This article has been accepted for inclusion in a future issue of this journal.

Content is final as presented, with the exception of pagination.

IEEE TRANSACTIONS ON CYBERNETICS 1

Adaptive Neural Network Control of an Uncertain

Robot With Full-State Constraints
Wei He, Member, IEEE, Yuhao Chen, and Zhao Yin, Student Member, IEEE

Abstract—This paper studies the tracking control problem for tracking control of uncertain nonlinear multiinput and mul-
an uncertain n-link robot with full-state constraints. The rigid tioutput (MIMO) systems in the discrete-time form is studied
robotic manipulator is described as a multiinput and multioutput in [22]. Yang et al. [23] and Cui et al. [24] investigated the
system. Adaptive neural network (NN) control for the robotic sys-
tem with full-state constraints is designed. In the control design, optimized adaptive control and trajectory generation for a class
the adaptive NNs are adopted to handle system uncertainties and of wheeled inverted pendulum models of vehicle systems.
disturbances. The Moore–Penrose inverse term is employed in However, most of the papers considered the nonlinear system
order to prevent the violation of the full-state constraints. A bar- without constraints, and used the NNs to approximate those
rier Lyapunov function is used to guarantee the uniform ultimate uncertainties for achieving the control objective. The system
boundedness of the closed-loop system. The control performance
of the closed-loop system is guaranteed by appropriately choos- performance might be influenced by these methods when there
ing the design parameters. Simulation studies are performed to are constraints.
illustrate the effectiveness of the proposed control. In order to overcome the drawbacks caused by constraints
Index Terms—Adaptive control, barrier Lyapunov func- and guarantee the performance of closed-loop control, we must
tion (BLF), full-state constraint, neural networks (NNs), non- deal with the constraints of the robotic system efficiently. For
linear system, robot. the sake of this, we use the adaptive neural network control
design to compensate for the effects of both parametric and
disturbance uncertainties and achieve uniform ultimate bound-
I. I NTRODUCTION
edness. A barrier Lyapunov function (BLF) can analyze the
OR THE past few years, there has been increas-
F ing attention paid to the control problem of the
constrained robot [1]–[4]. Universally, there are constraints
stability of the closed-loop system, and also close the door
upon the violation of the full-state constraints.
Because of requiring relatively less information of the
in most physical systems with various forms, such as system dynamics, neural networks turn out to be a clever
physical stoppages, saturation, performance, and safety move. Artificial neural networks have been proven to be
specifications [5], [6]. Obviously, severe performance degra- capable of approximating various nonlinear functions to any
dation, safety problems, and other issues can be caused by desired degree of accuracy under certain conditions [25], [26].
violation of these constraints. Therefore, to narrow these Thus, in the control design of uncertain nonlinear sys-
gaps in the control design for the robot becomes a serious tems, artificial neural networks have already been widely
problem. The representative works on mobile manipulators employed. For example, in [6], an adaptive neural network
are given in [7]–[10]. Recently, using adaptive neural net- tracking controller is designed for the uncertain MIMO non-
work (NN) control design techniques, several interesting solu- linear systems, in [27], an adaptive neural network friction
tions have been presented for the nonlinear systems [11]–[17]. compensator is presented for servo control of hard disk
Ren et al. [18] and Liu et al. [19] dealt with the problem of drives, and Dai et al. [28] presented a novel contribution to
adaptive NN control for a class of uncertain nonlinear sys- solve the learning control problem for ocean surface ship
tems. In [20] and [21], an adaptive tracking control design in uncertain dynamical environments. The relevant applica-
is studied for a class of nonlinear discrete-time systems tions for this approach based on Lyapunov’s stability theory
with dead-zone input. An adaptive neural output feedback include [29] and [30]. He et al. [31] have presented the neu-
ral network control for a rehabilitation robot by both state
Manuscript received February 2, 2015; accepted February 25, 2015. This
work was supported in part by the National Natural Science Foundation of feedback and output feedback.
China under Grant 61203057 and Grant 61403063, in part by the National Given unknown system dynamics and constraints, as well as
Basic Research Program of China (973 Program) under Grant 2014CB744206, interactions with the human and environment, the control of an
in part by the National High Technology Research and Development Program
of China (863 Program) under Grant 2015AA042304, and in part by the uncertain robot still needs more assiduous efforts in the con-
Fundamental Research Funds for the China Central Universities of University trol design and stability analysis. Li et al. [1] have presented
of Electronic Science and Technology of China under Grant ZYGX2013Z003. the results for the impedance control of the constrained robots
This paper was recommended by Associate Editor J. Q. Gan.
The authors are with the School of Automation Engineering and Center based on Lyapunov’s direct method. The hybrid force/position
for Robotics, University of Electronic Science and Technology of China, control is also widely employed for the robotic systems with
Chengdu 611731, China (e-mail: [email protected]). constraints [32]–[34]. To ensure the position tracking and the
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. boundedness of force errors, adaptive position/force control
Digital Object Identifier 10.1109/TCYB.2015.2411285 is proposed for an uncertain constrained flexible joint robot
2168-2267 c 2015 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.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

2 IEEE TRANSACTIONS ON CYBERNETICS

in [35]. In [36], robust adaptive position/force control is inves- containing the origin, that is continuous, positive definite, has
tigated for a nonholonomic robot under holonomic constraints. continuous first-order partial derivatives at every point of D,
Thanks to the research progress of control design for has the property V(x) → ∞ as x approaches the boundary
the n-link robot with uncertainties, this paper is provided of D, and satisfies V(x(t)) ≤ b, ∀t ≥ 0 along the solution of
valuable motivation. However, most of exciting papers con- ẋ = f (x) for x(0) ∈ D and some positive constant b.
sidered the nonlinear system without constraints [37], [38], Lemma 1 [46], [47]: For bounded initial conditions, if
nonlinear system with output constraints [39], [40], flexible there exists a C1 continuous and positive definite Lyapunov
systems with output constraints [41]–[43], or the dynam- function V(x) satisfying κ1 ( x ) ≤ V(x) ≤ κ2 ( x ), such that
ics of the nonlinear system is represented by a single-input V̇(x) ≤ −ρV(x) + c, where κ1 , κ2 : Rn → R are class K
and single-output (SISO) system [44]. Tee and Ge [44] functions, and ρ and c are two positive constants, then the
addressed the tracking problem of strict-feedback systems solution x(t) is uniformly bounded.
with constraints on the states via an elegant control scheme. Lemma 2 [48]: For adaptive law (17), there exits a com-
Ren et al. [45] presented a novel adaptive neural control pact set
method by employing the state estimation observer for a    
  si
class of output feedback nonlinear systems in the presence ω1 = Ŵi |Ŵi  ≤ (1)
of unknown functions. There are several interesting insights σi
and many novel ideas in the above two papers, which give where Si (Z) ≤ si with φi > 0, such that Ŵi (t) ∈ ω1 ,
us inspirations to choose the BLF to solve our problem in ∀t ≥ 0 provided that Ŵi (0) ∈ ω1 .
this paper. However, the authors only consider the constraints Lemma 3 [49]: For any positive constant vector kb ∈ Rn ,
with output or partial state, and especially the authors solve the following inequality holds for any vector x ∈ Rn in the
the problem with a SISO system. Therefore, in this paper, interval |x| < |kb |:
we propose the Moore–Penrose inverse term and design the
adaptive neural network control to address the tracking prob- kbT kb xT x
log ≤ .
lem of a MIMO robotic system. The adaptive neural networks kbT kb − xT x kbT kb− xT x
are used to approximate the unknown model of the robot in
order to dispose the system uncertainties and improve the B. Problem Formulation
robustness of the system, and the BLF is designed for pre-
venting the violation of the constraints. Compared with the The dynamics of an n-link rigid robotic system are described
existing work, this paper has several main contributions as by [50]
follows. M(q)q̈ + C(q, q̇)q̇ + G(q) = τ (t) − J T (q)f (t) (2)
1) Adaptive neural network control with full-state feed-
back is proposed for an uncertain robot with constraints, where q, q̇, q̈ ∈ Rn are the position, velocity, and accel-
which can guarantee the performance and improve the eration vectors, respectively, τ ∈ Rn is the input torque,
robustness of closed-loop system effectively. M(q) ∈ Rn×n represents a symmetric positive definite iner-
2) The constraints are not violated with the proposed con- tia matrix, C(q, q̇)q̇ ∈ Rn denotes the centripetal and Coriolis
trol, and the Moore–Penrose inverse term is used to torques, and G(q) ∈ Rn is the gravitational force, J(q)
design the control law to prevent the violation of the describes the Jacobian matrix which is assumed to be nonsin-
full-state constraints. gular, and f (t) ∈ Rn is the vector of constrained force exerted
3) With the proposed control, all the signals of the closed- by the human and environment, which is 0 when there is no
loop system are proved to be uniformly ultimately contact for the robot with the human and environment.
bounded, the closed-loop system states eventually con- Property 1 [1], [50]: The matrix M(q) is symmetric and
verge to a compact set, and the asymptotic tracking is positive definite.
achieved successfully. Assumption 1: We assume that the constrained force f (t) is
This paper is organized as follows. The preliminaries and uniformly bounded, i.e., there exists a constant f̄ ∈ R+ , such
the dynamics of an n-link uncertain robotic system are given in that | f (t)| ≤ f̄ , ∀t ∈ [0, ∞).
Section II. Control design and stability analysis are discussed Remark 1: This is a reasonable assumption as the time-
by employing a BLF in Section III, where it is shown that varying constrained force f (t) is bounded from an engineering
the uniform boundedness of the closed-loop system can be point of view. Thus, the knowledge of the exact value for
achieved by the proposed control and the output constraints f (t) is not required, and the force sensors mounted on the
are never violated. Simulations are carried out to illustrate interaction points are not needed as well. As such, different
the effectiveness of the proposed control in Section IV. The constrained forces up to various levels of interactions can be
conclusion is given in Section V. applied without affecting the control design or analysis.
Since the positions and velocities of robot are known, we
can obtain the accelerations logically. Then, we let x1 = q,
II. P RELIMINARIES AND P ROBLEM F ORMULATION
x2 = q̇, we have the description of the robot as
A. Useful Technical Lemmas and Definitions
Definition 1 [40]: A BLF is a scalar function V(x), defined ẋ1 = x2
 
with respect to the system ẋ = f (x) on an open region D ẋ2 = M −1 (x1 ) τ − J T (x1 )f − C(x1 , x2 )x2 − G(x1 ) = a. (3)
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HE et al.: ADAPTIVE NN CONTROL OF AN UNCERTAIN ROBOT WITH FULL-STATE CONSTRAINTS 3

The control objective is to track a desired trajectory of According to the Moore–Penrose inverse, we can obtain
the earth-frame positions xd (t) = [qd1 , qd2 , . . . , qdn (t)]T , and 
 + 0, z2 = [0, 0, . . . , 0]T
desired trajectory of the velocities α = [α1 , α2 , . . . , αn ]T . zT2 zT2 = (14)
While ensuring that all signals are bounded and that the full- 1, Otherwise
state constraints are not violated, i.e., |x1 | ≤ kc1 , |x2 | ≤ 
when z2 = [0, 0, . . . , 0]T , V̇2 = − 3i=1 k1i z21i ≤ 0.
kc2 , ∀t ≥ 0, where kc1 = [kc11 , kc12 , . . . , kc1n ]T kc2 = Then asymptotic stability of the system can still be
[kc21 , kc22 , . . . , kc2n ]T are positive constant vectors. drawn by the Barbalat lemma in [51]. Otherwise in case
of z2
= [0, 0, . . . , 0]T , we design the model-based control
III. C ONTROL D ESIGN (MB control) as
A. Model-Based Control
τ = J T (x1 )f + C(x1 , x2 )x2 + G(x1 ) + M(x1 )α̇ − K2 z2
In case that the parameters M(x1 ), J T (x1 ), C(x1 , x2 ), and
G(x1 ) are known, we denote z1 = [z11 , z12 , . . . , z1n ]T =  T + n
z1i z2i  T + n
z2i (ai − α̇i )
− z2 2 − z2
− z2 2 − z2
(15)
x1 − xd and z2 = [z21 , z22 , . . . , z2n ]T = x2 − α. Choosing k
i=1 ai 1i i=1
kbi 2i
the asymmetric BLF as
where K2 is the control gain, and k2i , i = 1, 2, . . . , n are
1
n
k2 positive constants. Then, substituting (13) and (15) into (12),
V1 = log 2 ai 2 (4)
2 kai − z1i
i=1
we can obtain

n
where ka = kc1 − X0 = [ka1 , ka2 , . . . , kan ]T , then differentiat- V̇2 = − k1i z21i − zT2 K2 z2 < 0. (16)
ing of V1 with respect to time, we have i=1
n
z1i ż1i According to Lemma 1, we know the signal z1 remains in
V̇1 = . (5)
k − z21i
2 the interval −ka ≤ z1 ≤ ka , ∀t > 0, similarly, the signal z2
i=1 ai
remains in the interval −kb ≤ z2 ≤ kb , ∀t > 0.
Differentiating of z1 with respect to time, we have
ż1 = ẋ1 − ẋd B. Adaptive Neural Network Control With
= z2 + α − ẋd (6) Full-State Feedback
ż1i = z2i + αi − ẋdi . (7) The adaptive law is given as follows:
 
˙ =
S (Z )z − σ |z |Ŵ , i = 1, 2, . . . , n
We propose α as Ŵi i i i 2,i i 2i i (17)
α = ẋd − A (8)
where Ŵ = [Ŵ1 , Ŵ2 , . . . , Ŵn ]T are the weights of the neu-
where ral network, S(Z) = [S(Z)1 , S(Z)2 , . . . , S(Z)n ]T are the basis
⎡
2 − z2 k z
 ⎤ functions, Z = [x1T , x2T , α T , α̇ T ] are the inputs of the neural
ka1 11 11 11
⎢ 2  ⎥ networks,
i =
iT > 0 (i = 1, 2, . . . , n) are the constant gain
⎢ k − z212 k12 z12 ⎥ matrices, and σi > 0 (i = 1, 2, . . . , n) are small constants. The
A = ⎢ a2 ⎥ (9)
⎣ · · · ⎦ neural network Ŵ T S(Z) is used to approximate W ∗T S(Z)
 2
ka3 − z21n k1n z1n
W ∗T S(Z) = Ŵ T S(Z) − (Z)
k1i , i = 1, 2, . . . , n are positive constants. Substituting (7)–(9) 
into (5), we can obtain = − J T (x1 )f + C(x1 , x2 )x2 + G(x1 )


n n + M(x1 )α̇ − (Z) (18)
z1i z2i
V̇1 = − k1i z21i + . (10)
k − z21i
2 where W̃i = Ŵi − Wi∗ and W̃i , Ŵi , Wi∗ are the neural network
i=1 i=1 ai
weight errors, estimate, and actual value, respectively. (Z) is
Then, we consider a BLF candidate as the approximation error which is bounded over the compact
1 set, i.e., |(Z)| ≤ ¯ (Z), ∀Z ∈ z with ¯ (Z) > 0 as an unknown
n
k2 1
V2 = V1 + log 2 bi 2 + zT2 Mz2 (11) constant.
2
i=1
kbi − z2i 2
Then, we propose the following control as:
where kb = kc2 − Y0 = [kb1 , kb2 , . . . , kbn ]T , then differentiat- 
 T + n
z2i (ai − α̇i ) k1i z21i k2i z22i
ing (11) with respect to time leads to τ = − z2 + +
  2 − z2
kbi 2 − z2
kai 2 − z2
kbi
n n
z1i z2i z2i ż2i
i=1 2i

1i 2i
V̇2 = − k1i z1i +
2
2 − z2
+ 2 + zT2 Mż2 . z1i z2i
i=1 i=1
kai 1i kbi − z 2
2i + 2 − Ŵ T S(Z) − K3 z2 (19)
kai − z21i
(12)
where K3 is the control gain. In the following part, we are
Differentiating z2 with respect to time, we have
  ready to present the stability theorem of the closed-loop
ż2 = M −1 (x1 ) τ − J T (x1 )f − C(x1 , x2 )x2 − G(x1 ) − α̇. (13) system.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

4 IEEE TRANSACTIONS ON CYBERNETICS

Consider the following Lyapunov candidate function: where

⎛ ⎛  2 ⎞
1  T −1 σi2 Wi∗ 
n
V3 = V2 + W̃i
i W̃i . (20) ρ = min⎝min(2k1i ), min(2k2i ), min⎝  ⎠
2 −1
i=1 λmax
i
⎞
Differentiating V3 with respect to time yields
2λmin (K3 − I) ⎠

n n
z1i z2i n
z2i ż2i λmax (M)
V̇3 = − k1i z21i + +
k − z1i
2 2 k − z22i
2    
i=1 i=1 ai i=1 bi 1 n
σi2  4 s  ∗ 4

n C = (Z)
¯ 2
+ W  +
∗ i
+ Wi  (27)
+ zT2 Mż2 + ˙ .
W̃iT
i−1 W̃ (21) 2 4 i
σi
i i=1
i=1
where λmin (•) and λmax (•) denote the minimum and maximum
Taking ż2 = ẋ2 − α̇ = a − α̇ into (21), we can get eigenvalues of matrix •, where λ(A) are real, respectively. To
ensure ρ > 0, the control gain K3 is chosen to satisfy the

n n
z1i z2i n
z2i (ai − α̇i )
V̇3 = − k1i z21i + + following conditions:
i=1
k − z1i
2
i=1 ai
2
i=1
2 − z2
kbi 2i
  λmin (K3 − I) > 0. (28)
+ zT2 τ − J T (x1 )f − C(x1 , x2 )x2 − G(x1 ) − M α̇
 n Remark 2: If C can equal to zero, and we can say that the
+ ˙ .
W̃iT
i−1 W̃ (22) system could achieve the exponential stability. However, for
i 
i=1 our controller, the C = 1/2 ¯ (Z) 2 + ni=1 σi2 /4( Wi∗ 4 +
W̃i 4 ), where σi are the control parameters designed in the
Substituting (17)–(19) into (22), we can obtain
adaptive law, and σi are used to improve the robustness of

n 
n 
n
the system. We can also design another adaptive laws without
V̇3 = − k1i z21i + W̃iT Si (Z)z2,i − W̃iT |z2i |Ŵi the parameter of σi , but it will influence the robustness of the
i=1
 +
i=1
 +
i=1 proposal control. In other words, σi in C can be set as zero.

n
zT2 zT2 k1i z21i 
n
zT2 zT2 k2i z22i And for the term 1/2 ¯ (Z) 2 , it derives from the approxi-
− −
2 − z2
kai 2 − z2
kbi mation error of neural network, which is a positive constant.
i=1 1i i=1 2i
 T + Therefore, we just achieve the stability for our system, and
n
z2i (ai − α̇i ) − z2 z2 z2i (ai − α̇i )
T
could not achieve the exponential stability.
+ 2 − z2
kbi 2i
According to Lemma 3, we have |z1 | < |kb |, ∀t > 0,
i=1
  provided the initial conditions satisfy z1 (0) < kb . Then
+ zT2 W ∗T S(Z) − Ŵ T S(Z) + (Z) − zT2 K3 z2 (23) it is straightforward to show, from x1 (t) = z1 + xd , and
 |xd (t)| ≤ A0 , that |x1 (t)| < kb + A0 = kc . Hence, we can state
when z2 = [0, 0, . . . , 0]T , V̇2 = − 3i=1 k1i z21i . According to that the output constraint will never be violated.
the Barbalat lemma in [51], we can still draw the asymp- Theorem 1: For the system dynamics described by (2),
totic stability of the system. Otherwise, in case of z2
= under Assumption 1, with state feedback control law (19)
[0, 0, . . . , 0]T , we have together with adaption law (17), for initial conditions sat-
isfy z1 (0) ∈ 0 := {z1 ∈ R3 : −ka < z1 < ka }, and
 1 T 
n
V̇3 ≤ − k1i z21i − zT2 K3 z2 + z2 z2 + ¯ (Z) 2 z2 (0) ∈ 0 := {z2 ∈ R3 : −kb < z2 < kb }, i.e., the initial con-
2 ditions are bounded. The signals of the closed-loop system are
i=1

n 
n
k1i z21i 
n
k2i z22i semiglobally uniformly bounded (SGUB). And the asymptotic
− W̃iT |z2i |Ŵi − 2 − z2
− 2 − z2 tracing is achieved, i.e., x1 (t) → xd (t) and x2 (t) → α(t) as
kai kbi
i=1 i=1 1i i=1 2i t → ∞. The multiple full-state constraints are never violated,

n
k1i z21i 
n
k2i z22i i.e., |x1 | < kc1 , |x2 | < kc2 , ∀t > 0, and the closed-loop error
≤ −zT2 (K3 − I)z2 − 2 − z2
2 − z2
− signals z1 and z2 will remain within the compact sets z1 , z2 ,
i=1
kbikai 2i1i i=1 respectively, defined by

n  
σi  ∗ 4
2        
Wi + W̃i 4 − 2Wi∗  W̃i 
2
+
2   
4 z1 := z1 ∈ R3  z1i ≤ kai 2 1 − e−D , i = 1, 2, 3 (29)
i=1
1 ! " #
+ ¯ (Z) 2 . (24) 
3 D
2 z2 := z2 ∈ R z2 ≤ , i = 1, 2, 3
λmin (M)
$  
From Lemma 2, we can get
  
        si   z2 ∈ R3  z2 ≤ kbi 2 1 − e−D , i = 1, 2, 3
W̃i  =    
Ŵi − Wi∗  ≤ Ŵi  + Wi∗  ≤ + Wi∗ . (25)
σi
(30)
Thus, we can obtain
where D = 2(V3 (0) + C/ρ), ρ and C are two positive
V̇3 ≤ − ρV3 + C (26) constants.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HE et al.: ADAPTIVE NN CONTROL OF AN UNCERTAIN ROBOT WITH FULL-STATE CONSTRAINTS 5

TABLE I
PARAMETERS OF THE ROBOT

and
G11 = (m1 lc2 + m2 l1 )g cos q1 + m2 lc2 g cos(q1 + q2 )
Fig. 1. Schematic of the two-degree of freedom knee rehabilitation robotic
G21 = m2 lc2 g cos(q1 + q2 )
system.  
From the above analysis, we can conclude that the signals M11 = m1 lc1
2
+ m2 l12 + lc22
+ 2l1 lc2 cos q2 + I1 + I2
 
z1 , z2 , W̃i (i = 1, 2, . . . , n) are SGUB. From the boundedness
M12 = m2 lc22
+ l1 lc2 cos q2 + I2
of x1r in Assumption 1, we know that x is bounded. Since ẋ1r  
is also bounded, it follows that α1 is bounded and in turn x2 M21 = m2 lc22
+ l1 lc2 cos q2 + I2
is bounded. With Wi∗ (i = 1, 2, . . . , n) as slow time varying,
we know that Ŵi (i = 1, 2, . . . , n) are also bounded. For com- M22 = m2 lc2
2
+ I2
pleteness, the details of the proof, similar to [47], are provided C11 = −m2 l1 lc2 q̇2 sin q2
here. C12 = −m2 l1 lc2 (q̇1 + q̇2 ) sin q2
IV. S IMULATION C21 = m2 l1 lc2 q̇1 sin q2
Considering a robot with two revolute joints in the vertical C22 = 0.
plane as shown in Fig. 1, simulations are carried out to illus- The kinematics of the robot and the Jacobian matrix are
trate the effectiveness of the proposed control. Let mi and li be written as
the mass and length of link i, lci be the distance from joint i − 1 % & % &
φ11 J11 J12
to the center of mass of link i, as indicated in the figure, and Ii φ(q) = , J(q) = (37)
φ21 J21 J22
be the moment of inertia of link i about an axis coming out of
the page passing through the center of mass of link i, i = 1, 2. where
We define φ11 = l1 cos q1 + l2 cos(q1 + q2 )
% & % &
q θ φ21 = l1 sin q1 + l2 sin(q1 + q2 )
q= 1 = 1 . (31)
q2 θ2
J11 = −l1 sin q1 + l2 sin(q1 + q2 )
Then, we have the kinetic energy given as
J12 = −l2 sin(q1 + q2 )
1 1 1
K(q, q̇) = 2 2
m1 lc1 q̇1 + I1 q̇21 + m2 l12 q̇21 J21 = l1 cos q1 + l2 cos(q1 + q2 )
2 2 2
1 J22 = l2 cos(q1 + q2 ).
+ m2 l1 lc2 q̇1 (q̇1 + q̇2 ) cos q2 + m2 lc2
2
(q̇1 + q̇2 )2
2 Parameters of the robot are listed in Table I.
1 The initial positions of the robot are given as
+ I2 (q̇1 + q̇2 )2 . (32)
2
The potential energy is written as q1 (0) = q2 (0) = q̇1 (0) = q̇2 (0) = 0. (38)

P(q) = m1 glc2 sin q1 + m2 g[l1 sin q1 + lc2 sin(q1 + q2 )]. The desired trajectory tracking a circular path is given as
qd = [sin(0.5t), 2 cos(0.5t)]T , where t ∈ [0, tf ] and tf = 30 s.
(33) We consider the robot under external disturbance composed
Using Lagrange’s equation d/dt∂(K − P)/∂ q̇ − by the Gaussian white noise. f (t) is given as [sin(t) + 1 +
∂(K − P)/∂q = 0, the dynamics of the robot can be d(t), 2 cos(t) + 0.5 + d(t)]T , where d(t) ∼ N (0, 1) is a white
expressed as (2), where Gaussian noise of power 0 dBW. As −0.783 ≤ d(t) ≤ 0.818,
% & we have f (t) ≤ [2.818, 3.3180]T , and we choose f̄ = [3, 3.4]T .
G11
G(q) = (34) Three different cases are evaluated for the simulation stud-
G21
% & ies. First, we examine the MB control designed in (15).
M11 M12 Second, the proposed adaptive neural network control with the
M(q) = (35)
M21 M22 full-state feedback (19) is considered. Finally, we conduct the
% &
C11 C12 simulation of proportional-derivative (PD) control to compare
C(q, q̇) = (36)
C21 C22 with the proposed NN control.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

6 IEEE TRANSACTIONS ON CYBERNETICS

Fig. 2. Comparison between x1 and xd (MB control). Fig. 4. Comparison between x2 and x2d (MB control).

Fig. 3. Tracking error z1 (MB control). Fig. 5. Tracking error z2 (MB control).

A. MB Control
For the MB control, in order to guarantee the state con-
straints |q| ≤ kc1 = [1.17, 2.17]T , |q̇| ≤ kc2 = [0.68, 1.18]T ,
and make the desired trajectory satisfy −X0 ≤ qd ≤
X0 , −Y0 ≤ q̇d ≤ Y0 , we choose X0 = [1, 2]T , Y0 = [0.5, 1]T ,
thus, the constraint of z1 is ka = kc1 − Y0 = [0.17, 0.17]T , the
constraint of z2 is kb = kc2 − Y0 = [0.18, 0.18]T , the control
parameters are chosen as K1 = diag[k11 , k12 ] = diag[40, 5],
K2 = diag[k21 , k22 ] = diag[28, 21]. For the approximation-
based control, a number of 256 nodes are used for each Si (Z)
with centers chosen in the area of [−1, 1]×[−1, 1]×[−1, 1]×
[−1, 1] × [−1, 1] × [−1, 1] × [−1, 1] × [−1, 1]. The initial
conditions are given as q = [0, 2]T and q̇ = [0.5, 0]T .
The tracking performance of the closed-loop system for the Fig. 6. Local amplification of Fig. 5 (t = [5, 25], z2 = [−0.2, 0.2])
(MB control).
robot are given in Figs. 2 and 4. From the two figures, we
can obtain that all the x1 and x2 can successfully track the
desired trajectory. According to Figs. 3 and 5, we can state
the system errors are converging to a small value close to zero. |q| ≤ kc1 = [1.17, 2.17]T , |q̇| ≤ kc2 = [0.68, 1.18]T , and
And Fig. 6 is the local amplification of Fig. 5 (t = [5, 25], make the desired trajectory satisfy −X0 ≤ qd ≤ X0 ,
z2 = [−0.2, 0.2]). The corresponding control input is given −Y0 ≤ q̇d ≤ Y0 , we choose X0 = [1, 2]T , Y0 = [0.5, 1]T ,
in Fig. 7. thus, the constraint of z1 is ka = kc1 − Y0 = [0.17, 0.17]T , the
constraint of z2 is kb = kc2 − Y0 = [0.18, 0.18]T , the control
parameters are chosen as K1 = diag[40, 1], K2 = diag[5, 16],
B. Adaptive Neural Network Control K3 = diag[70, 19], σ1 = 0.001, σ2 = 0.001,
1 = 50 × I, and
For the adaptive neural network control with full-
2 = 100 × I. A number of 256 nodes are chosen as well. The
state feedback, in order to guarantee the state constraints initial conditions are given as q = [0, 2]T and q̇ = [0.5, 0]T .
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HE et al.: ADAPTIVE NN CONTROL OF AN UNCERTAIN ROBOT WITH FULL-STATE CONSTRAINTS 7

Fig. 7. Control input τ (MB control). Fig. 10. Comparison between x2 and x2d (NN control).

Fig. 8. Comparison between x1 and xd (NN control). Fig. 11. Tracking error z2 (NN control).

Fig. 9. Tracking error z1 (NN control).

Fig. 12. Local amplification of Fig. 11 (t = [5, 30], z2 = [−0.2, 0.2])
(NN control).

Results of the simulation are shown in Figs. 8–13.

Figs. 8 and 10 show that x1 and x2 can successfully track
NN control, and the control gains of the PD control are chosen
the desired trajectory. From Figs. 9 and 11, we can state that
as Kp = diag[40, 4] and Kd = diag[0.4, 1].
the system errors are converging to a small value close to
The simulation results are shown in Figs. 14–18. It can
zero. Fig. 12 is the local amplification of Fig. 11 (t = [5, 30],
be seen that the PD control is able to make the robotic sys-
z2 = [−0.2, 0.2]). The control input is shown in Fig. 13.
tem track the desired trajectory. From the previous cases, we
know that the trajectory tracking is achieved with the pro-
C. PD Control posed NN control, and tracking errors remain in the sets
The PD control law is designed as τ = −Kp z1 − Kd ż1 . All z1 ∈ (−0.17, 0.17) and z2 ∈ (−0.18, 0.18) when the BLF
parameters of the robotic system are the same as the proposed is used. However, when we use the PD control without the
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

8 IEEE TRANSACTIONS ON CYBERNETICS

Fig. 13. Control input τ (NN control). Fig. 16. Comparison between x2 and x2d (PD control).

Fig. 17. Tracking error z2 (PD control).

Fig. 14. Comparison between x1 and xd (PD control).

Fig. 18. Control input τ (PD control).

Fig. 15. Tracking error z1 (PD control).

V. C ONCLUSION
barrier term, and keep the same initial conditions, the state This paper studies the adaptive neural network control for an
constraints are violated. uncertain robot with unknown dynamics. Full-state feedback
From the above comparison, we can see that the errors of control design is covered in the study of the robotic system
the PD control are larger than the proposed NN control when with full-state constraints, we have employed the adaptive neu-
the parameters of the system are unknown, because of the NN ral networks to approximate the unknown model of the robot,
learning ability. Moreover, the tracking performance of the PD shut the door on the violation of the constraints, and compen-
control is not as good as the proposed NN control, and the sate for the uncertainties of the dynamic model of the robot.
PD control cannot guarantee z1 < ka or z2 < kb . By tuning the weights of the NNs, we have managed to prove
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

HE et al.: ADAPTIVE NN CONTROL OF AN UNCERTAIN ROBOT WITH FULL-STATE CONSTRAINTS 9

all the signals of the closed-loop system to be uniformly ulti- [7] Z. Li, Y. Yang, and J. Li, “Adaptive motion/force control of mobile
mately bounded. Finally, from the simulation results, we have under-actuated manipulators with dynamics uncertainties by dynamic
coupling and output feedback,” IEEE Trans. Control Syst. Technol.,
made the exciting conclusion that the proposed control is able vol. 18, no. 5, pp. 1068–1079, Sep. 2010.
to track a desired trajectory with an excellent performance. [8] Z. Li, J. Li, and Y. Kang, “Adaptive robust coordinated control of
multiple mobile manipulators interacting with rigid environments,”
Automatica, vol. 46, no. 12, pp. 2028–2034, 2010.
A PPENDIX [9] R. Cui, B. Gao, and J. Guo, “Pareto-optimal coordination of mul-
tiple robots with safety guarantees,” Auton. Robot., vol. 32, no. 3,
Proof of Theorem 1 pp. 189–205, 2012.
Proof: Multiplying (26) by eρt yields [10] Z. Li, S. S. Ge, and Z. Wang, “Robust adaptive control of coor-
dinated multiple mobile manipulators,” Mechatronics, vol. 18, no. 5,
d  ρt  pp. 239–250, 2008.
V3 e ≤ Ceρt . (39) [11] C. L. P. Chen, G.-X. Wen, Y.-J. Liu, and F.-Y. Wang, “Adaptive consen-
dt sus control for a class of nonlinear multiagent time-delay systems using
Integrating the above inequality, we obtain neural networks,” IEEE Trans. Neural Netw. Learn. Syst., vol. 25, no. 6,
  pp. 1217–1226, Jun. 2014.
C −ρt C C [12] C. L. P. Chen, Y.-J. Liu, and G.-X. Wen, “Fuzzy neural network-based
V3 ≤ V3 (0) − e + ≤ V3 (0) + . (40) adaptive control for a class of uncertain nonlinear stochastic systems,”
ρ ρ ρ IEEE Trans. Cybern., vol. 44, no. 5, pp. 583–593, May 2014.
[13] B. Xu, Z. Shi, C. Yang, and F. Sun, “Composite neural dynamic surface
Then, we have control of a class of uncertain nonlinear systems in strict-feedback form,”
IEEE Trans. Cybern., vol. 44, no. 12, pp. 2626–2634, Dec. 2014.
1 V3 (0) + Cρ [14] W. Chen and L. Jiao, “Adaptive tracking for periodically time-varying
||z2 ||2 ≤ (41) and nonlinearly parameterized systems using multilayer neural net-
2 λmin (M) works,” IEEE Trans. Neural Netw., vol. 21, no. 2, pp. 345–351,
C 1 k2 Feb. 2010.
V3 (0) + ≥ ln 2 bi2 . (42) [15] W. Chen, S. S. Ge, J. Wu, and M. Gong, “Globally stable adaptive
ρ 2 kbi z2i backstepping neural network control for uncertain strict-feedback sys-
tems with tracking accuracy known a priori,” IEEE Trans. Neural Netw.
Taking exponentials on both side of the inequality yields Learn. Syst., in press, 2015. DOI: 10.1109/TNNLS.2014.2357451.
   [16] Y.-J. Liu, S. Tong, and C. L. P. Chen, “Adaptive fuzzy control via
2 V3 (0)+ Cρ observer design for uncertain nonlinear systems with unmodeled dynam-
z2i ≤ kbi 1 − e
2 2
. (43) ics,” IEEE Trans. Fuzzy Syst., vol. 21, no. 2, pp. 275–288, Apr. 2013.
[17] B. Xu, C. Yang, and Z. Shi, “Reinforcement learning output feedback
For z1 , we have NN control using deterministic learning technique,” IEEE Trans. Neural
Netw. Learn. Syst., vol. 25, no. 3, pp. 635–641, Mar. 2014.
C 1 k2 [18] B. Ren, S. S. Ge, C.-Y. Su, and T. H. Lee, “Adaptive neural control
V3 (0) + ≥ ln 2 ai2 (44) for a class of uncertain nonlinear systems in pure-feedback form with
ρ 2 kai z1i hysteresis input,” IEEE Trans. Syst., Man, Cybern. B, Cybern., vol. 39,
   no. 2, pp. 431–443, Apr. 2009.
2 V (0)+ Cρ
z21i ≤ kai
2
1−e 3 . (45) [19] Z. Liu, C. Chen, Y. Zhang, and C. L. P. Chen, “Adaptive neural control
for dual-arm coordination of humanoid robot with unknown nonlin-
earities in output mechanism,” IEEE Trans. Cybern., vol. 45, no. 3,
Hence, z1 converges to the compact set z1 , and the bound pp. 521–532, Mar. 2015.
of z2 can be obtained similarly. [20] Y.-J. Liu and S. Tong, “Adaptive NN tracking control of uncertain
nonlinear discrete-time systems with nonaffine dead-zone input,” IEEE
Trans. Cybern., vol. 45, no. 3, pp. 497–505, Mar. 2015.
[21] Y.-J. Liu, L. Tang, S. Tong, and C. L. P. Chen, “Adaptive NN
ACKNOWLEDGMENT controller design for a class of nonlinear MIMO discrete-time sys-
The authors would like to thank the Editor-in-Chief, the tems,” IEEE Trans. Neural Netw. Learn. Syst., in press, 2015.
DOI: 10.1109/TNNLS.2014.2330336.
Associate Editor, and the anonymous reviewers for their con- [22] Y.-J. Liu, C. P. Chen, G. Wen, and S. Tong, “Adaptive neural output
structive comments which helped to improve the quality and feedback tracking control for a class of uncertain discrete-time nonlinear
presentation of this paper. systems,” IEEE Trans. Neural Netw., vol. 22, no. 7, pp. 1162–1167,
Jul. 2011.
[23] C. Yang, Z. Li, and J. Li, “Trajectory planning and optimized adaptive
R EFERENCES control for a class of wheeled inverted pendulum vehicle models,” IEEE
Trans. Cybern., vol. 43, no. 1, pp. 24–36, Feb. 2013.
[1] Y. Li, S. S. Ge, and C. Yang, “Learning impedance control for phys- [24] R. Cui, J. Guo, and Z. Mao, “Adaptive backstepping control of wheeled
ical robot–environment interaction,” Int. J. Control, vol. 85, no. 2, inverted pendulums models,” Nonlin. Dyn., vol. 79, no. 1, pp. 501–511,
pp. 182–193, 2012. 2015.
[2] L. Wang, Z. Liu, C. L. P. Chen, Y. Zhang, S. Lee, and X. Chen, [25] A. U. Levin and K. S. Narendra, “Control of nonlinear dynamical
“Energy-efficient SVM learning control system for biped walking systems using neural networks. II. Observability, identification, and
robots,” IEEE Trans. Neural Netw. Learn. Syst., vol. 24, no. 5, control,” IEEE Trans. Neural Netw., vol. 7, no. 1, pp. 30–42, Jan. 1996.
pp. 831–837, May 2013. [26] S. S. Ge, C. C. Hang, and T. Zhang, “Adaptive neural network control
[3] R. Cui, J. Guo, and B. Gao, “Game theory-based negotiation for multiple of nonlinear systems by state and output feedback,” IEEE Trans. Syst.,
robots task allocation,” Robotica, vol. 31, no. 6, pp. 923–934, 2013. Man, Cybern. B, Cybern., vol. 29, no. 6, pp. 818–828, Dec. 1999.
[4] B. Xu, “Robust adaptive neural control of flexible hypersonic flight vehi- [27] P. San, B. Ren, S. S. Ge, T. H. Lee, and J. Liu, “Adaptive neural network
cle with dead-zone input nonlinearity,” Nonlin. Dyn., in press, 2015. control of hard disk drives with hysteresis friction nonlinearity,” IEEE
DOI: 10.1007/s11071-015-1958-8. Trans. Control Syst. Technol., vol. 19, no. 2, pp. 351–358, Mar. 2011.
[5] K. P. Tee, B. Ren, and S. S. Ge, “Control of nonlinear systems with time- [28] S.-L. Dai, C. Wang, and F. Luo, “Identification and learning control of
varying output constraints,” Automatica, vol. 47, no. 11, pp. 2511–2516, ocean surface ship using neural networks,” IEEE Trans. Ind. Informat.,
2011. vol. 8, no. 34, pp. 801–810, Nov. 2012.
[6] M. Chen, S. S. Ge, and B. Ren, “Adaptive tracking control of uncertain [29] B. V. E. How, S. S. Ge, and Y. S. Choo, “Dynamic load positioning for
MIMO nonlinear systems with input constraints,” Automatica, vol. 47, subsea installation via adaptive neural control,” IEEE J. Ocean. Eng.,
no. 3, pp. 452–465, 2011. vol. 35, no. 2, pp. 366–375, Apr. 2010.
 This article has been accepted for inclusion in a future issue of this journal. Content is final as presented, with the exception of pagination.

10 IEEE TRANSACTIONS ON CYBERNETICS

[30] M. Chen, S. S. Ge, and B. Ren, “Robust attitude control of helicopters Wei He (S’09–M’12) received the B.Eng. degree in
with actuator dynamics using neural networks,” IET Control Theor. automatic control from the College of Automation
Appl., vol. 4, no. 12, pp. 2837–2854, Dec. 2010. Science and Engineering, South China University
[31] W. He, S. S. Ge, Y. Li, E. Chew, and Y. S. Ng, “Neural network control of Technology, Guangzhou, China, and the Ph.D.
of a rehabilitation robot by state and output feedback,” J. Intell. Robot. degree in electrical and computer engineering
Syst., pp. 1–17, in press, 2015. DOI: 10.1007/s10846-014-0150-6. from the National University of Singapore (NUS),
[32] Z. Wang, S. S. Ge, and T. H. Lee, “Robust motion/force control of Singapore, in 2006 and 2011, respectively.
uncertain holonomic/nonholonomic mechanical systems,” IEEE/ASME From 2011 to 2012, he joined the Department
Trans. Mechatronics, vol. 9, no. 1, pp. 118–123, Mar. 2004. of Electrical and Computer Engineering, NUS, as a
[33] Z. Li, S. S. Ge, and A. Ming, “Adaptive robust motion/force control Research Fellow. He is currently with the Center for
of holonomic-constrained nonholonomic mobile manipulators,” IEEE Robotics and the School of Automation Engineering,
Trans. Syst., Man, Cybern. B, Cybern., vol. 37, no. 3, pp. 607–616, University of Electronic Science and Technology of China, Chengdu, China.
Jun. 2007. His current research interests include distributed parameter systems, intelli-
[34] C. Su, T. P. Leung, and Q. Zhou, “Force/motion control of constrained gent control systems, and robotics.
robots using sliding mode,” IEEE Trans. Autom. Control, vol. 37, no. 5, Prof. He is an Editor of the Journal of Intelligent & Robotic Systems
pp. 668–672, May 1992. (Springer).
[35] L. Huang, S. S. Ge, and T. H. Lee, “Position/force control of uncer-
tain constrained flexible joint robots,” Mechatronics, vol. 16, no. 2,
pp. 111–120, 2006.
[36] Z. Li, S. S. Ge, M. Adams, and W. Wijesoma, “Robust adaptive
control of uncertain force/motion constrained nonholonomic mobile
manipulators,” Automatica, vol. 44, no. 3, pp. 776–784, 2008.
[37] S. S. Ge, C. C. Hang, T. H. Lee, and T. Zhang, Stable Adaptive Neural
Network Control. Boston, MA, USA: Kluwer, 2001.
[38] Y. Liu, S. Tong, D. Wang, T.-S. Li, and C. Chen, “Adaptive neural output
feedback controller design with reduced-order observer for a class of
uncertain nonlinear SISO systems,” IEEE Trans. Neural Netw., vol. 22,
no. 8, pp. 1328–1334, Aug. 2011.
[39] W. He, S. Zhang, and S. S. Ge, “Adaptive control of a flexible crane
system with the boundary output constraint,” IEEE Trans. Ind. Electron.,
vol. 61, no. 8, pp. 4126–4133, Aug. 2014.
[40] K. P. Tee, S. S. Ge, and E. H. Tay, “Barrier Lyapunov functions for the
control of output-constrained nonlinear systems,” Automatica, vol. 45, Yuhao Chen is currently pursuing the B.Eng.
no. 4, pp. 918–927, 2009. degree with the School of Automation Engineering,
[41] W. He, C. Sun, and S. S. Ge, “Top tension control of a flexible marine University of Electronic Science and Technology of
riser by using integral-barrier Lyapunov function,” IEEE/ASME Trans. China, Chengdu, China.
Mechatronics, vol. 20, no. 2, pp. 497–505, Apr. 2015. His current research interests include flexible
[42] W. He, S. S. Ge, and D. Huang, “Modeling and vibration control for robotic manipulator, neural network control, and
a nonlinear moving string with output constraint,” IEEE/ASME Trans. adaptive control.
Mechatronics, in press, 2015. DOI: 10.1109/TMECH.2014.2358500.
[43] W. He and S. S. Ge, “Vibration control of a flexible string with
both boundary input and output constraints,” IEEE Trans. Control Syst.
Technol., in press, 2015. DOI: 10.1109/TCST.2014.2362718.
[44] K. P. Tee and S. S. Ge, “Control of nonlinear systems with partial state
constraints using a barrier Lyapunov function,” Int. J. Control, vol. 84,
no. 12, pp. 2008–2023, 2011.
[45] B. Ren, S. S. Ge, K. P. Tee, and T. H. Lee, “Adaptive neural control for
output feedback nonlinear systems using a barrier Lyapunov function,”
IEEE Trans. Neural Netw., vol. 21, no. 8, pp. 1339–1345, Aug. 2010.
[46] W. He, S. S. Ge, B. V. E. How, and Y. S. Choo, Dynamics and Control of
Mechanical Systems in Offshore Engineering. London, U.K.: Springer,
2014.
[47] S. S. Ge and C. Wang, “Adaptive neural network control of uncertain
MIMO non-linear systems,” IEEE Trans. Neural Netw., vol. 15, no. 3,
pp. 674–692, May 2004.
[48] W. Meng, Q. Yang, D. Pan, H. Zheng, and G. Wang, “NN-based
Zhao Yin (S’14) received the B.Eng. degree in
asymptotic tracking control for a class of strict-feedback uncertain non-
electric engineering and automation from Hangzhou
linear systems with output constraints,” in Proc. 51st IEEE Conf. Decis.
Dianzi University, Zhejiang, China, in 2013. He is
Control, Maui, HI, USA, 2012, pp. 5410–5415.
currently pursuing the M.E. degree with the School
[49] Z. Zhao, W. He, and S. S. Ge, “Adaptive neural network control of a fully
of Automation Engineering, University of Electronic
actuated marine surface vessel with multiple output constraints,” IEEE
Science and Technology of China, Chengdu, China.
Trans. Control Syst. Technol., vol. 22, no. 4, pp. 1536–1543, Jul. 2014.
His current research interests include neural net-
[50] S. S. Ge, T. H. Lee, and C. J. Harris, Adaptive Neural Network Control
work control, adaptive control, and robotics.
of Robotic Manipulators. Singapore: World Scientific, 1998.
[51] J. Slotine and W. Li, Applied Nonlinear Control. Englewood Cliffs, NJ,
USA: Prentice Hall, 1991.