IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL.

22, 2025 15139

Constrained Tracking Control of MIMO Nonlinear

Systems With Discontinuous References and
Unknown Dynamics
Jin-Xi Zhang , Member, IEEE, Wei-Li Qi , and Tianyou Chai , Life Fellow, IEEE

Abstract—This paper is dealt with the tracking control prob- they fail to be applied to discontinuous references, which are
lem for the multi-input multi-output (MIMO) block-triangular common in vehicle guidance, industrial electronics and robotic
nonlinear systems with output constraints under discontinuous applications. The related tracking control approaches are effec-
references. It is focused on the case where the system exhibits tive for the regular discontinuity of the reference, the bounded
inherent nonlinearities, e.g., radically unbounded nonlinearities, nonlinearities of the system, the known system dynamics, or the
and totally unknown dynamics. This renders the existing solu- known bounding functions of the nonlinearities. To overcome
tions infeasible. To surmount this challenge, a novel hybrid the above challenges, a low-complexity hybrid tracking control
tracking control strategy composed of a robust decoupling strategy composed of a robust decoupling constrained controller
constrained controller and a proportional controller is devised in and a proportional controller is developed in this paper. It steers
this paper. It guarantees that the system outputs evolve within the system outputs to track the discontinuous references without
the prescribed constraint bands and track the discontinuous constraint violation. The result of a comparative experiment on
references with the tunable settling time and accuracy. Moreover, a 2DSFL robot show the effectiveness and advantage of the
the controller shows a significant simplicity. No attempt is made proposed approach.
for parameter identification, function approximation, disturbance
estimation, derivative calculation or command filtering, despite Index Terms—Discontinuous references, MIMO nonlinear sys-
the unknown system dynamics and the recursive control design. tems, output constraints, tracking control, unknown dynamics.
The theoretical findings are validated by a comparative experi-
I. I NTRODUCTION
ment on a 2-DOF serial flexible link (2DSFL) robot.

Note to Practitioners—MIMO nonlinear systems are widely

present in practice, e.g., robotic manipulators, ground vehicles
N UMEROUS practical plants are multi-input multi-output
(MIMO) nonlinear systems, which include but are not
limited to robotic manipulators, ground vehicles, surface
and aircrafts. For operational specifications and safety consid- vehicles, aircraft, and spacecraft. From the perspectives of
erations, they are always subject to output constraints. On the operational specifications and safety considerations, these sys-
other hand, building an exact mathematical model for a practical tems are always subject to output constraints. For instance, the
system is unfulfillable. The existing constrained control methods
trajectory of a ground vehicle has to be constrained within
for uncertain MIMO nonlinear systems work for the case where
the control distribution matrix is norm-bounded or the control a safe boundary, on the purpose of obstacle and collision
distribution matrix, additive nonlinearity, Jacobian matrix or avoidance. A widely adopted constraint-handling method is the
disturbance bound is known. Moreover, the control algorithms barrier Lyapunov function (BLF) [1], [2]. On the other hand,
are complex due to the dependence on parameter identification, it is well documented that building an exact mathematical
function approximation and disturbance estimation. Additionally, model for a practical system is unfulfillable, especially for a
complex system (e.g., the MIMO nonlinear system with inputs
Received 13 December 2024; revised 13 February 2025; accepted 29 April coupling). Therefore, model uncertainties are inevitable in the
2025. Date of publication 2 May 2025; date of current version 16 May control design.
2025. This article was recommended for publication by Associate Editor Up to now, a variety of output-constrained control solutions,
G. Michieletto and Editor M. Dotoli upon evaluation of the reviewers’
comments. This work was supported in part by the National Natural based on BLFs, for the MIMO nonlinear systems under model
Science Foundation of China under Grant 62473089, in part by the uncertainties have been reported [3], [4], [5], [6], [7], [8], [9],
Research Program of Liaoning Liaohe Laboratory under Grant LLL23ZZ- [10], [11], [12], [13], [14]. For the Euler-Lagrange systems,
05-01, in part by the 111 Project 2.0 of China under Grant B08015, a series of approximation-based controllers were developed
in part by the National Key Research and Development Program of China
under Grant 2022YFB3305905, in part by the Xingliao Talent Program [3], [4], [5], [6], [7], [8], in which neural networks [3], [4],
of Liaoning Province of China under Grant XLYC2203130, in part by [5], [6] or fuzzy logic systems [7], [8] were employed to
the Natural Science Foundation of Liaoning Province of China under deal with the unknown nonlinearities of the plants. To be
Grant 2023JH26/10200011 and Grant 2024JH3/10200012, in part by the specific, the constant and symmetric constraint [3], [7], the
Open Research Project of the State Key Laboratory of Industrial Con- constant yet asymmetric constraint [4], [5], and the time-
trol Technology of China under Grant ICT2024B12, and in part by
the Fundamental Research Funds for the Central Universities of China varying and asymmetric constraint [6], [8] are addressed,
under Grant N2424016 and Grant N25YJS002. (Corresponding author: respectively. In the presence of nonparametric uncertainties,
Tianyou Chai.) an approximation-free robust control law was constructed [9].
The authors are with the State Key Laboratory of Synthetical Automation For the block-triangular nonlinear systems, an adaptive fixed-
for Process Industries, Northeastern University, Shenyang 110819,
China (e-mail: zhangjx@[Link]; qiweili@[Link]; time control strategy with the universal BLF was proposed
tychai@[Link]). [10]; an adaptive control scheme with the global performance
Digital Object Identifier 10.1109/TASE.2025.3566443 guarantee was devised [11]; a neural network control approach
1558-3783 © 2025 IEEE. All rights reserved, including rights for text and data mining, and training of artificial intelligence and
similar technologies. Personal use is permitted, but republication/redistribution requires IEEE permission.
See [Link] for more information.
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.
 15140 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 22, 2025

with the given performance specifications was developed [12]; of the system [19], the known system dynamics [15] and the
an event-triggered neural network control policy [13] and a known bounding functions of the system nonlinearities [22]
quantized neural network control scheme [14] were designed are admissible.
for economizing resource. The above observations reveal that thus far, both the
Despite the above-mentioned achievement in output- low-complexity control problem for MIMO systems with
constrained control of MIMO nonlinear systems, it should inherent nonlinearities and unknown dynamics under output
be noted that the admissible nonlinearity is limited in the constraints and the tracking control problem for unknown
sense that the nonlinear control distribution matrix should inherent MIMO nonlinear systems with irregular discontinuous
be norm-bounded [3], [8], [10], [14]. Moreover, partial key references remain open. This paper presents a solution via a
model information is necessary, e.g., the known control dis- novel hybrid tracking control strategy composed of a robust
tribution matrix [6], [11], [13], additive nonlinearity [10], constrained controller and a proportional feedback controller.
[11], Jacobian matrix [4], [7] and disturbance bound [4], Its superiority is threefold.
[7]. Additionally, the complexity of the control algorithm 1) Both the additive nonlinearities and the multiplicative
is also noteworthy. Adaptive mechanisms were employed to nonlinearities of the system are allowed to be radically
identify model parameters [10], [11]; disturbance observers unbounded, rather than norm-bounded [3], [8], [10],
[11] were adopted to estimate disturbances; neural networks [14], [19]. The irregular discontinuity of the references
[3], [4], [5], [6], [12], [13], [14] or fuzzy logic systems [7], is admissible, unlike the known and fixed quantization
[8] were used to approximate unknown nonlinearities. They step [23].
increase the computational burden and complicate the con- 2) Specific information of neither the system nonlinearities
troller implementation, as a large number of parameters should and external disturbances nor their bounding functions
be updated online. Furthermore, the usage of neural/fuzzy or bounds is required, unlike [4], [6], [7], [10], [11],
systems necessitates a compact set condition, which was not [13], [15], [22].
rigorously warranted in many references. Last but not least, all 3) No effort needs to be paid for parameter identification
of the aforesaid results work for the case where the references [10], [11], [25], function approximation [3], [4], [5],
are continuously differentiable in time. [6], [7], [8], [12], [13], [14], [26], [27], [28], [29],
Reference tracking is a typical control task, and in vari- disturbance estimation [11] and command filtering [12],
ous scenarios, the reference is piecewise continuous rather [21], leading to a control simplicity.
than uniformly continuous. A representative example is the The rest of the paper is organised as follows. Section II for-
guidance and control of a ground vehicle [15], in which mulates the problem under consideration. Section III presents
the reference trajectory generated online includes geomet- the controller development. Section IV gives the theoretical
ric singularities due to sharp corners or obstacle avoidance. analysis to validate its effectiveness. In Section V, an exper-
This leads to discontinuity in position, orientation or cur- iment is conducted to illustrate the theoretical result. In the
vature of the target trajectory. Other examples incorporate end, Section VI concludes this paper.
but are not exclusive to industrial electronics [16], [17]
and robotic applications [18]. It is well documented that II. P ROBLEM F ORMULATION
the direct application of the conventional control methods A. System Description
to the case of discontinuous references may lead to per- Consider a sort of MIMO nonlinear systems in the block-
formance deterioration [15], instability [19], or even safety triangular form:
accidents [20].
< ẋi = Fi (xi ) + Gi (xi )xi+1 + Di (t), i = 1, · · · , n − 1,
8
Intuitively, the rough reference can be filtered to generate
a smooth trajectory [21]. However, the resulting reference ẋ = Fn (xn ) + Gn (xn )u + Dn (t), (1)
may differ widely from the original reference, due to active : n
y = x1 ,
damping of the nonzero frequency. Moreover, if the derivatives
of the original reference are unavailable, the fungible reference where xi = [xi1 , · · · , xim ]T ∈ Rm and xi = [x1T , · · · , xiT ]T ∈ Rim ,
is disable to exactly recover the real reference even without i = 1, · · · , n; xn represents the system state; u is a vector of the
discontinuity. In place of filtering, a unified lateral guidance control inputs; y = [y1 , · · · , ym ]T ∈ Rm is a vector of the system
technique based on dynamic and geometric approaches was outputs; Fi (xi ) ∈ Rm is a vector of the nonlinear functions and
devised [15], in which the online adapted virtual reference Gi (xi ) ∈ Rm×m is the nonlinear control distribution matrix,
is incorporated into the dynamic lateral guidance scheme. i = 1, · · · , n; Di (t) ∈ Rm , i = 1, · · · , n, are the external
It improves the control performance against the lateral error disturbances, which are locally Lipschitz-continuous in time
induced by the discontinuous reference. For the noninstanta- and norm-bounded.
neous impulsive fractional-order systems with discontinuous Assumption 1: [8], [10], [11], [12], [13], [14], [30], [31]
references, a P-type iterative learning control strategy was Both Fi (xi ) and Gi (xi ) are continuous with respect to xi , i =
proposed to accomplish accurate tracking [22]. It requires only 1, · · · , n.
a small number of iterations in a short time. By introducing Assumption 2: [10], [14], [30], [32], [33], [34], [35] Let
the impulsive dynamics to the prescribed performance control
1
Gi xi = Gi xi + GTi xi , i = 1, · · · , n,




design, an impulsive closed-loop system is constituted [19]. (2)
Under the discontinuous reference, fast accurate tracking con- 2
trol is achieved. By involving a smooth trajectory generator each of which is either positive definite or negative definite
to the funnel control law, the predefined tracking performance for any xi ∈
Rim . Without loss of generality, consider the case
with the quantized reference is realized [23]. Nonetheless, only where Gi xi is positive definite, i = 1, · · · , n.
the regular discontinuity of the reference (e.g., the known and Remark 1: Assumptions 1 and 2 warrant the existence
fixed quantization step) [23], [24], the bounded nonlinearities of solutions and the controllability of (1). Otherwise, the

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: CONSTRAINED TRACKING CONTROL OF MIMO NONLINEAR SYSTEMS 15141

system dynamics does not meet the local Lipschitz condition sense. Item 2) describes discontinuity of the reference, which
such that no solution exists; a singular control distribution can also be replaced with limt→tk− r j (t) = r j (tk ), j = 1, · · · , m,
matrix leads to the system uncontrollable. Diverse practical k = 0, 1, 2, · · · . Item 3) indicates that despite discontinuity,
plants fulfill Assumptions 1 and 2, which include but are the reference is still piecewise continuous and differentiable
not limited to the multiple-link robotic manipulator [36], the in time with the bounded derivative. This excludes the case of
marine engineering vessel [37] and the permanent magnet random references like noise. On the other hand, a reference
synchronous motor [38]. Moreover, their control distribution with the unbounded derivative is meaningless for a stable
matrices are positive-define, by feat of the inherent property tracking control system in which all the signals should be
of the inertia matrix. bounded. Besides, it is notable that the references need only
Remark 2: This study is concentrated on the case where to be available at each time instant during operation, and
the dynamics of (1) is highly unknown and exhibits inherent their derivatives are not necessarily accessible. In the related
nonlinearities. On the one hand, the knowledge of neither Fi (·), works, e.g., [3], [8], [11], and [12], however, the (high-order)
Gi (·) and Di (t), i = 1, · · · , n, nor their bounding functions or derivatives of the references have to be available even without
bounds is available, except for Assumption 2. On the other involving discontinuity.
hand, the general nonlinear dynamics are admissible, e.g., the Problem 1: For the MIMO nonlinear system in (1) subject
radically unbounded nonlinearity. to the output constraints in (3), develop a controller such that
the system outputs track the discontinuous reference vector
B. Output Constraints without constraint violation and all the signals in the control
system are bounded.
The evolution of the outputs of (1) needs to satisfy
y j (t) < y j (t) < y j (t), j = 1, · · · , m, ∀t ≥ 0, (3) III. C ONTROL D ESIGN
where y j (t) and y j (t) are the upper and lower boundaries, To solve problem 1, a novel hybrid tracking control scheme
consisting of a robust constrained controller and a proportional
respectively, which are allowed to be asymmetric. The initial
controller is developed in this section.
outputs are reasonably constrained by
On the purpose of (3), employ a group of barrier functions:
y j (0) < y(0) < y j (0), j = 1, · · · , m. (4) !
y j (t) − y j (t)
η1 j (t) = ln , j = 1, · · · , m. (6)
Assumption 3: [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], y j (t) − y j (t)
[11], [12], [13], [14], [20], [39] There exists no t ≥ 0 such
that To handle the inputs coupling of (1), the partial derivatives of
lim y j (t) = lim y j (t), j = 1, · · · , m. (5) (6) are involved:
t→t t→t
y j (t) − y j (t)
Furthermore, y j (t), y j (t), ẏ j (t) and ẏ j (t) are bounded for t ≥ 0. β1 j (t) = , j = 1, · · · , m. (7)
(y j (t) − y j (t))(y j (t) − y j (t))
Remark 3: Without (5), the constraint boundaries may be
infinitely close together, e.g., Further, let
y j (t) = e , y j (t) = −e .
−t −t
η11 (t) β11 (t)
2 3 2 3
6 .. 7 ..
η1 (t) = 4 . 5 , β1 (t) = 4 . 5. (8)
6 7
When time tends to infinity, the feasible solution to the above
constraint problem becomes an empty. On the other hand, if η1m (t) β1m (t)
the constraint boundary is unbounded, this is equivalent to
Then, the intermediate control law is designed by
the absence of constraints, making the output constraint no
sense. Besides, in various practical applications, the output α1 (t) = [α11 (t), · · · , α1m (t)]T = C1 β1 (t)η1 (t), (9)
constraint is generated by a trajectory planning device, i.e.,
the output constraint is determined by the user or designer. In where C1 = diag{c11 , · · · , c1m } denotes the control gain matrix
this case, it is feasible to plan differentiable constraints with with c1 j > 0, j = 1, · · · , m, being constants. Proceed with
bounded derivatives. zi (t) = xi (t) − αi−1 (t) = [zi1 (t), · · · , zim (t)]T , (10)
 
pi j − zi j (t)
C. Discontinuous Reference ηi j (t) = ln , j = 1, · · · , m, (11)
pi j + zi j (t)
On the basis of (3), the control objective for (1) is
that the system outputs track a reference vector, r(t) = 2pi j
βi j (t) = 2 , j = 1, · · · , m, (12)
[r1 (t), · · · , rm (t)]T ∈ Rm . pi j − z2i j (t)
Assumption 4: [9], [19], [20], [39] There hold 2
ηi1 (t)
3 2
βi1 (t)
3
1) y j (t) < r j (t) < y j (t), j = 1, · · · , m, t ≥ 0,
ηi (t) = 4 ... 5 , βi (t) = 4 ..
. 5, (13)
6 7 6 7
2) limt→t+ r j (t) = r j (tk ), j = 1, · · · , m, k = 0, 1, 2, · · · ,
k
3) |ṙ j (t)| < ∞, tk < t < tk+1 , j = 1, · · · , m, k = 0, 1, 2, · · · , ηim (t) βim (t)
where t0 = 0; tk+1 − tk ≥ µ, k = 0, 1, 2, · · · ; µ > 0 is a αi (t) = [αi1 (t), · · · , αim (t)] = Ci βi (t)ηi (t),
T
(14)
constant. for i = 2, · · · , n, one by one, where Ci = diag{ci1 , · · · , cim } with
Remark 4: Item 1) ensures a feasible reference tracking ci j > 0, j = 1, · · · , m, being constants; pi j > 0, j = 1, · · · , m,
under output constraints. Otherwise, steering the output to are constants which should meet
track a reference outside the constraint band based on the
operational specification and safety consideration makes no |zi j (0)| < pi j , i = 2, · · · , n, j = 1, · · · , m. (15)

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.
 15142 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 22, 2025

The proportional controller is designed recursively by Proof: Through (18), (6) and (7) are rewritten by
 
8
e1 (t) = x1 (t) − r(t), p1 j (t) − z1 j (t)
ˆ η1 j (t) = ln , j = 1, · · · , m, (22)
ρ1 (t) = −B1 e1 (t), p1 j (t) + z1 j (t)
ˆ
<
(16) 2p1 j (t)
e (t) = xi (t) − ρi−1 (t), i = 2, · · · , n, β1 j (t) = 2 , j = 1, · · · , m.
: i (23)
ˆ
ˆ
ρi (t) = −Bi ei (t), i = 2, · · · , n, p1 j (t) − z21 j (t)

where Bi = diag{bi1 , · · · , bim } stands for the control gain matrix Based on (9), there is
with bi j > 0 being a constant, i = 1, · · · , n, j = 1, · · · , m. Based α1 j (t) = c1 j β1 j (t)η1 j (t), j = 1, · · · , m. (24)
on (14) and (16), the final control law is obtained by
Under (21), differentiating (24) via (22) and (23) yields
u(t) = αn (t) + ρn (t). (17)
α̇1 j (t) = c1 j β̇1 j (t)η1 j (t) + c1 j β1 j (t)η̇1 j (t),
Remark 5: In the existing studies on constrained tracking for j = 1, · · · , m, where
control [3], [6], [7], [8], [9], [11], [12], [39], the output
constraint is in general transformed to the tracking error 2 ṗ1 j (t)
β̇1 j (t) = 2
constraint. By this means, the constraint satisfaction and p1 j (t) − z21 j (t)
the error stabilization are ensured by a unified controller

4p1 j (t) p1 j (t) ṗ1 j (t) − z1 j (t)ż1 j (t)
design. This however becomes infeasible under discontinuous −  2 , (25)
references in the recursive constrained control design, because 2 2
p1 j (t) − z1 j (t)
the discontinuous references result in jumps of the intermediate  
errors, which may cause violation of the constraints [3], z1 j (t) ṗ1 j (t)
η̇1 j (t) = − β1 j (t) ż1 j (t) − . (26)
[7], [9], [39]. Therefore, a hybrid tracking control scheme p1 j (t)
is developed in this paper, where the constrained controller
in (6)–(15) aims at guaranteeing the output constraints and In the case of (21), η1 j (t) in (22) and β1 j (t) in (23) are bounded
the proportional controller in (16) dedicates to stabilizing the over [0, τ), j = 1, · · · , m. By Assumption 3 and (18), we have
tracking errors under the discontinuous references. On the p1 j (t) ∈ L∞ , ṗ1 j (t) ∈ L∞ and p11j (t) ∈ L∞ , j = 1, · · · , m. Under
2 (t) is bounded over [0, τ). Hence, β̇1 j (t) in (25)
1
other hand, the conventional decoupling constrained control (21), p2 (t)−z
1j 1j
law adopts the identical control gain [9], [30], [40], [41], and η̇1 j (t) in (26) are bounded over [0, τ) under (20) and (21),
[42], [43], i.e., the entries of each Ci , i = 1, · · · , n, in (9) j = 1, · · · , m. Hence, α1 j (t) in (24) is bounded over [0, τ),
and (14) are same. In contrast, the independent control gains j ∈ {1, · · · , m}. Following the same line to analyze α̇i j (t) by
are admissible by our approach, which renders the controller (14), i ∈ {2, · · · , n}, j ∈ {1, · · · , m}, one can deduce that it is
design more flexible. The theoretical analysis below shows its also bounded on [0, τ), under (20) and (21). 
effectiveness. Theorem 1: Under Assumptions 1–5 and the initial condi-
Remark 6: It is noted from (6)–(17) that the control law tions in (4) and (15), the developed controller in (9)–(14), (16)
does not involve the specific information of the system model and (17) solves problem 1.
in (1) or the derivatives of the references and intermediate Proof: It starts from claiming that
control signals. Nonetheless, no effort is paid for parameter
estimation, function approximation, disturbance estimation or |zi j (t)| < pi j (t), i = 1, · · · , n, j = 1, · · · , m, ∀t ≥ 0, (27)
command filtering. Accordingly, a low-complexity and less which is to be established by contradiction. From (4) and (18),
demanding control algorithm is devised. there hold |z1 j (0)| < p1 j (0), j = 1, · · · , m. This together with
(15) yields (27) for t = 0. Note that the state variables of (1)
IV. T HEORETICAL A NALYSIS are uniformly continuous. The same holds for y j (t) and y j (t)
For ease of analysis, two lemmas are provided as follows. by Assumption 3, j = 1, · · · , m. Hence, z1 j (t) and p1 j (t) in
Lemma 1: Let (18) are uniformly continuous, j = 1, · · · , m. This means by
(22)–(24) that α1 j (t) is continuous as long as |z1 j (t)| < p1 j (t),
z1 j (t) = y j (t) − 0.5(y j (t) + y j (t)), j = 1, · · · , m. Then, z2 j (t) in (10) is continuous under the same
(

p1 j (t) = 0.5(y j (t) − y j (t)),

(18) condition, j = 1, · · · , m. Follow the same line to analyze zi j (t),
i = 3, · · · , n, j = 1, · · · , m. Then, it can be concluded that zi j (t)
for j = 1, · · · , m and is continuous in the case of |zk j (t)| < pk j (t), k = 1, · · · , i − 1,
j = 1, · · · , m. The above findings imply that if (27) is violated,
z1 (t) = [z11 (t), · · · , z1m (t)]T . (19) then there should exist t∗ > 0 such that
Then, (3) is equivalent to |z1 j (t)| < p1 j (t), j = 1, · · · , m, t ≥ 0. lim |zi j (t)| = lim∗− pi j (t), i ∈ {1, · · · , n}, j ∈ {1, · · · , m}, (28)
t→t∗− t→t
Proof: It is straightforward and thus omitted. 
Lemma 2: For each i ∈ {1, · · · , n}, each j ∈ {1, · · · , m} and and
any τ > 0, α̇i j (t) is bounded over [0, τ), if |zi j (t)| < pi j (t), i = 1, · · · , n, j = 1, · · · , m, t < t∗ . (29)
|żi j (t)| < ∞, t < τ, (20) Next, we suppose (28) with (29) and enumerate each scenario
in (28) for justification. For simplicity of notation, we omit
and there exists a constant δi j ∈ (0, pi j ) such that the time or state dependence of some variables in the sequel.
|zi j (t)| ≤ pi j − δi j , t < τ, (21) For ease of analysis, construct
1
where p1 j is time-varying and pi j is constant, i = 2, · · · , n. Vi = ηTi Ci ηi , i = 1, · · · , n. (30)
2
Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: CONSTRAINED TRACKING CONTROL OF MIMO NONLINEAR SYSTEMS 15143

Differentiating (30) yields where λ1 is the minimum eigenvalue of G1 . Through (45), we

obtain that for t < t∗ , V̇1 < 0 when
V̇i = ηTi Ci η̇i , i = 1, · · · , n. (31)
h1
From (10), (12), (18) and (23), the differential equations of kC1 β1 η1 k > . (46)
λ1
(11) and (22) are obtained by
  By (8) with (23), we have
zi j ṗi j
η̇i j = −βi j żi j − , i = 1, · · · , n, j = 1, · · · , m. v 0 !2 1
pi j
u
m
2c1 j p1 j
uX
kC1 β1 η1 k = t η21 j A
u
Note from (8), (10), (13) and (19) that
@
2 2
j=1
p 1j − z 1j
η̇i = −βi (żi − Ai ) , i = 1, · · · , n, (32) v 0
u
m
!2 1
where 2p1 j
uX
η21 j A,
u

z11 ṗ11 z1m ṗ1m
T ≥ c1 t @
2 2
(47)
A1 = ,··· , , (33) j=1
p 1j − z1j
p11 p1m
and Ai = 0, i = 2, · · · , n. Substituting (32) into (31) gives where c1 = min{c11 , · · · , c1m }. Under (29), there holds
V̇i = −ηTi Ci βi (żi − Ai ), i = 1, · · · , n. (34) pi j 1
≥ , i = 1, · · · , n, j = 1, · · · , m. (48)
p2i j − z2i j pi j
Step 1: At the outset, we analyze
Due to Assumption 3 and (18), there is a positive constant p1
V̇1 = −ηT1 C1 β1 (ż1 − A1 ). (35) so that
Differentiating (19) with (18) by (1) yields p1 j ≤ p1 , j = 1, · · · , m. (49)
ż1 = F1 + G1 x2 + D1 − 0.5(ẏ + ẏ). (36) With (48) for i = 1 and (49), (47) becomes
From (9), (10) and (14), there is kC1 β1 η1 k ≥ k1 kη1 k, t < t∗ , (50)
2c1
xi+1 = zi+1 + Ci βi ηi , i = 1, · · · , n − 1. (37) where k1 = p1 . It follows from (46) and (50) that for t < t∗ ,
Substituting (37) for i = 1 into (36) yields V̇1 < 0 if
h1
kη1 k > .
ż1 = F1 + G1 z2 + G1C1 β1 η1 + D1 − 0.5(ẏ + ẏ). (38) λ1 k1
Inserting (38) into (35) yields Note from (30) that
V̇1 = ηT1 C1 β1 H1 − ηT1 C1 β1G1C1 β1 η1 , ci ci
(39) kηi k2 ≤ Vi ≤ kηi k2 , i = 1, · · · , n, (51)
2 2
where
where ci = max{ci1 , · · · , cim }, i = 1, · · · , n. Hence, it is further
H1 = 0.5(ẏ + ẏ) − F1 − D1 + A1 − G1 z2 . (40) c h2
obtained that on [0, t∗ ), V̇1 < 0 as V1 ≥ 2λ12 k12 . This yields
1 1
The inequalities in (29) imply by Lemma 1 that kx1 k is
bounded for t < t∗ . On this basis, Assumption 1 ensures the c1 h21
 
boundedness of kF1 k and kG1 k on [0, t∗ ). Note by (18) with V1 (t) ≤ ζ1 := max V1 (0), 2 2 , t < t∗ .
2λ1 k1
Assumption 3 and (29) that kA1 k in (33) and kz2 k in (10) are
From (51) for i = 1, we further have kη1 k ≤ ζ1 /c1 , t < t∗ .
p
bounded on [0, t∗ ). Substituting the above facts together with
D1 ∈ L∞ and Assumption 3 into (40) yields kH1 k < ∞, t < t∗ . This in turn implies by (8) and (22) that
Let
|z1 j (t)| ≤ p1 j (t) − o1 j < p1 j (t), j = 1, · · · , m, t < t∗ , (52)
h1 = sup kH1 k. (41)
t∈[0,t∗ )
where o1 j is a positive constant, j = 1, · · · , m. On the
Since β1 in (8) and C1 in (9) are diagonal matrices, by (41), other hand, the boundedness of kx1 k on [0, t∗ ) together with
(39) is scaled by Assumption 4 implies by (16) that
V̇1 ≤ h1 kC1 β1 η1 k − ηT1 β1C1G1C1 β1 η1 , t < t∗ . (42) kρ1 k < ∞, t < t∗ , (53)
From (2), there is which enables us to analyze V̇n below.
Step 2: Continue to analyze V̇2 in (34):
Gi − GTi
Gi = Gi + , i = 1, · · · , n. (43) V̇2 = −ηT2 C2 β2 ż2 . (54)
2
Based on (43) for i = 1, (42) is rewritten by Computing the derivative of z2 in (10) via (1) yields
G1 − GT1
 
V̇1 ≤ h1 kC1 β1 η1 k − ηT1 β1C1 G1 + C1 β1 η1 . (44) ż2 = F2 + G2 x3 + D2 − α̇1 . (55)
2

Inserting (37) for i = 2 into (55) gives
Recall Assumption 2 and note that G1 − GT1 is a skew
symmetric matrix. Then, (44) is further scaled by ż2 = F2 + G2 z3 + G2C2 β2 η2 + D2 − α˙1 . (56)

V̇1 ≤ h1 kC1 β1 η1 k − λ1 kC1 β1 η1 k2 Substituting (56) into (54) leads to

= kC1 β1 η1 k (h1 − λ1 kC1 β1 η1 k) , t < t∗ , (45) V̇2 = ηT2 C2 β2 H2 − ηT2 C2 β2G2C2 β2 η2 , (57)

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.
 15144 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 22, 2025

where Step i (i = 3, · · · , n − 1): With minor variation of the argu-

H2 = α̇1 − F2 − D2 − G2 z3 . (58) ment in Step 2 to analyze V̇3 , · · · , V̇n−1 , it can be concluded
that there exist constants oi j > 0, i = 3, · · · , n−1, j = 1, · · · , m,
Except for C1 β1 η1 , the boundedness of the other variables in such that
(38) for t < t∗ has been established in last step. One sees from |zi j (t)| ≤ pi j − oi j < pi j , t < t∗ , (66)
(8), (22), (23) and (52) that there hold kη1 k < ∞ and kβ1 k < ∞
on [0, t∗ ). From (9), we further have and
kρi k < ∞, t < t∗ , (67)
kα1 k = kC1 β1 η1 k < ∞, t < t∗ . (59)
for i = 3, · · · , n − 1 and j = 1, · · · , m.
This ensures kż1 k in (38) is bounded for t < t∗ . It thus Step n: In the end, we analyze V̇n in (34):
follows from (52) and Lemma 2 that kα̇1 k is bounded on [0, t∗ ).
Further, by (37) for i = 2, we know that kx2 k is bounded over V̇n = −ηTn Cn βn żn . (68)
[0, t∗ ) due to (29) for i = 2 and (59). Recall the boundedness
of kx1 k for t < t∗ established above. There thus hold kF2 k < ∞ By (1) and (17), the differential equation of zn in (10) is
and kG2 k < ∞ on [0, t∗ ) according to Assumption 1. The above obtained as follows
facts together with D2 ∈ L∞ imply that under (29), (58) is żn = Fn + Gn u + Dn − α̇n−1 . (69)
bounded over [0, t∗ ). Denote
Substituting (14) for i = n and (69) into (68) yields
h2 = sup kH2 k. (60)
t∈[0,t∗ ) V̇n = ηTn Cn βn Hn − ηTn βnCnGnCn βn ηn , (70)
Note that β2 in (13) and C2 in (14) are diagonal matrices. By where
(60), (57) is scaled by Hn = α̇n−1 − Fn − Dn − Gn ρn . (71)
V̇2 ≤ h2 kC2 β2 η2 k − ηT2 C2 β2G2C2 β2 η2 , t < t∗ . (61) The essential difference between (58) and (71) lies in ρn .
According to (43) for i = 2, (61) is further bounded by Except for ρn , the boundedness of the other variables in (71)
over [0, t∗ ) can be verified in the same manner as in Step 2.
G2 − GT2
 
Therein, the boundedness of kxn k on [0, t∗ ) together with (67)
V̇2 ≤ h2 kC2 β2 η2 k − η2 C2 β2 G2 +
T
C2 β2 η2
2 for i = n − 1 ensures by (16) that kρn k < ∞, t < t∗ . Hence,
≤ kC2 β2 η2 k (h2 − λ2 kC2 β2 η2 k) , t < t∗ , kHn k is bounded on [0, t∗ ). Then, let

where λ2 is the minimum eigenvalue of G2 . It shows that for hn = sup kHn k. (72)
t∈[0,t∗ )
t < t∗ , V2 < 0 if
h2 Based on (72), (70) is scaled by
kC2 β2 η2 k > . (62)
λ2 V̇n ≤ hn kCn βn ηn k − ηTn βnCnGnCn βn ηn , t < t∗ .
Note from (13) with (12) that
v Following the same line as in Step 2, it can be established that
u 0 !2 1
uX m
2p2 j |zn j (t)| ≤ pn j − on j < pn j , j = 1, · · · , m, t < t∗ , (73)
kC2 β2 η2 k ≥ c2 t η22 j A,
u @
2 2
(63)
j=1
p2j − z 2j where on j is a positive constant, j = 1, · · · , m. Now, a contra-
diction between (28) and (52), (65), (66) and (73) happens.
where c2 = min{c21 , · · · , c2m }. From (48) for i = 2 and (63), This means that (28) is false, and instead,
we have
kC2 β2 η2 k ≥ k2 kη2 k, t < t∗ , (64) |zi j (t)| ≤ pi j (t) − oi j < pi j (t), t ≥ 0, (74)
2c
where k2 = p 2 with p2 = max{p21 , · · · , p2m }. It further shows for i = 1, · · · , n and j = 1, · · · , m. Obviously, our claim in
2 (27) is true. According to Lemma 1, the output constraints are
from (62) and (64) that for t < t∗ , V̇2 < 0 if guaranteed.
h2 Due to (74), it follows from (11), (12), (22) and (23) that
kη2 k > . ηi j and βi j are uniformly bounded, i = 1, · · · , n, j = 1, · · · , m.
λ2 k2
Thereby, the same holds for α1 in (9) and αi in (14), i =
Hence, it is further obtained from (51) for i = 2 that on [0, t∗ ), 2, · · · , n. This in turn implies by (10) and (27) that xi ∈ L∞ ,
c h2
V̇2 < 0 as V2 ≥ 2λ22 k22 . This yields i = 1, · · · , n. Apply it together with r ∈ L∞ in Assumption 4
2 2
to (16). Through a recursive analysis, we obtain ei ∈ L∞ and
c2 h22
 
V2 (t) ≤ ζ2 := max V2 (0), 2 2 , t < t∗ . ρi ∈ L∞ , i = 1, · · · , n. As a result, kuk in (17) is uniformly
2λ2 k2 bounded.
What remains to be shown is the reference tracking perfor-
From (51) for i = 2, we further have kη2 k ≤ ζ2 /c2 , t < t∗ ,
p
mance. Choose
which in turn implies by (11) for i = 2 that
V1 j = 0.5 e21 j , j = 1, · · · , m. (75)
|z2 j (t)| ≤ p2 j − o2 j < p2 j , j = 1, · · · , m, t < t∗ , (65)
where o2 j is a positive constant, j = 1, · · · , m. The bounded- Define F1 = [ f11 , · · · , f1m ] , G1 = {gi j } ∈ R
T
, m×m

ness of kx2 k on [0, t∗ ) and (53) ensure by (16) that e1 = [e11 , · · · , e1m ]T , D1 = [d11 , · · · , d1m ]T and ρ1 =
[ρ11 , · · · , ρ1m ]T . Differentiating (75) with Assumption 4 and
kρ2 k < ∞, t < t∗ , (16) yields
V̇1 j = e1 j ẋ1 j − ṙ j , tk < t < tk+1 ,


which facilitates us to analyze V̇n below.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: CONSTRAINED TRACKING CONTROL OF MIMO NONLINEAR SYSTEMS 15145

for k = 0, 1, 2, · · · and j = 1, · · · , m. It further follows from

(1) that
m
!
X
V̇1 j = e1 j f1 j + d1 j − ṙ j + g ji x2i , tk < t < tk+1 , (76)
i=1

for k = 0, 1, 2, · · · and j = 1, · · · , m. By (16), there is

x2 j = e2 j − b1 j e1 j , j = 1, · · · , m. (77)
Substituting (77) into (76) yields
V̇1 j = e1 j µ1 j − b1 j g j j e1 j , tk < t < tk+1 ,


(78)
for k = 0, 1, 2, · · · and j = 1, · · · , m, where
Xm Xm
µ1 j = f1 j + d1 j − ṙ j + g ji e2i + g ji ρ1i , j = 1, · · · , m. (79)
i=1 i=1
i, j

By feat of x1 ∈ L∞ and the continuity of f1 j and gi j in x1 ,

there hold f1 j ∈ L∞ and gi j ∈ L∞ , i = 1, · · · , m, j = 1, · · · , m.
This in conjunction with ρ1 , e2 , D1 ∈ L∞ and Assumption 4
yields the boundedness of (79). For convenience of notation,
Fig. 1. Experimental setup of 2DSFL robot.
define
( )
for k = 0, 1, 2, · · · and j = 1, · · · , m. This in turn implies by
µ := max sup |µ1 j |, k = 0, 1, 2, · · · ,
t∈(tk ,tk+1 ) (75) that
s
for j = 1, · · · , m. Then, (78) is scaled by µ2
|e1 j | ≤ 2V(tk )e−b1 j gt + 2 2 , tk < t < tk+1 , (83)
b1 j g
V̇1 j ≤ e1 j µ − b1 j g j j e21 j , tk < t < tk+1 , (80)
for k = 0, 1, 2, · · · and j = 1, · · · , m. Due to Assumption 2, for k = 0, 1, 2, · · · and j = 1, · · · , m. It is seen from (83)
there is a constant g so that that the larger b1 j is, the smaller |e1 j | becomes on [tk , tk+1 ),
j = 1, · · · , m, k = 0, 1, 2, · · · . As b1 j in (16) is freely chosen,
0 < g ≤ g j j , j = 1, · · · , m. (81) there certainly exists a suitable b1 j such that e1 j is arbitrarily
small, j = 1, · · · , m. In other words, the convergence speed and
Based on (81), (80) becomes the ultimate bound of |e1 j | on [tk , tk+1 ) can be tuned arbitrarily
V̇1 j ≤ |e1 j |(µ − b1 j g|e1 j |), tk < t < tk+1 , fast and small, respectively, k = 0, 1, 2, · · · , j = 1, · · · , m. 

for k = 0, 1, 2, · · · and j = 1, · · · , m. For any constants q j > V. E XPERIMENTAL V ERIFICATION

0, j = 1, · · · , m, there is
To validate the theoretical findings, an experiment is con-
1 2 qj ducted on a 2-DOF Serial Flexible Link (2DSFL) Robot,
V̇1 j ≤ e1 j + µ2 − b1 j ge21 j
2q 2 manufactured by Quanser Inc. Canada, as shown in Fig. 1(a).
 j 
1 qj
= − b1 j g e21 j + µ2 , tk < t < tk+1 , (82) A. Experimental Setup
2q j 2
The 2DSFL is actuated by two DC motors driving via
for k = 0, 1, 2, · · · and j = 1, · · · , m. Let q j = 1
b1 j g and recall harmonic gearboxes (zero backlash) with a two-bar serial
(75). Then, (82) becomes flexible linkage. The primary link is rigidly clamped to the
b1 j g µ2 first drive (a.k.a. Joint 1) and carries at its end the second
V̇1 j ≤ − e21 j + drive (a.k.a. Joint 2) to which another flexible link is attached.
2 2b1 j g Both motors are equipped with quadrature optical encoders to
µ2 measure the angles of the two joints. The control algorithm is
= −b1 j gV1 j + , tk < t < tk+1 , implemented into a MATLAB-function file in Simulink, which
2b1 j g
is converted to C language code through QUARC software
for k = 0, 1, 2, · · · and j = 1, · · · , m. We further have to generate digital control signals in real time. The digital
! signals are converted to analog signals via Q-8 USB Data
µ2 µ2 Acquisition Device, which are then amplified by a Quanser
V1 j ≤ V(tk ) − 2 2 e−b1 j gt + 2 2
2b1 j g 2b1 j g AMPAQ current amplifier into drive current to drive the two
motors. It is documented that the 2DSFL is described by the
µ2
= V(tk )e−b1 j gt + 1 − e−b1 j gt Euler-Lagrange equation [43], which is in the form of (1). On


2b21 j g2 this basis, to test the robustness of the control system against
µ2 external disturbances, a flexible load is suspended at the end
≤ V(tk )e−b1 j gt + , tk < t < tk+1 , of the second link of the 2DSFL robot as shown in Fig. 1(a).
2b21 j g2 The flexible load means that the load has an impact on the

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.
 15146 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 22, 2025

Fig. 2. Reference tracking of Joint 1 under output constraint. Fig. 3. Reference tracking of Joint 2 under output constraint.

movement of the second link, and the acting force is varying

with the speed change of the second link.

B. Effectiveness Test
The control objective for the 2DSFL robot is to make the
angles of the two joints to track the following references
respectively:
if t < 10,

0.5 sin (0.3πt) + 0.1,
r1 =
−0.3 cos (0.2π(t + 2)) , otherwise,
0.3 cos (0.25πt) − 0.5, if t < 15,

r2 =
0.1 cos (0.1πt) , otherwise,
while satisfying the following constraints:
−0.6 < q1 < 0.7, t ≥ 0, Fig. 4. Intermediate error and prescribed bound.
−0.9 < q2 < 0.3, t ≥ 0.
According to Theorem 1, a controller is obtained with y1 = 0.7,
y1 = −0.6, y2 = 0.3, y2 = −0.9, c11 = 3, c12 = 5, c21 = 4,
c22 = 3, p21 = 150, p22 = 150, b11 = 75, b12 = 120,
b21 = 1.1 and b22 = 1.1. Applying the control scheme to
the 2DSFL robot, the experimental results are displayed in
Figures 2– 7. It is observed from Fig. 2 and Fig. 3 that
reference tracking of the joint angles is accomplished with
both performance and constraint satisfactions, in spite of the
discontinuous references and the unknown system dynamics.
Fig. 4 and Fig. 5 exhibit that the intermediate errors evolve
inside the respectively predefined bounds. The angular veloc-
ities of Joints 1 and 2 are bounded, as shown in Fig. 6
and Fig. 7, respectively. Therefore, the experimental results
demonstrate the effectiveness of the theoretical findings.

Fig. 5. Intermediate error and prescribed bound.

C. Comparative Experiment
On the purpose of comparison, the well-established Linear
Quadratic Regulator (LQR) control method, suggested by the However, this is due to the fact that the parameters of the LQR
manufacturer Quanser, is applied to the 2DSFL robot under controller are well tuned based on the model information. In
the same condition and control objective. As detailed in the contrast, our control scheme is nearly model-free. On the other
manual [36], the LQR control gains are K1 = [15.81, 0.49]T hand, careful inspection of Fig. 10 and Fig. 11 reveals that
and K2 = [15.81, 0.35]T for Joint 1 and Joint 2, respec- under the LQR controller,
tively. Applying the LQR controllers to the 2DSFL robot, • before the discontinuity of the references, the upper and
Figures 8– 11 display the experimental results. It is observed lower bounds of e1 are 0.09 and −0.07, respectively;
from Fig. 8 and Fig. 9 that the joints of the 2DSFL robot the upper and lower bounds of e2 are 0.11 and −0.01,
track the discontinuous references without constraint violation. respectively.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: CONSTRAINED TRACKING CONTROL OF MIMO NONLINEAR SYSTEMS 15147

Fig. 6. Angular velocity of Joint 1. Fig. 9. Constrained reference tracking of Joint 2 by the LQR controller.

Fig. 7. Angular velocity of Joint 2.

Fig. 10. Tracking errors e1 of Joint 1 by the LQR controller and e1 of Joint
1 by our controller.

Fig. 8. Constrained reference tracking of Joint 1 by the LQR controller.

Fig. 11. Tracking errors e2 of Joint 2 by the LQR controller and e2 of Joint
2 by our controller.
• at the discontinuity of the references, the peak of e1 is
0.28, and the settling time is 0.46; the peak of e2 is 0.26, In contrast, under the designed controller,
and the settling time is 0.63.
• after the discontinuity of the references, the upper and • before the discontinuity of the references, the upper and
lower bounds of e1 are 0.06 and −0.04, respectively; lower bounds of e1 are 0.01 and −0.01, respectively;
the upper and lower bounds of e2 are 0.01 and −0.01, the upper and lower bounds of e2 are 0.01 and −0.01,
respectively. respectively.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.
 15148 IEEE TRANSACTIONS ON AUTOMATION SCIENCE AND ENGINEERING, VOL. 22, 2025

• at the discontinuity of the references, the peak of e1 is [8] Z. Ruan, Q. Yang, S. S. Ge, and Y. Sun, “Adaptive fuzzy fault tolerant
0.2, and the settling time is 0.17; the peak of e2 is 0.29, control of uncertain MIMO nonlinear systems with output constraints
and the settling time is 0.24. and unknown control directions,” IEEE Trans. Fuzzy Syst., vol. 30,
no. 5, pp. 1224–1238, May 2022.
• after the discontinuity of the references, the upper and [9] J.-X. Zhang and G.-H. Yang, “Fault-tolerant output-constrained control
lower bounds of e1 are 0.01 and −0.01, respectively; of unknown Euler–Lagrange systems with prescribed tracking accuracy,”
the upper and lower bounds of e2 are 0.01 and −0.01, Automatica, vol. 111, Jan. 2020, Art. no. 108606.
respectively. [10] X. Jin, “Adaptive fixed-time control for MIMO nonlinear systems with
asymmetric output constraints using universal barrier functions,” IEEE
The quantitative comparison of these performance indices Trans. Autom. Control, vol. 64, no. 7, pp. 3046–3053, Jul. 2019.
shows that our controller yields the better tracking perfor- [11] L. Kong, W. He, Z. Liu, X. Yu, and C. Silvestre, “Adaptive tracking
mance than the LQR controller in regard to the smaller control with global performance for output-constrained MIMO nonlinear
steady-state error, the smaller error peak, and the shorter set- systems,” IEEE Trans. Autom. Control, vol. 68, no. 6, pp. 3760–3767,
Jun. 2023.
tling time. Therefore, the results of the comparative experiment [12] Y. Song and S. Zhou, “Neuroadaptive control with given performance
verify the advantages of our controller. specifications for MIMO strict-feedback systems under nonsmooth actu-
ation and output constraints,” IEEE Trans. Neural Netw. Learn. Syst.,
vol. 29, no. 9, pp. 4414–4425, Sep. 2018.
VI. C ONCLUSION [13] L.-B. Wu, J. H. Park, X.-P. Xie, and Y.-J. Liu, “Neural network adaptive
tracking control of uncertain MIMO nonlinear systems with output
This paper presents a hybrid tracking control approach for constraints and event-triggered inputs,” IEEE Trans. Neural Netw. Learn.
the MIMO block-triangular nonlinear systems with output con- Syst., vol. 32, no. 2, pp. 695–707, Feb. 2021.
straints, discontinuous references and unmatched disturbances. [14] K. Zhao and J. W. Chen, “Adaptive neural quantized control of
It is effective for the cases of the general nonlinearities of MIMO nonlinear systems under actuation faults and time-varying output
constraints,” IEEE Trans. Neural Netw. Learn. Syst., vol. 31, no. 9,
the system, the irregular discontinuity of the references, and pp. 3471–3481, Sep. 2020.
the totally unknown system dynamics. The control scheme [15] M. Boudali, R. Orjuela, and M. Basset, “Unified dynamic and geomet-
consists only of a robust decoupling constrained controller rical vehicle guidance strategy to cope with the discontinuous reference
and a proportional controller, without the needs for parameter trajectory,” Veh. Syst Dyn., vol. 58, no. 11, pp. 1629–1656, 2020.
identification, function approximation, disturbance estimation, [16] M. Reichhartinger and M. Horn, “Application of higher order sliding-
mode concepts to a throttle actuator for gasoline engines,” IEEE Trans.
derivative calculation and command filtering. It steers the Ind. Electron., vol. 56, no. 9, pp. 3322–3329, Sep. 2009.
system outputs to track the discontinuous references with [17] H. Gao, Z. Li, X. Yu, and J. Qiu, “Hierarchical multiobjective heuristic
the tunable speed and accuracy, without constraint violation. for PCB assembly optimization in a beam-head surface mounter,” IEEE
A comparative experiment on a 2DSFL robot is performed, Trans. Cybern., vol. 52, no. 7, pp. 6911–6924, Jul. 2022.
and the result verifies the effectiveness and advantage of the [18] M. Posa and R. Tedrake, “Direct trajectory optimization of rigid
body dynamical systems through contact,” in Proc. Algorithmic Found.
proposed approach. The current approach allows for tuning Robot., X. E. Frazzoli, T. Lozano-Perez, N. Roy, and D. Rus, Eds., Jan.
the transient and steady-state tracking performance via the 2013, pp. 527–542.
controller parameters, however in a qualitative manner. As a [19] F. Fotiadis and G. A. Rovithakis, “Prescribed performance control for
result, the anticipated tracking behavior fails to be prescribed. discontinuous output reference tracking,” IEEE Trans. Autom. Control,
Moreover, the qualitative tuning is conservative leading to vol. 66, no. 9, pp. 4409–4416, Sep. 2021.
[20] J.-X. Zhang and T. Chai, “Robust tracking control of unknown nonlinear
a large control amplitude, which may violate the actuator systems with discontinuous references under output constraints,” IEEE
saturation of a practical plant. Therefore, working toward the Trans. Syst., Man, Cybern., Syst., vol. 54, no. 11, pp. 7015–7026, Nov.
extension to the prescribed performance control design under 2024.
input and output constraints deserves future investigations. [21] C. G. L. Bianco, “An efficient algorithm for the real-time generation
of synchronous reference signals,” IEEE Trans. Ind. Electron., vol. 64,
no. 6, pp. 4621–4630, Jun. 2017.
R EFERENCES [22] S. Liu, J. Wang, and W. Wei, “Iterative learning control based on a
noninstantaneous impulsive fractional-order system,” J. Vib. Control,
[1] K. P. Tee, S. S. Ge, and E. H. Tay, “Barrier Lyapunov functions for the vol. 22, no. 8, pp. 1972–1979, 2016.
control of output-constrained nonlinear systems,” Automatica, vol. 45, [23] J.-X. Zhang and T. Chai, “Global prescribed performance control of
no. 4, pp. 918–927, Apr. 2009. unknown strict-feedback systems with quantized references,” IEEE
[2] K. P. Tee, B. Ren, and S. S. Ge, “Control of nonlinear systems with time- Trans. Syst., Man, Cybern., Syst., vol. 53, no. 10, pp. 6257–6267, Jul.
varying output constraints,” Automatica, vol. 47, no. 11, pp. 2511–2516, 2023.
Nov. 2011. [24] D. Zheng, X. Song, S. Song, and V. Stojanovic, “Quantized control
[3] K. Zhao, Y. D. Song, T. D. Ma, and L. He, “Prescribed perfor- for interconnected PDE systems via mobile measurement and control
mance control of uncertain Euler-Lagrange systems subject to full-state strategies,” J. Franklin Inst., vol. 361, no. 14, Sep. 2024, Art. no.
constraints,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 8, 107070.
pp. 3478–3489, Aug. 2018. [25] J. Wang, H. Pan, and W. Sun, “Event-triggered adaptive output constraint
[4] S. Zhang, Y. Dong, Y. Ouyang, Z. Yin, and K. Peng, “Adaptive tracking control of uncertain MIMO nonlinear systems with sensor
neural control for robotic manipulators with output constraints and and actuator faults,” IEEE Trans. Autom. Sci. Eng., vol. 21, no. 4,
uncertainties,” IEEE Trans. Neural Netw. Learn. Syst., vol. 29, no. 11, pp. 6774–6786, Oct. 2024.
pp. 5554–5564, Nov. 2018. [26] S. Liu, B. Niu, G. Zong, X. Zhao, and N. Xu, “Adaptive neural dynamic-
[5] Y. Liu, X. Chen, Y. Wu, H. Cai, and H. Yokoi, “Adaptive neural memory event-triggered control of high-order random nonlinear systems
network control of a flexible spacecraft subject to input nonlinearity with deferred output constraints,” IEEE Trans. Autom. Sci. Eng., vol. 21,
and asymmetric output constraint,” IEEE Trans. Neural Netw. Learn. no. 3, pp. 2779–2791, Jul. 2024.
Syst., vol. 33, no. 11, pp. 6226–6234, Nov. 2022. [27] L. Liu, M. Sun, R. Wang, C. Zhu, and Q. Zeng, “Finite-time neural
[6] Y. J. Liu, S. M. Lu, and S. C. Tong, “Neural network controller design control of stochastic active electromagnetic suspension system with
for an uncertain robot with time-varying output constraint,” IEEE Trans. actuator failure,” IEEE Trans. Intell. Vehicles, early access, Apr. 16,
Syst., Man, Cybern., Syst., vol. 47, no. 8, pp. 2060–2068, Aug. 2017. 2024, doi: 10.1109/TIV.2024.3386693.
[7] X. Yu, W. He, H. Li, and J. Sun, “Adaptive fuzzy full-state and output- [28] L. Liu, Z. Li, Y.-J. Liu, and S. Tong, “Adaptive fuzzy output feedback
feedback control for uncertain robots with output constraint,” IEEE control of switched uncertain nonlinear systems with constraint con-
Trans. Syst., Man, Cybern., Syst., vol. 51, no. 11, pp. 6994–7007, Nov. ditions related to historical states,” IEEE Trans. Fuzzy Syst., vol. 30,
2021. no. 12, pp. 5091–5103, Dec. 2022.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.
 ZHANG et al.: CONSTRAINED TRACKING CONTROL OF MIMO NONLINEAR SYSTEMS 15149

[29] Z. Zhang, K. Zhang, X. Xie, and V. Stojanovic, “ADP-based prescribed- Jin-Xi Zhang (Member, IEEE) received the B.S.
time control for nonlinear time-varying delay systems with uncertain degree in automation and the Ph.D. degree in control
parameters,” IEEE Trans. Autom. Sci. Eng. (from July 2004), vol. 22, theory and control engineering from Northeastern
pp. 3086–3096, 2025. University, Shenyang, China, in 2014 and 2020,
[30] J.-X. Zhang, E.-Y. Cui, and P. Shi, “Low-complexity high-performance respectively.
control of unknown block-triangular MIMO nonlinear systems,” IEEE From 2019 to 2020, he was a Research Fellow
Trans. Ind. Electron., early access, Dec. 27, 2024, doi: 10.1109/ with the Institute for Intelligent Systems, Faculty of
TIE.2024.3515269. Engineering and the Built Environment, University
[31] X. Song, Z. Peng, S. Song, and V. Stojanovic, “Interval observer design of Johannesburg, South Africa. He is currently a Dis-
for unobservable switched nonlinear partial differential equation systems tinguished Associate Professor with the State Key
and its application,” Int. J. Robust Nonlinear Control, vol. 34, no. 16, Laboratory of Synthetical Automation for Process
pp. 10990–11009, Nov. 2024. Industries, Northeastern University. His research interests include nonlinear
[32] Y. Song, X. Huang, and C. Wen, “Robust adaptive fault-tolerant PID control, intelligent control, prescribed performance control, multi-agent sys-
control of MIMO nonlinear systems with unknown control direction,” tems, fault diagnosis and fault-tolerant control, and image processing. He is
IEEE Trans. Ind. Electron., vol. 64, no. 6, pp. 4876–4884, Jun. 2017. an Associate Editor of International Journal of Fuzzy Systems and Journal of
[33] S. Zhou and Y. Song, “Prescribed performance neuroadaptive fault- Automation and Intelligence.
tolerant compensation for MIMO nonlinear systems under extreme
actuator failures,” IEEE Trans. Syst., Man, Cybern., Syst., vol. 51, no. 9,
pp. 5427–5436, Sep. 2021. Wei-Li Qi received the B.S. degree in mechanical
[34] X. Wang, Q. Wang, and C. Sun, “Prescribed performance fault-tolerant design manufacturing and automation from China
control for uncertain nonlinear MIMO system using actor-critic learn- University of Petroleum (East China), Qingdao,
ing structure,” IEEE Trans. Neural Netw. Learn. Syst., vol. 33, no. 9, China, in 2021. He is currently pursuing the M.S.
pp. 4479–4490, Sep. 2022. degree in artificial intelligence with Northeastern
[35] Z. Shao, Y. Wang, D. Luo, and Y. Song, “A unified approach for tracking University, Shenyang, China. His research interests
control of MIMO nonlinear systems under unknown control directions include nonlinear systems, fault-tolerant control, and
and irregular constraints,” IEEE Trans. Autom. Sci. Eng., vol. 21, no. 4, prescribed performance control.
pp. 6845–6854, Oct. 2024.
[36] Reference Manual 2-DOF Serial Flexible Link Robot, Quanser Inc.,
Markham, ON, Canada, 2012.
[37] R. Skjetne, T. I. Fossen, and P. V. Kokotovic, “Adaptive maneuvering,
with experiments, for a model ship in a marine control laboratory,” Tianyou Chai (Life Fellow, IEEE) received the
Automatica, vol. 41, no. 2, pp. 289–298, Feb. 2005. Ph.D. degree in control theory and engineering from
[38] J.-X. Zhang and G.-H. Yang, “Low-complexity adaptive tracking con- Northeastern University, Shenyang, China, in 1985.
trol of MIMO nonlinear systems with unknown control directions,” He became a Professor with Northeastern Univer-
Int. J. Robust Nonlinear Control, vol. 29, no. 7, pp. 2203–2222, sity in 1988. He was the Director of the Department
May 2019. of Information Science, National Natural Science
[39] J.-X. Zhang, K.-D. Xu, and Q.-G. Wang, “Prescribed performance track- Foundation of China, from 2010 to 2018. He is
ing control of time-delay nonlinear systems with output constraints,” currently the Founder and the Director of the Center
IEEE/CAA J. Autom. Sinica, vol. 11, no. 7, pp. 1557–1565, Jul. 2024. of Automation, which became the National Engi-
[40] A. K. Kostarigka, Z. Doulgeri, and G. A. Rovithakis, “Prescribed neering and Technology Research Center and the
performance tracking for flexible joint robots with unknown dynamics State Key Laboratory of Synthetical Automation for
and variable elasticity,” Automatica, vol. 49, no. 5, pp. 1137–1147, May Process Industries. He has developed control technologies with applications to
2013. various industrial processes. He has published 297 peer-reviewed international
[41] J.-X. Zhang, J. Ding, and T. Chai, “Fault-tolerant prescribed performance journal articles. His current research interests include modeling, control,
control of wheeled mobile robots: A mixed-gain adaption approach,” optimization, and integrated automation of complex industrial processes. He
IEEE Trans. Autom. Control, vol. 69, no. 8, pp. 5500–5507, Aug. 2024. is a member of the Chinese Academy of Engineering and an IFAC Fellow.
[42] E. Psomopoulou, A. Theodorakopoulos, Z. Doulgeri, and For his contributions, he has won five prestigious awards of the National
G. A. Rovithakis, “Prescribed performance tracking of a variable Natural Science, the National Science and Technology Progress, and the
stiffness actuated robot,” IEEE Trans. Control Syst. Technol., vol. 23, National Technological Innovation, the 2007 Industry Award for Excellence
no. 5, pp. 1914–1926, Sep. 2015. in Transitional Control Research from IEEE Multiple-Conference on Systems
[43] J.-X. Zhang, J.-G. Song, Q. Chen, and F. Yang, “Mixed-gain adaption- and Control, and the 2017 Wook Hyun Kwon Education Award from the Asian
based fault-tolerant funnel control of robotic manipulators with unknown Control Association. His paper titled Hybrid intelligent control for optimal
dynamics and sensor faults,” IEEE Trans. Autom. Sci. Eng., vol. 22, operation of shaft furnace roasting process was selected as one of three best
pp. 3044–3055, 2025. papers for the Control Engineering Practice Paper Prize during 2011–2013.

Authorized licensed use limited to: INDIAN INSTITUTE OF TECHNOLOGY KANPUR. Downloaded on June 26,2025 at 09:57:14 UTC from IEEE Xplore. Restrictions apply.