This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCSYS.2020.2997600, IEEE Control
Systems Letters

Function Approximation Technique Based

Immersion and Invariance Control for Unknown
Nonlinear Systems
Yang Bai1 , Yujie Wang2 , Mikhail Svinin1 , Evgeni Magid3 , and Ruisheng Sun4

Abstract—A function approximation technique based immer- which is model-free, and thus applicable to a wide class
sion and invariance (FATII) control method is proposed in this of systems and robust to uncertainties. Also, the asymptotic
letter. Firstly, an unknown control system is restructured as the stability of the FATII method can be proved.
combination of an auxiliary system and a variation term from the
original system. The variation term is treated as a time-varying The FATII method is based on the function approximation
uncertainty and parameterized by a group of weighted chosen technique (FAT) [7]–[9], inspired by its applications on the
basis functions. These weights are estimated at every time instant adaptive control problems. The FAT-based approaches recon-
and the change of the estimates is governed by an update law. struct the time-varying uncertainties in the control systems
The update law is defined based on an immersion and invariance as the combination of a group of weighted basis functions
approach such that both the system state and the estimation
error converge to zero. The FATII method is model-free and thus and a remainder term. Then, model reference adaptive control
applicable to a wide range of systems. The asymptotic stability of (MRAC) techniques are utilized to deal with the unknown
the proposed method is established and its feasibility is verified weights, and the effects of the remainder term can be elimi-
under simulations. nated by a sliding mode controller [10]–[14]. Note that FAT-
Index Terms—Robust adaptive control, Uncertain systems based designs have a drawback that in the absence of persistent
excitation (PE) condition, the estimation for the weights is
I. I NTRODUCTION not guaranteed to converge to the actual value. The drawback
would deteriorate the control accuracy and lead to an undesired
T HE need for the control of nonlinear systems arises in
many practical applications that include the use of ships,
underwater vehicles, aircraft, satellites, flexible joint robots,
transient response of the closed loop system [15].
Noticing that the immersion and invariance (I&I) technique
[16], [17] can treat the state stabilization and the parameter
hyper-redundant and snake-like manipulators, walking robots,
estimation in a unified manner, for the design of the FATII
hybrid machines, etc.
method, we improve the FAT-based approaches by utilizing
To control these systems, various methods have been pro-
the I&I technique instead of the MRAC, such that not only
posed in the literature, which can be classified into two main
the system state, but also the error between the uncertainty and
types: model-based and model-free methods. Model-based
its estimation, are steered to zero. Nevertheless, the remainder
control methods include energy shaping [1], partial feedback
term generated in the function approximation process cannot
linearization [2], backstepping [3], sliding mode control [4],
be eliminated by a sliding mode controller when one directly
[5], etc. They commonly rely upon the model information of
incorporates the conventional I&I technique into the FAT-
the control systems, which limits their range of applications.
based controller. Regarding this problem, we modify the
Also, they are not sufficiently robust to large system uncer-
conventional I&I technique with the use of switching in order
tainties. Instead of utilizing the model information, a number
to deal with the remainder term.
of model-free methods [6] have been developed for designing
controllers directly from the input-output data, bypassing the The process for designing FATII controllers is as follows.
modelling step. Although they are widely applicable and A control system is reconstructed as the combination of an
adaptive to uncertainties, the asymptotic stability of these approximated one and the variation from it. The variation
methods is difficult to prove. term is then estimated with chosen basis functions weighted
To feature both the stability and the adaptiveness, in this by unknown parameters. Update laws are defined based on
letter, we propose a novel control method (the FATII method) the modified I&I technique such that the parameters can be
automatically adjusted and the effect of the uncertainty term
This research was supported, in part, by the Japan Science and Technol- to the control process can be eliminated.
ogy Agency, the JST Strategic International Collaborative Research Program, The rest of this letter is organized as follows. First, in Sec-
Project No. 18065977, and by the Russian Foundation for Basic Research,
Project No. 19-58-70002. tion II we state the control problem, illustrate the process for
1 Y. Bai and M. Svinin are with the Information Science and Engineering constructing the FATII controller, and establish its asymptotic
Department, Ritsumeikan University, 1-1-1 Noji-higashi, Kusatsu, Shiga 525- stability. The feasibility of the proposed control method is
8577, Japan. [email protected],
[email protected] verified under simulations in Section III. Finally, conclusions
2 Y. Wang is with the Department of Electrical and Computer Engi- are drawn in Section IV.
neering, University of Illinois Urbana-Champaign, Urbana, IL 61801, USA.
[email protected]
3 E. Magid is with the Department of Intelligent Robotics, Kazan
II. C ONTROLLER DESIGN PROCESS
Federal University, Kremlyovskaya str. 35, Kazan 420008, Russian Federation.
[email protected] In this section, the FAT approach is firstly utilized to convert
4 R. Sun is with the School of Energy and Power Engineering,
Nanjing University of Science and Technology, Nanjing 210094, China. general control systems into a unified form. Next, the FATII
[email protected] control method is proposed, and its stability is established.

2475-1456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 01:40:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCSYS.2020.2997600, IEEE Control
Systems Letters

A. Statement of problem Note that the controllability of the restructured system (5)
Given a control system written in the state-space form implies the existence of bounded u∗ for the desired x such
that d is definable and bounded. Otherwise, the function
ẋ = f (x) + G(x)u + ξ, (1) approximate technique through (7) to (8) is not applicable.
For the control of (8), several remarks are in order.
where x ∈ Rn represents the state, u ∈ Rm with n > m,
represents the input, and ξ ∈ Rn for the external disturbance Remark 1. The control problem requires to eliminate the
which is assumed to be bounded, define an asymptotically effect of di in (8). For this purpose, di , at each time instant
stabilizing law u for system (1). t, is estimated by dˆi (t) through the MRAC techniques in
Assume that fe and G e respectively stand for the nominal the FAT-based controls. However, dˆi (t) is not guaranteed to
drift term and control matrix such that converge to di in the absence of the PE condition. The error
between the actual parameters and the parameter estimation
f = fe + δf , G = G e + δG, (2)
can produce large uncertainty, which would deteriorate the
where δf and δG denote the difference between the nominal control accuracy and lead to an undesired transient response
and actual values. By substituting (2) into (1), one obtains of the closed-loop system.
A unique feature of the I&I framework is that it allows to
ẋ = fe + Gu
e + ξ,
e (3) treat the state stabilization and the parameter estimation in a
unified manner [17]. For the convergence of both the system
where ξe = δf +δGu+ξ is the lumped error [18]. Note that as
state and the estimation error, we propose in Section II-B the
stated in [19], the term “unknown system” is utilized in place
FATII method based on the I&I approach.
of “uncertain system” to stress that f and G are not known a
priori and thus are possibly subject to large uncertainty ξ,
e so Remark 2. In the FAT-based controller design, the effect of 
that model-based control would not be applicable. in (8) is commonly covered by a sliding mode controller, but
By mimicking the pole placement method for linear sys- this operation cannot be accomplished when the conventional
tems, one introduces an auxiliary input u∗ ∈ Rn to match the I&I technique is incorporated into the FAT-based controller.
dimension of input with that of the output, such that Therefore, in the design of the FATII controller, we modify the
I&I technique with the use of switching to deal with .
u = G∗ u∗ , (4)
∗
Remark 3. Several candidates for the basis function ψi in (7)
where the auxiliary matrix G is chosen to be a full rank can be chosen to approximate the nonlinear functions, and in
m × n matrix. Then, the control system (1) is rewritten as this letter, we select the Fourier series [20], [21]. From the
e ∗ u∗ + ξ,
ẋ = fe + GG e (5) Weierstrass theorem, when N → ∞, the approximation error
 would be infinitesimally small. By selecting enough number
where the number of the inputs equals that of the states. Note of the basis functions, the estimation of the uncertainty can
that an essential condition for the selection of matrix G∗ is be sufficiently accurate such that  in (10) is negligible [20].
that the reconstructed system (5) is required to be controllable, Thus, it is reasonable to have the following [10]–[14]
otherwise the design of u∗ cannot guarantee the convergence
Assumption 1. The error  is bounded such that kk1 ≤ E1
of x. Methods for the design of G∗ that render system (5)
and kk2 ≤ E2 where E1 and E2 are positive constants.
controllable, are specified in Appendix.
To further simplify (5), one can rearrange it as B. FATII based controller design
ẋ = u∗ + d(x, t, u∗ ), (6) Define in the extended space (x, dˆi ) the manifold
where d(x, t, u∗ ) = fe+(GG e ∗ −I)u∗ + ξ.e Through the above Mi = {(x, dˆi ) ∈ Rn | di − dˆi − βi = 0}, (9)
rearrangement, the original system (1) is restructured as the where βi (x, t) is a continuous function to be specified. The
combination of two parts, a trivial linear system ẋ = u∗ re- motivation for this definition is described as follows. The
ferring to the auxiliary system, and d, which can be viewed as dynamics of (8) restricted to the manifold Mi (provided it
an uncertainty term to the auxiliary system. Thus the original is invariant) is described by
problem is reformulated to the adaptive control problem for a
N
linear system with time-varying uncertainties, which is stated X
as designing a locally asymptotically stabilizing law u∗ for ẋ = u∗ + (dˆi + βi )ψi + . (10)
(6), with d unknown. i=1
To tackle the stated problem, we utilize the weighted basis Note from (10), the unknown vector di is excluded from the
functions to approximate d in the control system (6) at each expression of ẋ, which is important in the controller design
time instant as process since di cannot appear in the control law.
N However, (10) is equivalent to (6) only when the system
dynamics stay in the manifold Mi . By defining the off-the-
X
d(x, t, u∗ ) = di ψi (x, t) + , (7)
i=1
manifold variable
where di is constant and ψi consists of x and t, and , referring zi = di − dˆi − βi , (11)
to the approximation error, describes the deviation between the where zi ∈ Rn×1 and βi ∈ Rn×1 , zi = 0 implies that the
uncertainty d and the weighted basis functions. Substituting system dynamics stay in the manifold Mi . With the off-the-
(7) into (6) yields manifold variable, the state equation is transformed to
N
X N
X
ẋ = u∗ + ẋ = u∗ + zi + dˆi + βi ψi + ,


di ψi (x, t) + . (8) (12)
i=1 i=1

2475-1456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 01:40:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCSYS.2020.2997600, IEEE Control
Systems Letters

PN
where i=1 zi ψi represents the estimation error of the system By substituting (13) into (12), one obtains
uncertainty. Thus, the original control problem can be restated N
as defining an asymptotically stabilizing law u∗ for both the ẋ = −Kx − γsgn(x) +
X
zi ψi + , (16)
system state and the estimation error.
i=1
Remark 4. As mentioned in Remark 1, by selecting a suf- and the substitution of (16) into the derivative of (11) gives
ficiently large N ,  in (10) is negligible [7], [20]. On the   
other hand, taking  into account will reduce the number −ψ PN z ψ +  , σ = 1
i j=1 j j
of basis functions and thus, increase the efficiency of the żi = gi,σ = (17)
estimation process. To address both cases, in what follows, −xψi . σ=2

FATII controllers are proposed respectively when  is not (see
> >
By introducing y = x> z1> · · · zN
 
Section II-B1) and is negligible (see Section II-B2). , (16), (17) can
be written in the augmented state space form as ẏ = f (y, t),
1) Controller design when  is not negligible: When the the subsystem of which is described as
number of basis function N is selected small (seeking for a
low computational load) and thus  is not negligible, define ẏ = fσ (y, t), σ = 1, 2 (18)
the following controller where
 PN 
N
X −Kx − γsgn(x) + i=1 zi ψi + 
u∗ = −Kx − γsgn(x) − (dˆi + βi )ψi ,  g1,σ (y, t) 
fσ (y, t) =  . (19)
 
i=1 ..
(    . 
˙ −xψ̇i + Kx + γsgn(x) ψi , kxk2 > δ gN,σ (y, t)
d̂i =
xψi , kxk2 ≤ δ According to (14) and (15), define Fσ (x, t) = F[fσ ](x, t) as

xψi , kxk2 > δ the Filippov regularization of the subsystem (18), satisfying
βi = (13) that Fσ (y, t) ⊆ Fσ0 (y, t), where
0, kxk2 ≤ δ
 PN 
{−Kx + i=1 zi ψi + } − γF[sgn](x)
where δ is an arbitrary positive constant, γ > E1 , K = K1 +  {g1,σ } 
K2 + I, and K1 , K2 are positive definite matrices satisfying Fσ0 (y, t) = 

.

.
that λmin (K2 ) ≥ E22 /δ 2 .  .. 
Remark 5. It can be seen that the closed-loop system for- {gN,σ }
mulated by the plant (12) and controller (13) is a switched To study the generalized solutions of the closed-loop system
nonsmooth system. Thus, the traditional Lyapunov techniques formulated by (12) and (13) based on its afore-generated
are not applicable. To address this problem, the generalized subsystems, one constructs the following Lyapunov candidate
Lasalle-Yoshizawa theorem in [22] is employed, where the function
main idea is stated as follows: one can establish asymptotic N
properties for the generalized solutions of the switched nons- 1 > 1 1X >
V = y y = x> x + z zi . (20)
mooth system by using those of its subsystems. 2 2 2 i=1 i
In addition, for a differential equation ẋ = h(x, t) with
discontinuous right-hand side, an absolutely continuous func- Since V , expressed by (20), is smooth, the Clarke gradient
tion x(t) is called a generalized solution (Filippov solution) is reduced to the standard gradient, as ∂V = {y}. It should
to it on [t0 , t1 ] if ẋ ∈ F[h](x, t), where be noted that x> F[sgn](x) = {kxk1 } [22]. Thus, a bound
on the generalized time derivative of the Lyapunov candidate
function V˙ σ (σ = 1, 2) can be expressed as
\ \  
F[h](x, t) = co h B(x, ρ) − N , t , (14)
ρ>0 µ(N )=0
V˙ σ = max y> q
q∈Fσ (y,t)
represents the Filippov regularization [23]. In (14), co denotes  N 
the convex closure, B(x, ρ) the open ball of radius ρ centered
X
> >
T ≤ max y q=x − Kx + z i ψi + 
at x, and µ(N )=0 the intersection over all sets N of q∈Fσ0 (y,t)
i=1
Lebesgue measure zero. It has the following property [24] N
X
−γkxk1 + zi> gi,σ . (21)
F[h1 + h2 ](x, t) ⊆ F[h1 ](x, t) + F[h2 ](x, t), (15)
i=1

which will be used in the following For both σ = 1 and σ = 2, we need to analyze whether the
corresponding V˙ σ is negative semidefinite.
Theorem 1. Every maximal solution of the Filippov regular-
When σ = 1, by substituting the expression of gi,σ as (17)
ization of the switched nonsmooth system formulated by (12)
into (21), one obtains
and (13), regardless of the initial condition, is complete [25],
bounded, and satisfies limt→∞ x(t) = 0.  N 
˙
Vσ ≤x >
− Kx +
X
zi ψi + 
Proof. Define the state-dependent switching signal σ(x), i=1
where σ = 1 represents kxk2 > δ, and σ = 2 denotes N
X N
>  X 
kxk2 ≤ δ. The following steps are based on the notation of − zi ψi zi ψi +  − γkxk1 ,
the switching signals. i=1 i=1

2475-1456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 01:40:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCSYS.2020.2997600, IEEE Control
Systems Letters

a rearrangement of which gives where K is a positive definite matrix satisfying λmin (K) > 14 .
The stability of the proposed controller (27) is proved in the
V˙ σ ≤ −x> (K1 + K2 )x − γkxk1 + >  following
 N N N 
1 X X X
Theorem 2. The closed loop system, formulated by (10) and
− ( zi ψi )>( zi ψi )+2( zi ψi )> +> 
2 i=1 (27), is asymptotically stable, and the uncertainty estimation
i=1 i=1
 N N N  converges to the actual uncertainty asymptotically.
1 > >
X X
>
X
− x x−2x zi ψi +( z i ψi ) ( zi ψi ) Proof. The substitution of (27) into (26) yields
2 i=1 i=1 i=1
  N
X
1 ẋ = −Kx + zi ψi . (28)
− x> x − 2x>  + >  , (22)
2 i=1

where K = K1 + K2 + I and K1 , K2 are positive-definite To prove the stability of the closed-loop system, the Lyapunov
matrices. One can further simplify (22) as candidate function can be formulated as
N
!
1 1
V˙ σ ≤ −x> (K1 + K2 )x + kk2 − γkxk1 − kx − k2
2 2
X
V = x> x + zi> zi , (29)
2 2
2 2 i=1
N N
1 X 1 X the derivative of which is calculated as
− z i ψi − x − zi ψi + 
2 i=1 2 i=1 XN
2 2 >
V̇ = x ẋ + zi> żi . (30)
≤ −x> (K1 + K2 )x + kk22 i=1
≤ −λmin (K1 )kxk22 − λmin (K2 )kxk22 + E22 . (23) The derivative of zi is computed as
As λ1 = λmin (K1 ) > 0, λmin (K2 ) ≥ E22 /δ 2 , and σ(x) = 1 ˙ ∂βi ∂βi
żi = −d̂i − ẋ −
(kxk2 > δ), one obtains V˙ σ ≤ −λ1 kxk22 . ∂x ∂t
When σ = 2, by substituting (17) into (21), the generalized  N
X 
time derivative of the Lyapunov candidate function V˙ σ is = xψ̇i − Kxψi − xψ̇i − ψi − Kx + zj ψj
transformed into j=1
N
! N N
X 
˙
Vσ ≤x >
−Kx +
X
zi ψi +  − γkxk1 −
X
zi> xψi = −ψi zj ψj . (31)
i=1 i=1 j=1

≤ −x> Kx + x  − γkxk1 . >

(24) By substituting (28) and (31) into (30), one obtains
> !> N
The Cauchy-Schwarz inequality gives x  ≤ kxk2 kk2 , and N
! N
!
X X X
>
as k.k2 ≤ k.k1 , x>  ≤ kxk1 kk1 . Since K = K1 + K2 + I, V̇ = −x Kx − zi ψi − zi ψi zi ψi
one obtains i=1 i=1 i=1
1
V˙ ≤ −x>K x−(γ −kk )kxk ≤ −λ kxk2 −(γ −kk )kxk .
σ 1 1 1 1 2 1 1 = −x Kx + x> x
>
4
As kk1 ≤ E1 , by selecting γ > E1 , V̇ ≤ when −λ1 kxk22 N
!> N
!
σ = 2.
X 1 X 1
− zi ψi − x zi ψi − x
Thus, in conclusion, for both cases when σ = 1 and σ = 2, i=1
2 i=1
2
V˙ σ ≤ −λ1 kxk22 . (25) 
1
 N
X 1
2

≤ − λmin (K) − kxk22 − zi ψi − x , (32)

According to the generalized LaSalle-Yoshizawa Theorem 4 2
i=1 2
[22, Theorem 3], for switched nonsmooth systems with finite
number of subsystems being active in a small neighborhood where k.k2 represents the l2 norm for a vector.
of the state space, all maximal generalized solutions of the Since λmin (K) > 14 , V̇ is negative semi-definite, and
switched nonsmooth system formulated by (16) and (17) are therefore, x, zi are bounded. According to (28) and (31), as
complete [25], bounded, and satisfy limt→∞ x = 0. all components of ẋ and żi are bounded, so are ẋ and żi .
Moreover, it can be obtained from (32) that
2) Controller design when  is negligible: When there are a
1
sufficient number of basis functions,  is negligible, and thus, V̈ = −2x> K ẋ + x> ẋ
the plant can be simplified from (12) into 2
X N >  X N 
N 1 1
X −2 zi ψi − x (żi ψi + zi ψ̇i ) − ẋ ,
ẋ = u∗ + zi + dˆi + βi ψi .


(26) 2 2
i=1 i=1
i=1
the components of which are proved bounded. Therefore, so is
The controller is designed as V̈ , implying that V̇ is uniformly continuous. Thus, one obtains
N
X V̇ → 0, indicating
∗
u = −Kx − (dˆi + βi )ψi , N N
X X
i=1 lim x = 0, lim di ψi = lim (dˆi + βi )ψi , (33)
˙ t→∞ t→∞ t→∞
d̂i = −xψ̇i + Kxψi , i=1 i=1

βi = xψi , (27) according to the Barbalat’s lemma [26, Lemma 4.3].

2475-1456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 01:40:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCSYS.2020.2997600, IEEE Control
Systems Letters

III. C ASE STUDY B. Nonholonomic system

In this section, we test the validity of the constructed FATII The control of nonholonomic systems is difficult because,
controllers applied on a chaotic system and a nonholonomic according to the Brockett’s condition [28], there exists no
system. Note that the robustness of the proposed controllers continuous state feedback asymptotically stabilizing such type
are also demonstrated in the simulations based on the fact of systems at the equilibrium.
that the main portions of systems are treated as time-varying To demonstrate the feasibility of the proposed FATII con-
uncertainties. troller on nonholonomic systems, a rolling ball on a plane is
selected. The kinematics of the rolling system is described
A. Chaotic system by Montana’s equations [29], [30], as ẋ = G(x)u, where
Dealing with chaotic motions of the libration angle of x = (ub , vb , uo , vo , ψ) and u = (ωx , ωy , ωz ) define respec-
a satellite is a significant research topic in the control of tively the configuration and the angular velocity of the rolling
spacecrafts ball, and
The satellite system can be modeled as [27] 
0 R 0


C φ̈ + cφ̇ + ζ(φ, t) = Mc , (34)  −R 0 0

G(x) = − sin ψ/ cos vo − cos ψ/ cos vo 0  . (35)
 
where ζ(φ, t) = 3ω 2 (B − A) sin φ cos φ + µm iIr−3 (2 sin φ  − cos ψ sin ψ 0
sin ωc t + cos φ cos ωc t), and φ is the the libration angle, Mc − sin ψ tan vo − cos ψ tan vo −1
the input, c, ωc , A, B, lm , I, r, i the constant parameters.
The motion equation (34) can be written in the state space In the simulation, the gain matrix K in the controller (13)
form as ẋ = f + gu where x = (φ, φ̇) is the state vector and is selected as diag(30, 10, 6, 5, 5). The trajectories for the
u = Mc the input. The gain matrix K in the controller (27) configuration of the rolling ball, including its position ub (t),
is selected as diag(20, 10) and matrix G∗ as [1, 1] such that vb (t), and its orientation uo (t), vo (t), ψ(t), converge to zero,
(A.2) is satisfied (see Appendix). The values of the constant as illustrated in Fig. 3. The initial conditions of the system
parameters in the model are selected the same as in [27]. Note state ub (t), vb (t), uo (t), vo (t), and ψ(t) are specified as 1m,
that ζ(φ, t) can be viewed as the system uncertainty which 1m, πrad, πrad, and πrad. The estimation errors between the
does not appear in the controller. auxiliary and the actual systems converge to zero, as illustrated
The initial conditions of the system state φ and φ̇ are spec- in Fig. 4. The initial values of the control parameters dˆi (t) are
as π rad, π rad/s. The estimation error (e1 (t), e2 (t))) =
ified P chosen to be zero. The input signals are also shown by Fig. 4.
N
d − i=1 di ψi between the auxiliary and the actual systems
2.0
converge to zero, as illustrated in Fig. 2, where i = 1, 2. The ub 3 uo
1.5
input signals are also shown by Fig. 2. Note that most works vo

Orientation
vb
Position

1.0 2
y
in the literature suppressed the vibration of the libration angle
0.5 1
whlie the FATII method completely eliminates it.
0.5 1.0 1.5 2.0 2.5 3.0 3.5
0.5 0.5 1.0 1.5 2.0 2.5 3.0 3.5
6 1 2 3 4 5 Time (s) Time (s)
Angular velocity
Libration angle

2
4 4
6
Figure 3. Trajectories for the position (left) and the orientation (right) of a
2 rolling ball under the FATII control.
8
10
1 2 3 4 5 12
Time (s) Time (s) 50
80 40 e1 20 wx
Estimation error

e2 40
Velocity input

60 20
Figure 1. Trajectories for the libration angle (left) and the angular velocity e3 30
10 wy
40 wz
(right) of a satellite under the FATII control. 20 e4 20 0
20 40 e5 10
0 0.2 0.4 0.6 10 0 0.2 0.4 0.6
0.5 1.0 1.5 2.0 2.5 3.0 3.5
20 0.5 1.0 1.5 2.0 2.5 3.0 3.5
20 10
150 e1
Estimation error

100 Time (s) Time (s)

e2
Input signal

1 2 3 4 5
50 20
40 Figure 4. Uncertainty estimation (left) and the input signals (right) under the
1 2 3 4 5
50 60 FATII control.
100 80
100
Time (s) Time (s)
IV. C ONCLUSIONS
Figure 2. Uncertainty estimation (left) and the input signals (right) under the
FATII control. The FATII control method has been proposed for a wide
class of nonlinear systems. The FATII method has the follow-
Note that a sliding mode control based strategy employing ing features. Firstly, it is model-free and thus, is applicable to
recursive least squares for nonlinear terms estimation [4], [5] a wide range of systems. Secondly, the design of the FATII
can also be used for this system, whereas the PE condition controller is based on a robust adaptive approach (FAT) and
is required to guarantee that dˆ converges to the actual un- therefore can reject the effect of the system uncertainties or
certainty. In comparison, the PE condition is dispensable for external disturbances to the control system. Thirdly, unlike
the proposed controller. However, the controller in [4], [5] can other model-free methods such as the soft computing tech-
deal with the situation that time-delay is included in the output niques, the stability for systems under the FATII control has
signal. been well established.

2475-1456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 01:40:12 UTC from IEEE Xplore. Restrictions apply.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/LCSYS.2020.2997600, IEEE Control
Systems Letters

The following issues need to be clarified in the future work. [5] ——, “Sliding mode observer for fault reconstruction of time-delay and
Firstly, a unified way of selecting G∗ needs to be developed. sampled-output systems–a time shift approach,” Automatica, vol. 106, pp.
390 – 400, 2019.
Secondly, we chose the Fourier functions to for estimating the [6] H. Hu, S. Song, and C. L. P. Chen, “Plume tracing via model-free
variation term d. However, the advantages and disadvantages reinforcement learning method,” IEEE Transactions on Neural Networks
for this type of basis function are not analyzed and more and Learning Systems, vol. 30, no. 8, pp. 2515–2527, Aug 2019.
[7] A.-C. Huang and Y.-C. Chen, “Adaptive sliding control for single-link
candidates need to be investigated. Thirdly, the delays in flexible-joint robot with mismatched uncertainties,” IEEE Transactions
the output signal need to be taken into account. Also, more on Control Systems Technology, vol. 12, no. 5, pp. 770–775, 2004.
systems will be tested under the FATII based control and [8] Y. Bai, M. Svinin, and M. Yamamoto, “Adaptive trajectory tracking
control for the ball-pendulum system with time-varying uncertainties,”
experimental works will be conducted. in 2017 IEEE/RSJ International Conference on Intelligent Robots and
Systems (IROS), Sep. 2017, pp. 2083–2090.
A PPENDIX [9] ——, “Function approximation based control for non-square systems,”
SICE Journal of Control, Measurement, and System Integration, vol. 11,
It should be noted that there is no unique way for defining no. 6, pp. 477–485, 2018.
G∗ in (4). Depending on the estimation of control systems by [10] M.-C. Chien and A.-C. Huang, “Adaptive control for flexible-joint elec-
trically driven robot with time-varying uncertainties,” IEEE Transactions
(3), it can be selected based on the following considerations. on Industrial Electronics, vol. 54, no. 2, pp. 1032–1038, 2007.
Firstly, for the estimated control system (3) whose tangent [11] S. Khorashadizadeh and M. M. Fateh, “Uncertainty estimation in robust
linearization preserves controllability, one can design G∗ as tracking control of robot manipulators using the fourier series expansion,”
Robotica, vol. 35, no. 2, pp. 310–336, 2017.
follows. The linearization of (3) at the equilibrium (xe , ue ), [12] M. M. Zirkohi, “Direct adaptive function approximation techniques
without the consideration of ξ,e gives based control of robot manipulators,” Journal of Dynamic Systems,
Measurement, and Control, vol. 140, no. 1, 2018.
ẋ = Ax + Bu, (A.1) [13] J.-W. Liang, H.-Y. Chen, and Q.-W. Wu, “Active suppression of pneu-
matic vibration isolators using adaptive sliding controller with self-tuning
fuzzy compensation,” Journal of Vibration and Control, vol. 21, no. 2,
where A = ∂(f∂x +Gu)
and B = ∂(f∂u +Gu)
e e e e
xe ,ue xe ,ue
. One pp. 246–259, 2015.
selects G∗ in (4) as a constant matrix (see Section III-A) such [14] D. Ebeigbe, T. Nguyen, H. Richter, and D. Simon, “Robust regressor-free
control of rigid robots using function approximations,” IEEE Transactions
that on Control Systems Technology, 2019.
[15] L. Sonneveldt, E. Oort, Q. Chu, and J. Mulder, “Immersion and Invari-
Re[λ(An×n − Bn×m G∗m×n Kn×n )] < 0, (A.2) ance Based Nonlinear Adaptive Flight Control,” in Proc. of AIAA Guid-
ance, Navigation, and Control Conference, Toronto, Ontario, Canada,
where K is positive definite. Thus, the closed loop system August 2010.
[16] A. Astolfi and R. Ortega, “Immersion and invariance: a new tool for sta-
ẋ = (A − BG∗ K)x, (A.3) bilization and adaptive control of nonlinear systems,” IEEE Transactions
on Automatic control, vol. 48, no. 4, pp. 590–606, 2003.
formulated by the linearized system (A.1) and the state feed- [17] R. Ortega, A. Astolfi, and L. Hsu, “Immersion and invariance model
back portion u = −G∗ Kx of the FATII controller, is stable. reference adaptive control: new parameterizations for the problem,” in
42nd IEEE International Conference on Decision and Control (IEEE Cat.
The effect of the variation between the linearized system (A.1) No.03CH37475), vol. 4, 2003, pp. 3239–3243.
and the original system (1) can be eliminated by the rest part [18] Y. Lv, J. Na, and X. Ren, “Online h-infinity control for completely
of the FATII controller. Note that the stability of (A.3) also unknown nonlinear systems via an identifier-critic-based adp structure,”
International Journal of Control, vol. 92, no. 1, pp. 100–111, 2019.
indicates the controllability of a linear system [19] S. Baldi, “Cooperative output regulation of heterogeneous unknown sys-
tems via passification-based adaptation,” IEEE Control Systems Letters,
ẋ = Ax + BG∗ u∗ , (A.4) vol. 2, no. 1, pp. 151–156, 2018.
[20] A.-C. Huang and Y.-S. Kuo, “Sliding control of non-linear systems con-
where u∗ ∈ Rn is viewed as the input. It is because the taining time-varying uncertainties with unknown bounds,” International
selection of u∗ = −Kx renders (A.4) to the stable form (A.3). Journal of Control, vol. 74, no. 3, pp. 252–264, 2001.
[21] A. Izadbakhsh, “Fat-based robust adaptive control of electrically driven
As (A.4) is the linearized system of (5) at the equilibrium, the robots without velocity measurements,” Nonlinear Dynamics, vol. 89,
controllability of (A.4) implies the local controllability of the no. 1, pp. 289–304, 2017.
restructured system (5) at the equilibrium [31]. [22] R. Kamalapurkar, J. A. Rosenfeld, A. Parikh, A. R. Teel, and W. E.
Dixon, “Invariance-like results for nonautonomous switched systems,”
Secondly, for control systems whose tangent linearization IEEE Transactions on Automatic Control, vol. 64, no. 2, pp. 614–627,
does not preserve controllability, such as the nonholonomic 2018.
systems (see Section III-B), one may select matrix G∗ as the [23] A. F. Filippov, Differential equations with discontinuous righthand sides:
e as G∗ = (G e > W G)
e −1 G e >W . control systems. Springer Science & Business Media, 2013, vol. 18.
weighted pseudoinverse of G [24] B. Paden and S. Sastry, “A calculus for computing filippov’s differential
The constant matrix W is designed such that system (5) is inclusion with application to the variable structure control of robot
controllable. The controllability proof of the restructured sys- manipulators,” IEEE transactions on circuits and systems, vol. 34, no. 1,
pp. 73–82, 1987.
tem (5) for typical nonholonomic systems such as the unicycle [25] R. Sanfelice, R. Goebel, and A. Teel, “A Feedback Control Motivation
system and the spherical rolling robot can be found in [30]. for Generalized Solutions to Hybrid Systems,” in Proc. of International
An alternative way is to select the nilpotent approximation for Workshop on Hybrid Systems: Computation and Control, Santa Barbara,
CA, USA, March 29-31, 2006.
nonholonomic systems as the nominal plant, and then design [26] J.-J. E. Slotine, W. Li et al., Applied nonlinear control. Prentice hall
G∗ through the process from (A.1) to (A.2). Englewood Cliffs, NJ, 1991, vol. 199, no. 1.
[27] A. Bobtsov, N. Nikolaev, and O. Slita, “Adaptive control of libration
angle of a satellite,” Mechatronics, vol. 17, no. 4, pp. 271 – 276, 2007.
R EFERENCES [28] R. W. Brockett, “Asymptotic stability and feedback stabilization,” in
Differential Geometric Control Theory. Birkhauser, 1983, pp. 181–191.
[1] R. Lozano and I. Fantoni, Non-Linear Control for Underactuated Me- [29] R. Murray, Z. Li, and S. Sastry, A Mathematical Introduction to Robotic
chanical Systems. Springer, 2001. Manipulation. Boca Raton: CRC Press, 1994.
[2] M. W. Spong, “Partial feedback linearization of underactuated mechanical [30] Y. Bai, M. Svinin, Y. Wang, and E. Magid, “Function approximation
systems,” in Proc. IEEE Int. Conference on Intelligent Robots and technique based control for a class of nonholonomic systems,” in Proc.
Systems, Munich, Germany, September 12–16 1994, pp. 314–321. of IEEE IEEE/SICE International Symposium on System Integration,
[3] P. V. Kokotovic, “The joy of feedback: nonlinear and adaptive,” IEEE Honolulu, Hawaii, USA, January 2020.
Control Systems Magazine, vol. 12, no. 3, pp. 7–17, June 1992. [31] J.-M. Coron, Control and Nonlinearity. Boston, MA, USA: American
[4] H. L. C. P. Pinto, T. R. Oliveira, and L. Hsu, “Fault reconstruction for Mathematical Society, 2007.
delay systems via least squares and time-shifted sliding mode observers,”
Asian Journal of Control, vol. 21, no. 1, pp. 52–69, 2019.

2475-1456 (c) 2020 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 01:40:12 UTC from IEEE Xplore. Restrictions apply.