v̇v − e3 ṙv . (16) A detailed procedure is given as follows: (a) Let l∗1 , l∗2 , k∗1 , k∗2 ,
k∗3 , ζ1∗ , ζ2∗ , ζ3∗ , ϑ1∗ , ϑ2∗ , and ϑ3∗ be equal to c0∗ . (b) Choose l2 , k1 , ϑ1 ,
On the other hand, using the dynamics (5), the derivative of Vo with
respect to the time is given by
ϑ2 , and ϑ3 such that l2 ≥ 1 + 2l∗2 , k1 ≥ 3/2 + k∗1 , ϑ1 = 1/2 + 2ϑ1∗ ,
ϑ2 = 1/2 + 2ϑ2∗ , and ϑ3 = 2ϑ3∗ , respectively. (c) Select k2 and k3
V̇o = −l1 x̃2 − l1 ỹ2 − l1 ψ̃ 2 − l2 ũ2 − l2 ṽ 2 − l2 r̃ 2 such that k2 ≥ 1 + p21 /2 + k∗2 and k3 ≥ l2 + k∗3 , respectively. This is
possible because p1 depends on k1 only, and k1 and l2 are already
+ ũ(W1∗T Θ1 (χ1 ) − W
1T Θ1 (
χ1 ) − xe cos ψ − ye sin ψ)
chosen in the previous step. (d) Finally, choose the remaining gains
+ ṽ(W2 Θ2 (χ2 ) − W2 Θ2 (
∗T  T
χ2 ) + xe sin ψ − ye cos ψ) such that l1 ≥ l2 rmax 2
/2 + Nh rmax
2
/2π + l∗1 , 1/ζ1 ≥ 1/2 + p22 /2 + ζ1∗ ,
 B.S. Park et al. / Automatica 77 (2017) 353–359 357

(a) Tracking result. (b) Tracking errors.

(c) Observer errors. (d) Norms of weighting vectors of neural networks.

Fig. 1. Simulation results. (a) Dotted: PD controller in Fossen (2002), solid: proposed method, dashed-dotted: reference trajectory; (b) dotted: xe , solid: ye , dashed-dotted:
ψe ; (c) solid: ũ, dotted: ṽ , dashed-dotted: r̃; and (d) solid: W
1 , dashed: W
2 , dotted: W
3 .

1/ζ2 ≥ 1/2 + p23 /2 + ζ2∗ , and 1/ζ3 ≥ 1/2 + p24 /2 + ζ3∗ . Noting that knowledge of the system parameters, we design l1 = 20, l2 = 3,
µmax > (WM2 1 + WM2 2 + WM2 3 )/2 and, by the previous procedure, k1 = k2 = k3 = 5, γ1 = 0.01, γ2 = 50, ζ1 = ζ2 = ζ3 = 0.1,
c1 /2c0∗ → (WM2 1 + WM2 2 + WM2 3 )/2 as c0∗ → ∞, one can design Tu = Tr = 4, and ϑ1 = ϑ2 = ϑ3 = 1. The disturbances are
the control gains that ensure c1 /2c0 < µmax . Together with (20), generated by the white Gaussian noise processes with zero mean.
this implies that the level set Ω1 = {VT ≤ c1 /2c0 } is contained in The saturation levels are chosen as τu,max = 2(N ), τu,min = −2(N ),
Ω2 = {VT ≤ µmax } and V̇T is negative on Ω2 \ Ω1 . Thus, the error τr ,max = 1.5(N ), and τr ,min = −1.5(N ). To estimate the unknown
terms ϕu − l2 r ỹ, ϕv − l2 r x̃, and ϕr , we employ the radial basis
signals of the closed-loop system initialized in Ω2 are ultimately
function networks (RBFNs) with Nh = 5.
bounded and converge to the set Ω1 . 
To show the proposed control system can track both straight
and curved line, the reference velocities are chosen as follows:
Remark 4. The gains and design parameters have some effects on
(1) 0 ≤ t < 40 : ud = 0.1, vd = 0, rd = 0; (2) 40 ≤ t < 80 : ud =
the stability and performance of the closed-loop system. First of
0.2, vd = 0, rd = 0; (3) 80 ≤ t < 120 : ud = 0.2, vd = 0, rd =
all, given the set {VT ≤ µmax }, the gains and design parameters
−0.1 sin(π t /20); and (4) t ≥ 120 : ud = 0.2, vd = 0, rd = 0. The
of the proposed controller have to be chosen sufficiently large to
initial conditions are ηd = [0, 0, π /4]T and η = [−5, 5, 0]T . In
compensate both of the uncertainties of the system (1) and the
order to verify the effectiveness of the proposed method, we com-
unwanted effects of the cross-product terms (appearing in (18))
pare with the PD controller in Fossen (2002). Simulation results
with their upper bounds estimated over the set {VT ≤ µmax }.
are plotted in Fig. 1. Fig. 1(a) shows the trajectory tracking result.
Although the actual procedure to compute those gains and design
Compared with the PD controller in Fossen (2002), the trajec-
parameters is given at the end of the proof, one may also find them
tory tracking of the proposed method is well established not only
by gradually increasing them and then by performing repeated
straight-line paths, but also curved-line paths despite the exis-
simulations. On the other hand, if we increase li , ki , and ϑi , and
tence of the uncertainties and the input saturation. Fig. 1(b) and
decrease ζi , then the bound c1 /2c0 of the errors will be reduced
(c) show the tracking and observer errors, both of them are ulti-
and approach its minimum (WM2 1 + WM2 2 + WM2 3 )/2. However, this
mately bounded as expected. The norms of the NN approximation
may increase the control efforts.
parameters are presented in Fig. 1(d). Fig. 2 shows the surge force
and yaw moment generated by (14) and their saturated values. In
4. Computer simulation the simulation, the saturated ones are applied to (1).
We now examine the performance of the control system
In order to perform a computer simulation, we take all the through a set of simulations which are given in Fig. 3. Except
parameters of the USV from Skjetne et al. (2005) and assume that for the gains k1 , k2 , l1 , and l2 , all other setups are the same as
they are unknown to the controller. In particular, without the in the previous case. The values of k1 , k2 , l1 , and l2 are given in
358 B.S. Park et al. / Automatica 77 (2017) 353–359

(a) Solid: σu (τu ), dotted: τu . (b) Solid: σr (τr ), dotted: τr .

Fig. 2. Controls with saturation and without saturation.

(a) k1 = 1, k2 = 3, l1 = 20, and l2 = 3. (b) k1 = 5, k2 = 5, l1 = 20, and l2 = 3.

(c) l1 = 10, l2 = 1, k1 = 5, and k2 = 5. (d) l1 = 20, l2 = 3, k1 = 5, and k2 = 5.

Fig. 3. Simulation results with different gains. The first and second rows show the tracking errors (dotted: xe , solid: ye , dashed-dotted: ψe ) and observation errors (solid: ũ,
dotted: ṽ , dashed-dotted: r̃), respectively.

each subfigure. Numerical computations by using the results of 5. Conclusion

Fig. 3 show that the times that the tracking errors lie in between
±0.25 are 65.21 s in case of Fig. 3(a) and 49.81 s in case of In this paper, the NN-based output feedback control has been
Fig. 3(b). Moreover, using the results of Fig. 3(c) and (d), one can proposed for reference tracking of USVs with input saturation and
see that the observation errors are within ±0.25 after 25.26 s uncertainties. To estimate the velocity data of the USV with un-
and 18.36 s, respectively. Therefore, the increments in gains yield certainties, the NN-based observer is developed. Based on the es-
fast convergence of the errors as expected. Note that due to these timated states of the observer, the output feedback controller is
increments in gains and the result of Theorem 1, the bound c1 /2c0 designed using the additional controllers that are able to deal with
of the errors is also reduced. Instead, we observed that the control the input saturation and underactuated problems simultaneously.
efforts are increased, but the corresponding figures are omitted due Furthermore, the approach angle that uses only position informa-
to the space limit. tion is developed to track any reference trajectories including a
 B.S. Park et al. / Automatica 77 (2017) 353–359 359

straight line. The boundedness of the tracking errors is proved us- Reyhanoglu, M. (1997). Exponential stabilization of an underactuated autonomous
ing Lyapunov analysis. It can be made small by increasing the gains surface vessel. Automatica, 33(12), 2249–2254.
Skjetne, R., Fossen, T. I., & Kokotovic, P. V. (2005). Adaptive maneuvering, with
of the proposed dynamic controller, but shown to have its mini- experiments, for a model ship in a marine control laboratory. Automatica, 41(2),
mum given by {VT ≤ (WM2 1 + WM2 2 + WM2 3 )/2}. The simulation re- 289–298.
sults show that the proposed control scheme is effective for USVs Wondergem, M., Lefeber, E., Pettersen, K. Y., & Nijmeijer, H. (2011). Output
feedback tracking of ships. IEEE Transactions on Control Systems Technology,
with input saturation and uncertainties. 19(2), 442–448.
Xie, W., & Ma, B. (2015). Robust global uniform asymptotic stabilization of
underactuated surface vessels with unknown model parameters. International
References Journal of Robust and Nonlinear Control, 25(7), 1037–1050.
Yu, R., Zhu, Q., Xia, G., & Liu, Z. (2012). Sliding mode tracking control of an
Antonelli, G., Caccavale, F., Cjiaverini, S., & Villani, L. (2000). Tracking control for underactuated surface vessel. IET Control Theory & Applications, 6(3), 461–466.
underwater vehicle-manipulator systems with velocity estimation. IEEE Journal Zhang, L. J., Jia, H. M., & Qi, X. (2011). NNFFC-adaptive output feedback trajectory
of Oceanic Engineering, 25(2), 399–413. tracking control for a surface ship at high speed. Ocean Engineering, 38(13),
Burger, M., Pavlov, A., Borhaug, E., & Pettersen, K.Y. (2009). Straight line path 1430–1438.
following for formation of underactuated surface vessels under influence of
constant ocean currents. In Proc. American Control Conference, St. Louis, MO
(pp. 3065–3070).
Bong Seok Park received his B.S., M.S., and Ph.D. degrees
Chen, M., & Ge, S. S. (2013). Direct adaptive neural control for a class of uncertain
in Electrical and Electronic Engineering from Yonsei Uni-
nonaffine nonlinear systems based on disturbance observer. IEEE Transactions
versity in 2005, 2008, and 2011, respectively. Since 2015,
on Cybernetics, 43(4), 1213–1225.
he has been with the Division of Electrical, Electronic, and
Chen, M., Ge, S. S., How, B. V. E., & Choo, Y. S. (2013). Robust adaptive position
Control Engineering, Kongju National University. His re-
mooring control for marine vessels. IEEE Transactions on Control Systems
search interests include nonlinear control, formation con-
Technology, 21(2), 395–409.
trol, and control of robots.
Chwa, D. (2011). Global tracking control of underactuated ships with input and
velocity constraints using dynamic surface control method. IEEE Transactions
on Control Systems Technology, 19(6), 1357–1370.
Do, K. D., Jiang, Z. P., & Pan, J. (2002a). Underactuated ship global tracking under
relaxed conditions. IEEE Transactions on Automatic Control, 47(9), 1529–1536.
Do, K. D., Jiang, Z. P., & Pan, J. (2002b). Universal controllers for stabilization and
tracking of underactuated ships. Systems & Control Letters, 47(4), 299–317.
Do, K. D., Jiang, Z. P., Pan, J., & Nijmeijer, H. (2004). A global output-feedback Ji-Wook Kwon received the B.S., M.S., and Ph.D. degrees in
controller for stabilization and tracking of underactuated ODIN: A spherical Electrical and Computer Engineering from Ajou University,
underwater vehicle. Automatica, 40(1), 117–124. Suwon, Korea, in 2005, 2007, and 2012, respectively. In
Do, K. D., & Pan, J. (2005). Global tracking control of underactuated ships with 2012, he was a senior researcher with Korea Institute of
nonzero off-diagonal terms in their system matrices. Automatica, 41(1), 87–95. Industrial Technology, Ansan, Korea. From 2013 to 2015,
Do, K. D., & Pan, J. (2006). Underactuated ships follow smooth paths with he was a senior researcher with YICT, Yonsei University,
integral actions and without velocity measurements for feedback: Theory and Incheon, Korea. Since 2015, he has been with R&D center,
experiments. IEEE Transactions on Control Systems Technology, 14(2), 308–322. Yujin Robot, Seoul Korea, where he is currently a manager.
Fossen, T. I. (2002). Marine control systems. Trondheim, Norway: Marine Cybernet- His research interests include mobile robot navigation,
ics. service robot and its applications, and multiple robot
Ge, S. S., Hang, C. C., Lee, T. H., & Zhang, T. (2001). Stable adaptive neural network
cooperation.
control. Norwell, MA: Kluwer.
Ghommam, J., Mnif, F., & Derbel, N. (2010). Global stabilisation and tracking control
of underactuated surface vessels. IET Control Theory & Applications, 4(1), 71–88.
Godhavn, J. M., Fossen, T. I., & Berge, S. P. (1998). Non-linear and adaptive Hongkeun Kim received his B.S. degree from Hanyang
backstepping designs for tracking control of ships. International Journal of University, Korea, in 2005, and Ph.D. degree from Seoul
Adaptive Control and Signal Processing, 12(8), 649–670. National University, Korea, in 2012. From 2014 to 2015, he
Jiang, Z. P. (2002). Global tracking control of underactuated ships by Lyapunov’s was a postdoctoral researcher at University of Groningen,
direct method. Automatica, 38(2), 301–309. the Netherlands. Since 2015, he has been with Korea
Lapierre, L., & Soetanto, D. (2007). Nonlinear path-following control of an AUV. University of Technology and Education, Korea, where he
Ocean Engineering, 38(1), 1734–1744. is currently an assistant professor.
Oussar, Y., Rivals, I., & Dreyfus, L. P. (1998). Training wavelet networks for nonlinear
dynamic input–output modeling. Neurocomputing, 20(1–3), 173–188.
Perera, L. P., & Soares, C. G. (2012). Pre-filtered sliding mode control for nonlinear
ship steering associated with disturbances. Ocean Engineering, 51, 49–62.