You are on page 1of 11

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 SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS 1

Adaptive Neural Impedance Control of a Robotic


Manipulator With Input Saturation
Wei He, Member, IEEE, Yiting Dong, Student Member, IEEE, and Changyin Sun

Abstract—In this paper, adaptive impedance control is devel- can compensate for the input saturation effect of the robotic
oped for an n-link robotic manipulator with input saturation system and maintain the stability of the robot.
by employing neural networks. Both uncertainties and input During the past decade, the control system of the robotic
saturation are considered in the tracking control design. In
order to approximate the system uncertainties, we introduce manipulator has been developed in a variety of robotic appli-
a radial basis function neural network controller, and the cations, such as rehabilitation, home service, education, and
input saturation is handled by designing an auxiliary sys- entertainment. Most of the robotic manipulator systems are
tem. By using Lyapunov’s method, we design adaptive neural integrated with fixed platforms, resulting in achieving physical
impedance controllers. Both state and output feedbacks are con- robot–environment interaction. The main challenge in this field
structed. To verify the proposed control, extensive simulations are
conducted. is to investigate the physical robot–environment interaction
problem intensively, which has become popular on the devel-
Index Terms—Adaptive neural network (NN) control, opment side [6], [7]. Focusing on the need of safe, fast, and
impedance control, input saturation, learning control, nonlinear
system, robot. accurate interaction, the position or velocity tracking method is
not essential in the robotic control task. Impedance control was
first proposed by Hogan [2]. The basic principle of this pro-
I. I NTRODUCTION posed method is to regulate the interaction between the manip-
N RECENT years, the control design and stability anal- ulator end effector and the force exerted on the environment
I ysis of a robot have received considerable attention. The
position or force control systems cannot be formulated eas-
by specifying a desired relation or impedance force, rather
than track the motion and force trajectory [8]. Many works
ily in order to control the robotic manipulator, given that have been done on learning impedance control in robotic plat-
the position and force control objectives cannot be consid- forms. In [9], an iterative learning impedance control problem
ered separately [1]. An impedance control is proposed so is formulated and solved for robotic manipulators. The learn-
that the control objective should not track the position or ing impedance control is designed such that the robotic control
force trajectory only but rather involve the regulation of the system has the ability to follow the desired response speci-
impedance of the robot end effector which relates both posi- fied by the target model as the actions are repeated. A robust
tion and force [2]–[4]. The main challenge of the impedance impedance control is proposed in [2], and this algorithm regu-
control design is the situation when it fails to provide ade- lates the dynamic behavior at the interaction point. In [10], two
quate power due to the input nonlinearities, especially input adaptive impedance controllers are presented, and it assumes
saturation. Saturation nonlinearity can be caused by the con- that some parameters in robotic manipulator dynamics may
straints of magnitude and the rate of the actuator inputs [5]. be uncertain. The robotic system is stable and the tracking
Higher overshoot and larger tracking errors may result from performance is guaranteed in the above methods, but these
neglecting the input saturation effect for a robotic system. It algorithms all require knowledge of robotic dynamics.
is therefore crucial to design an impedance controller which For the purpose of solving the problem, learning control is
presented which requires less knowledge [11]–[14]. In [15],
Manuscript received February 5, 2015; accepted April 2, 2015. This work adaptive neural networks (NNs) are used to approximate the
was supported in part by the National Natural Science Foundation of China unknown model of a rehabilitation robot. Colbaugh et al. [16]
under Grant 61203057 and Grant 61125306, in part by the National Basic presented an adaptive scheme for controlling the end-effector
Research Program of China (973 Program) under Grant 2014CB744206, in
part by the National High Technology Research and Development Program impedance of the robot manipulator. Adaptive control has been
of China (863 Program) under Grant 2015AA042304, and in part by the developed using a variable structure method with the presence
Fundamental Research Funds for the China Central Universities of the of uncertainties in [17]. In [18], target impedance reference
University of Electronic Science and Technology of China under Grant
ZYGX2013Z003 and Grant Y03001023601034. This paper was recommended trajectory is used to develop an NN impedance control effi-
by Associate Editor Z. Liu. ciently. In [19], adaptive control strategies are presented for
W. He and Y. Dong are with the School of Automation Engineering and mobile manipulators in the presence of uncertainties and dis-
Center for Robotics, University of Electronic Science and Technology of
China, Chengdu 611731, China (e-mail: weihe@ieee.org). turbances. For a class nonlinear systems, an adaptive fuzzy
C. Sun is with the School of Automation, Southeast University, control scheme [20] and adaptive NN control [21] are pro-
Nanjing 210096, China. posed to realize the control objective. A hybrid motion/force
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. control based on an adaptive NN is proposed successfully
Digital Object Identifier 10.1109/TSMC.2015.2429555 in [22] to compensate for the perturbations. In [23], a stable
2168-2216 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 SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

adaptive NN tracking control design technique is proposed for In order to suppress the riser’s vibration in the presence of
nonaffine nonlinear systems. In [24], an adaptive control for a top tension constraint, a boundary controller is studied via
class of nonlinear stochastic systems with unknown functions integral Lyapunov function in [53]. In addition, some out-
is studied. Lewis et al. [25] and Chen et al. [26] have ini- put information is not measurable in actual platforms. In this
tially applied backstepping control of nonlinear systems using paper, an auxiliary system is introduced to handle the effect
NN to a robotic platform. NNs have an important property of saturation.
of function approximation and can be used to approximate Based on the discussion above, we present adaptive neu-
nonlinear functions and unknown system dynamics [27]–[30]. ral impedance control for robot–environment interaction. The
In [31], to realize biped walking, a neural learning control primary contributions of this paper are summarized as follows.
system considering the energy cost is presented. 1) To solve the input saturation and external disturbance,
Those proposed controllers do not require the knowledge of we have designed an effective impedance controller
the structure of robotic systems and are implemented without which based on the Lyapunov function. Apart from that,
calculation of the robot inverse kinematic transformation. In the structures in both state and output feedbacks are
most of the control algorithms, it is assumed that the system presented.
states are available. However, it may not be practical to mea- 2) A learning approach for robotic control systems is pro-
sure all the values of system states. In [32], two adaptive NN posed, which does not require knowledge of robotic
decentralized output feedback control schemes are proposed dynamics. Furthermore, the stability of closed-loop sys-
for a class of nonlinear systems. Liu et al. [33] proposed tem can be ensured using Lyapunov stability theorem.
an adaptive output feedback control for uncertain nonlinear 3) The bounds of NN approximation errors, NN weights
single-input single-output systems. In [34], an adaptive neural and radial basis functions (RBFs) are not necessarily
output feedback tracking control of nonlinear multiple-input made known for control design in the design process.
multiple-output systems is studied. In [35] and [36], the high The rest of this paper is organized as follows. In Section II,
gain observer is used to estimate unmeasurable system states the preliminaries and system dynamics are presented.
effectively. A reduced-order observer is designed to estimate Section III illustrates the direct adaptive neural impedance
immeasurable states in [37]. control design process under the state and output feedback
Another challenge of the robotic manipulator control design methods. Section IV shows the performance of proposed
is how to deal with the input saturation which may result in methods by simulations. The last section concludes this paper.
higher overshoot and larger tracking errors. A number of the
methods are being used to handle the control design related II. P RELIMINARIES AND P ROBLEM F ORMULATION
with saturation. Santibañez et al. [38] proposed a saturated A. Preliminaries
nonlinear PID controller for industrial robotic manipulators.
Lemma 1 [54]: For bounded initial conditions, if there
However, the controller is based on the assumption that the
exists a C1 continuous and positive definite Lyapunov func-
information of saturator and system dynamics are known.
tion V(x) satisfying κ1 (x) ≤ V(x) ≤ κ2 (x), such that
Huang et al. [39] deal with the problem of global tracking
V̇(x) ≤ −κV(x) + c, where κ1 , κ2 : Rn → R are class K func-
and stabilization control of a robot, in the presence of input
tions, κ and c are two positive constants, then the solution x(t)
saturation and external disturbances. In [40], a learning control
is uniformly bounded.
scheme is proposed for nonlinear uncertain systems with input
Lemma 2 [55]: Consider the basis functions of Gaussian
saturation successfully. In [41] and [42], the adaptive control
RBFNN with Ẑ being the input vector, if Ẑ = Z −  ψ̄, where
is presented to handle the external disturbance, saturation and
ψ̄ is a bounded vector and constant  > 0, then we have
deadzone for nonlinear systems. Liu et al. [43] investigated the ⎡  T  ⎤
unknown direction hysteresis model with adaptive NN control. − Ẑ − μj Ẑ − μj
⎢ ⎥
Adaptive NN control is presented to compensate the effect sj (Ẑ) = exp⎣ ⎦, j = 1, 2, . . . , l
of output constraint for an uncertain robot in [44]. In [45], ηj2

adaptive tracking control is presented for a class of uncer-


tain nonlinear systems with input saturation. In [46], adaptive S(Ẑ) = S(Z) + St (1)
fuzzy controllers are designed via backstepping for nonlinear where St is a bounded vector function.
strict feedback systems with unknown deadzone and immea- Lemma 3 [56]–[58] (Rayleigh–Ritz Theorem): Let A ∈
surable states. Guo et al. [47] proposed two memristor-based Rn×n be a real, symmetric, and positive-definite matrix; there-
recurrent NNs with time delay via static or dynamic coupling. fore, all the eigenvalues of A are real and positive. Let λmin
In [48], an adaptive NN control method for the nonlinear and λmax denote the minimum and maximum eigenvalue of A,
multiagent systems with time-delay is proposed. In [49], in respectively; then for ∀x ∈ Rn , we have
the presence of input saturation and unknown external distur-
λmin ||x||2 ≤ xT Ax ≤ λmax ||x||2 (2)
bance, two robust adaptive control algorithms are developed.
Besides the works on ordinary differential equation systems where || · || denotes the standard Euclidean norm.
with input saturation, partial differential equation systems with In practice, the exact model of robot dynamics is not known.
constraints [50], [51] have also been studied in recent years. To approximate the exact values of dynamic terms, a Gaussian
He and Ge [52] presented boundary control to handle the effect RBFNN can be utilized as it can approximate the smooth
of output constraint for the string system by Lyapunov method. function [1].
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 NEURAL IMPEDANCE CONTROL OF ROBOTIC MANIPULATOR WITH INPUT SATURATION 3

B. Dynamics of the Robotic System A. Full State Feedback Control


When the robot comes in contact with the environment, We first consider the case where full state information
an interactive force develops between the robot and envi- x is available. Adaptive NN control is applied to approximate
ronment according to some user-defined dynamics, the target the unknown model of the robotic manipulator. Define the
impedance. The target impedance between the robot position Cartesian space signals as
x and the vector of forces and moments exerted by the robot
z1 = xd − x (6)
on the environment τe ∈ Rn is expressed as [1]
α1 = ẋd + K1 z1 (7)
Dd (ẍc − ẍ) + Cd (ẋc − ẋ) + Gd (xc − x) = τe (3)
z2 = α1 − ẋ = ż1 + K1 z1 . (8)
where xc ∈ Rn are the commanded trajectories which are
Taking the contact force τe into consideration, the Cartesian
bounded, smooth and twice differentiable, and the Dd , Cd , and
space dynamic NN based on adaptive control is modified, sup-
Gd ∈ Rn×n are also specified by the user. Dd , Cd , and Gd are
posing the dynamic model and the desired impedance model
the mass, damping, and stiffness matrices of the mass-spring
are completely known, by adding the contact force τe , to
damper system which provides the impedance relationship
between the robot contact forces τe and the position errors τ (t) = −z1 + D̂α̇1 + Ĉα1 + Ĝ
xc − x. + Kp (z2 + ζ ) + Kr sgn(z2 ) + τe (9)
It can be easily seen that when the robot is required to
move in free space, we have xc = x, ∀t > 0 and τe = 0. where Kr = diag[krii ], D̂ = D, Ĉ = C, Ĝ = G, and
τ =
On the other hand, while in contact with the environment, S(τ ) − τ (t). In actual implementation, the control input τ that
the end-effector response is to comply with the contact forces must be applied to the joint.
which defined by the target impedance (3). In general, the Substituting (9) into (4), we have
matrices Dd , Cd , and Gd are diagonal with constant posi- − Dż2 = −z1 + Cz2 + (D̂ − D)α̇1 + (Ĉ − C)α1 +
τ
tive elements. Nevertheless, there is considerable freedom to  
choose these matrices, which is one of the attractive features + Ĝ − G + Kp (z2 + ζ ) + Kr sgn(z2 ) (10)
of the impedance control scheme. where
τ = S(τ ) − τ (t). The following auxiliary design
The dynamics of an n-link rigid robotic system in Cartesian system is used to reduce the saturation effects:
space is described by [1] ⎧ T 1 T
⎨ z
τ +
τ
τ
2
D(q)ẍ + C(q, q̇)ẋ + G(q) = S(τ ) − τe (4) ζ̇ = −K ζ ζ − 2
ζ +
τ, ||ζ || ≥ μ
⎩ ||ζ || 2

where q ∈ Rn is the coordinate, τ ∈ Rn is the applied joint 0, ||ζ || < μ


torque, D(q) ∈ Rn×n is a symmetric positive definite inertia (11)
matrix, C(q, q̇)ẋ ∈ Rn denotes the Centripetal and Coriolis where Kζ = KζT > 0, μ is a small positive value and ζ ∈ Rn
torques, and G(q) ∈ Rn is the gravitational force, τe ∈ Rn is the state of auxiliary design system.
is the vector of constraint force exerted by the environment, Considering Lyapunov function candidate V1 = (1/2)zT1 z1 +
which is 0 when there is no contact between the manipulator (1/2)zT2 Dz2 + (1/2)ζ T ζ and the above auxiliary design func-
and environment. tion (11). Time derivative of V1 is
Property 1 [1], [59]: The matrix D(q) is symmetric and
1
positive definite. V˙1 = zT1 ż1 + zT2 Dż2 + zT2 Ḋz2 + ζ T ζ̇ . (12)
Property 2 [1], [59]: The matrix Ḋ(q) − 2C(q, q̇) is skew- 2
symmetric. Since ζ T
τ ≤ 1/2ζ T ζ + 1/2
τ T
τ , we have ζ T ζ̇ =
The desired impedance model given in the joint space is −ζ T Kζ ζ −|zT2
τ |−0.5
τ T
τ +ζ T
τ ≤ −ζ T Kζ ζ −|zT2
τ |+
given as 1/2ζ T ζ . Time derivative of V1 will then be
Dd (ẍd − ẍ) + Cd (q, q̇)(ẋd − ẋ) + Gd (xd − x) = −τe (5) 1
V˙1 ≤ −zT1 K1 z1 − zT2 (2Kp − I)z2
2
where xd is the desired trajectory, Dd , Cd , and Gd are the 1  
desired inertia, damping, and stiffness matrices, respectively. − ζ T 2Kζ − I − KpT Kp ζ. (13)
2
Assumption 1: The desired trajectories ||xd || ≤ d with To ensure the closed loop stability, control parameters Kp and
known constant d > 0 is smooth and differentiable. Therefore, Kζ should fulfill the following criteria: K1 > 0, 2Kp − I > 0,
define the time derivative of xd as: ||ẋd || ≤ d1 , ||ẍd || ≤ d2 . and 2Kζ − I − KpT Kp > 0. So V˙1 will be negative definite.
The control objective is to design control torques such that Since the uncertainties exist in D(q), C(q, q̇), and G(q), the
the impedance of the whole system tracks the given target model-based control design may not be realizable. Thus, the
impedance model (5). above model-based control is not applicable for the robot with
uncertainties. To overcome the challenge, NNs are utilized to
III. C ONTROL D ESIGN approximate the uncertainties and improve the performance of
Both state and output feedback schemes are presented in the system via the online estimation.
this section. The impedance control is aimed to reduce the Impedance control based on NN can improve the perfor-
saturation effect and track the desired trajectory. mance of the system even though the impedance model of
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 SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS



the system is time-varying. We will design the adaptive NN
zT2 W̃CT SC α1 = T
W̃Dk SCk α1 z2k (25)
impedance control to approximate the unknown model of the
k=0
constrained robot and adapt interactions between robots and ∞

humans. zT2 W̃GT SG = T
W̃Gk SGk z2k . (26)
We design the controller as k=0
τ (t) = −z1 + ŴDT SD (ZD )α̇1 + ŴCT SC (ZC )α1 + τe
Since −W̃ T Ŵ = −W̃ T (W ∗ + W̃) = −W̃ T W̃ − W̃ T W ∗
+ ŴGT SG (ZG ) + Kp (z2 + ζ ) + Kr sgn(z2 ). (14) and −W̃ T W ∗ ≤ (1/2)(W̃ T W̃ + W ∗T W ∗ ), we have
The NN updating laws are designed as −W̃ T Ŵ ≤ −(1/2)W̃ T W̃ + (1/2)W ∗T W ∗ . Meanwhile, consid-
  ering (24)–(26), with krii ≥ ||Exi ||, we have
˙ = S α̇ z − σ Ŵ
Ŵ (15)
Dk Dk Dk 1 2k Dk Dk
  1
˙ = S α z − σ Ŵ V̇2 ≤ −zT1 K1 z1 − zT2 (2Kp − I)z2
Ŵ (16) 2
1 T 
Ck Ck Ck 1 2k Ck Ck
 
˙ = S z − σ Ŵ − ζ 2Kζ − I − KpT Kp ζ
Ŵ Gk Gk Gk 2k Gk Gk (17) 2
∞ ∞
where Dk > 0, Ck > 0, Gk > 0, and σi is a small posi- 1  
− σDk W̃Dk
T
W̃Dk + σCk W̃Ck
T
W̃Ck
tive constant for improving the robustness [51]. ŴDT SD (ZD ) is 2
k=0 k=0
an approximation of WD∗T SD (ZD ), ŴCT SC (ZC ) is an approxi- ∞


mation of WC∗T SC (ZC ) and ŴGT SG (ZG ) is an approximation of + σGk W̃Gk W̃Gk
T

WG∗T SG (ZG ) and  ∞k=0 ∞


WD∗T SD (ZD ) = D + D (18) 1  ∗T ∗

∗T ∗
+ σDk WDk WDk + σCk WCk WCk
WC∗T SC (ZC ) = C + C (19) 2
k=0 k=0


WG∗T SG (ZG ) = G + G . (20) 
∗T ∗
+ σGk WGk WGk
Substituting the design controller (14) into (4), we have k=0
≤ −κV2 + C (27)
− Dż2 = W̃DT SD α̇1 + W̃CT SC α1 + W̃GT SG
+ Kp (z2 + ζ )
+ Kr sgn(z2 ) + Ex +
τ − z1 + Cz2 (21) where
  
where Kr = diag[krii ] and Ex = D α̇1 +C α1 +G . Considering κ = min 2K1 , λmin (2Kp − I), λmin 2Kζ − I − KpT Kp
the Lyapunov function candidate ⎞
1  T −1
∞ σDk σCk σGk
V2 =
1 T 1 1
z1 z1 + zT2 Dz2 + ζ T ζ + W̃Dk Dk W̃Dk ×  ,  ,   ⎠ (28)
−1 −1 −1
2 2 2 2 λmax Dk λmax Ck λmax Gk
k=0 ∞ ∞
∞ ∞
1 T −1 1 T −1 1  ∗T ∗

∗T ∗
+ W̃Ck Ck W̃Ck + W̃Gk Gk W̃Gk . (22) C= σDk WDk WDk + σCk WCk WCk
2 2 2
k=0
k=0 k=0

k=0
Time derivative of V2 is 
∗T ∗
+ σGk WGk WGk . (29)
V̇2 = −zT1 K1 z1 − zT2 Cz2 k=0

− zT2 W̃DT SD α̇1 + W̃CT SC α1 To ensure the closed loop stability, control parameters K1 ,

+ W̃GT SG + Kp (z2 + ζ ) + Kr sgn(z2 ) +
τ + Ex Kp , and Kζ should fulfill the following criteria: K1 > 0,

 2Kp − I > 0 and 2Kζ − I − KpT Kp > 0. So V̇2 will be negative
1
+ zT2 Ḋz2 + ζ T ζ̇ + T
W̃Dk SDk α̇1 z2k definite. Therefore, we have the following theorem.
2 Remark 1 [1]: In the proposed controller (14), the gain of
k=0

 ∞
 the term Kr , is chosen such that krii > ||Exi ||. For stability,
+ T
W̃Ck SCk α1 z2k + T
W̃Gk SGk z2k Kr is usually chosen to be conservatively large. This is not
k=0 k=0 very desirable due to the chattering introduced. The control

 ∞
 term Kr can be changed to
− σDk W̃Dk
T
ŴDk − σCk W̃Ck
T
ŴCk
k=0 k=0 Kr = kD α̇1 + kC α1 + kG (30)


− σGk W̃Gk
T
ŴGk (23) where kD ≥ ||D ||, kC ≥ ||C ||, and kG ≥ ||G ||. The gain
k=0 has the following advantages: 1) the values of elements kD ,
where Property 2 has been used, and noting that kC , and kG are only needed to be large enough to suppress the
∞ bounded approximation NN errors D , C , and G , respectively

zT2 W̃DT SD α̇1 = T
W̃Dk SDk α̇1 z2k (24) and 2) it is a function of α̇1 and α1 so that it decreased as
k=0
α̇1 and α1 diminish.
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 NEURAL IMPEDANCE CONTROL OF ROBOTIC MANIPULATOR WITH INPUT SATURATION 5

Theorem 1: For system dynamics described by (4), under Time derivative of V3 is


Assumption 1, and the control (14) with the adaptation
laws (15)–(17), given that the initial conditions are bounded, V̇3 = −zT1 K1 z1 + zT2 ((D − D̂)α̇1 + (C − Ĉ)α1
 
we can conclude that the desired impedance is achieved and + G − Ĝ − (C + Kp )z2 ) − zT2 Kp ζ2 − zT2 Kr sgn(z2 )
the impedance error will eventually converge to a small neigh-
1 −1 ˙
borhood around zero by appropriately choosing design param- − zT2
τ + zT2 Ḋz2 + ζ2T ζ˙2 + W̃DT D ŴD
eters. Furthermore, tracking error z1 converges asymptotically 2
√ ˙ + W̃ T −1 Ŵ
˙
to the compact set z1 := {z1 ∈ R2 | ||z1 ≤ B||} where + W̃CT C−1 Ŵ C G G G (41)
B = 2(V2∗ (0) + C/κ) with κ and C given in (28) and (29).
The proof of Theorem 1 is given in the Appendix. where
     
zT2 D − D̂ α̇1 + C − Ĉ α + G − Ĝ
B. Output Feedback Control  
The proposed control (9) is designed under the assumption = zT2 WD∗T SD − W̃DT ŜD − WD∗T ŜD α̇1
that all outputs are measurable. However, some output infor-  
mation may not be measurable due to practical issues such as + zT2 WC∗T SC − W̃CT ŜC − WC∗T ŜC α1
cost and dimensions. In this section, a high gain observer [60]  
+ zT2 WG∗T SG − W̃GT ŜG − WG∗T ŜG . (42)
is used to estimate the unmeasurable terms. According to [15],
the unmeasurable state ẋ can, therefore, be approximated as
Considering the effect of estimation error z̃2 , V = V3 + Vobs
π2 /γ . The estimate of the unmeasurable state vector z2 is
and Vobs = 1/2z̃T2 z̃2 . Based on (33), we have
designed as
π2 1 2 2
ẑ2 = − α1 (31) Vobs ≤ γ h2 . (43)
ε 2
where the dynamics of π2 are described as We just need to check the stability of V3 term as Vobs is stable.
επ̇1 = π2 (32) According to [1] and [55], we have
επ̇2 = −λ̄1 π2 − π1 + x (33) ŜD − SD = γ̄D SD , ŜC − SC = γ̄C SC , ŜG − SG = γ̄G SG (44)
where ε is any small constant. According to [15], there
where SD , SC , and SG are bounded vector functions with sat-
exist positive constants t∗ and rn such that ∀t > t∗ , we ∗ , ||S || < S∗ , ||S || < S∗ and γ̄ > 0,
isfying ||SD || < SD C G D
have |βn | ≤ εrn . So the z̃2 can be estimated as follows: C G
γ̄C > 0 and γ̄G > 0.
π2 Considering the first term of (42), we have
z̃2 = ẑ2 − z2 = − α1 − ẋ + α1 = βn . (34)
ε  
The control design is rewritten as follows: zT2 WD∗T SD − W̃DT ŜD − WD∗T ŜD α̇1
 
τ (t) = ŴDT ŜD α̇1 + ŴCT ŜC α1 + ŴGT ŜG = z̃T2 W̃DT ŜD ẍd − z̃T2 W̃DT ŜD K12 z1 + z̃T2 W̃DT ŜD K1 z2
   
+ Kp ẑ2 + ζ2 + Kr sgn ẑ2 + τe − z1 . (35) − ẑT2 W̃DT ŜD α̇1 − zT2 WD∗T γ̄D SD ẍd
The following auxiliary design system is used to examine the + zT2 WD∗T γ̄D SD K12 z1 − zT2 WD∗T γ̄D SD K1 z2 . (45)
saturation effect:
 Substituting ||Ŝi ||2 ≤ li into the above function, we have
|ẑT2
τ |+ 12
τ T
τ
ζ̇2 = −Kζ ζ2 − ||ζ2 ||2
ζ2 +
τ, ||ζ2 || ≥ μ
0, ||ζ2 || < μ. lD θ1 ||γ h2 ||2 2 d22
z̃T2 W̃DT ŜD ẍd ≤ W̃D +
(36) 4 θ1

lD θ2 K1 ||γ h2 || 2
4 2 1
The NN updating laws are designed as −z̃T2 W̃DT ŜD K12 z1 ≤ W̃D + zT1 z1
4 θ2
 
˙ = Ŝ α̇ ẑ − σ Ŵ
Ŵ (37) lD θ3 K12 ||γ h2 ||2 2 1
D D D 1 2 D D z̃T2 W̃DT ŜD K1 z2 ≤ W̃D + zT2 z2 . (46)
  4 θ3
˙ = Ŝ α ẑ − σ Ŵ
Ŵ (38)
C C C 1 2 C C
  Considering the above function, we have
˙ = Ŝ ẑ − σ Ŵ .
Ŵ (39)
G G G 2 G G
−zT2 WD∗T γ̄D SD ẍd + zT2 WD∗T γ̄D SD K12 z1 − zT2 WD∗T γ̄D SD K1 z2
Differentiating (8) yield ż2 = α̇1 − ẍ = α̇1 − D−1 (S(τ ) − G − 1 d 2 2 1
Cẋ − τe ). Considering the Lyapunov function candidate ≤ zT2 z2 + 2 WD∗ ||γ̄D SD ||2 + zT2 z2
2 2 2
1 T 1 1 1 1 T ∗ 2
V3 = −1
z z1 + zT2 Dz2 + ζ2T ζ2 + W̃DT D W̃D + z1 WD ||γ̄D SD || ||K1 || z1
2 4
2 1 2 2 2 2
1 1 2
1 1
+ W̃CT C−1 W̃C + W̃GT G
−1
W̃G . (40) + zT2 z2 + zT2 WD∗ ||γ̄D SD ||2 ||K1 ||2 z2 . (47)
2 2 2 2
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 SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

Therefore, we have (1/2)Wi∗T Wi∗ . Substituting these inequalities into (51),


  we have
zT2 WD∗T SD − W̃DT ŜD − WD∗T ŜD α̇1 
  θ2 + θ5 1  ∗ 2
1 1 2 V̇3 ≤ zT1 + WD ||γ̄D SD ||2 K14
≤ zT1 + WD∗ ||γ̄D SD ||2 ||K1 ||4 z1 θ2 θ5 2

θ2 2 ∗ 2 
 
1 ∗ 2 + WC ||γ̄C SC || K1 − K1 z1
2 2
1 3
+ z2 T
+ I+ WD ||γ̄D SD || ||K1 || z2
2 2
θ 2 2
 3   1 T T
lD ||γ h2 ||2 θ1 + θ2 K14 + θ3 K12 2 + ζ K Kp ζ2
+ W̃D 2 2 p 
4 T 1 1 ∗ 2
  + z2 + 3.5I − Kp + WD || ||
γ̄D SD K1 z2
2 2

θ3 2
d22 d22  
+ ||γ̄D SD ||2 WD∗ .
2
+ (48) lG θ6 ||γ h2 ||2 σG 2 d22 d2 1
θ1 2 + − W̃G + + 1 +
4 2 θ1 θ4 θ6
   
Considering the second term of (42), we have lD ||γ h2 ||2 θ2 K14 + θ3 K12 + θ4 σD 2
  + − W̃D
4 2
zT2 WC∗T SC − W̃CT ŜC − WC∗T ŜC α1  
 
θ4 2 d 2 lC ||γ h2 ||2 θ4 + θ5 K12 σC 2 1 T

1
lC ||γ h2 ||2 W̃C + 1 + zT1 z1 + − W̃C + ζ2 ζ2
4 θ4 θ5 4 2 2
2   2
θ5 1 2
+ lC ||γ h2 ||2 ||K1 ||2 W̃C − ẑT2 W̃CT ŜC α1 + d ||γ̄D SD ||2 + σD WD∗ − ζ2T Kζ ζ2
4 2 2
d2 2 1 2 
d1 ||γ̄C SC ||2 + σC WC∗
1 T 1 2
+ z2 z2 + 1 WC∗ ||γ̄C SC ||2 + zT2 z2 +
2
1 
2 2 2
1 2 + ||γ̄G SG ||2 + σG WG∗
2
+ zT1 WC∗ ||γ̄C SC ||2 ||K1 ||2 z1 . (49) 2
2
≤ −κV3 + C (52)
Considering the last term of (42), we have
  θ 2 1 where
zT2 −W̃GT ŜG − WG∗T γ̄G SG ≤ lG ||γ h2 ||2 W̃G +
6

θ6 
1  ∗ 2
4
θ2 + θ5
1 1 2 κ = min⎝2λmin − − WD ||γ̄D SD ||2 K14
+ zT2 z2 + WG∗ ||γ̄G SG ||2 θ2 θ5 2
2 2
− ẑ2 W̃G ŜG .
T T  
(50) ∗ 2

− WC ||γ̄C SC || K1 + K1 , 2 2

Therefore, time derivative of V3 is  


 1 1 ∗ 2
1  ∗ 2 2λmin − − 3.5I + Kp − WD ||γ̄D SD || K1 ,
2 2
T θ2 + θ5 θ3
V̇3 ≤ z1 + WD ||γ̄D SD ||2 K14  
2
θ2 θ5 2
  λmin −KpT Kp + 2Kζ − I ,
∗ 2

+ WC ||γ̄C SC || K1 − K1 z1
2 2  
lD ||γ h2 ||2 θ2 K14 + θ3 K12 + θ4 − 2σD
    ,
1 1 2 −1
+ zT2 + 3.5I − Kp + WD∗ ||γ̄D SD ||2 K12 z2 4λmax Dk
θ3 2  
    lC ||γ h2 ||2 θ4 + θ5 K12 − 2σC
1 lG θ6 ||γ h2 ||2 2   ,
+ ζ2T KpT Kp − 2Kζ + I ζ2 + W̃G −1
2 4 4λmax Ck
   ⎞
lD ||γ h2 ||2 θ2 K14 + θ3 K12 + θ4 2
+ W̃D lG θ6 ||γ h2 || − 2σG ⎠
2
4   (53)
−1
   4λmax Gk
lC ||γ h2 ||2 θ4 + θ5 K12 2
+ W̃C 1  d2
||γ̄D SD ||2 d22 + σD WD∗ + 2
2
4 C=
2 θ1
d22 2 d2 2  
+ ||γ̄D SD ||2 WD∗ + 1 ||γ̄C SC ||2 WC∗ 1 d2
||γ̄C SC ||2 d12 + σC WC∗ + 1
2
2 2 +
2 θ4
2 d2
1 d
+ ||γ̄G SG ||2 WG∗ + 2 + 1 +
2 1 1  1
||γ̄G SG ||2 + σG WG∗ + .
2
2 θ1 θ4 θ6 + (54)
2 θ6
− σD W̃DT ŴD − σC W̃CT ŴC − σG W̃GT ŴG . (51)
To ensure the closed loop stability, control parameters κ and
Since −W̃iT Ŵi = −W̃iT (Wi∗ + W̃i ) and −W̃iT Wi∗ ≤ C should fulfill the following criteria: κ > 0, C > 0. So V̇3
(1/2)(W̃iT W̃i + Wi∗T Wi∗ ), we have −W̃iT Ŵi ≤ −(1/2)W̃iT W̃i + will be negative definite.
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 NEURAL IMPEDANCE CONTROL OF ROBOTIC MANIPULATOR WITH INPUT SATURATION 7

TABLE I
Theorem 2: For the system dynamics described by (4), PARAMETERS OF THE ROBOT
under Assumption 1, and the control (35) with the adaptation
laws (37)–(39), given that the initial conditions are bounded,
we can conclude that the desired impedance is achieved
and the impedance error will eventually converge to a small
neighborhood around zero by appropriately choosing design
parameters. Furthermore, tracking error z1 converges asymp-

totically to the compact set z1 := {z1 ∈ R2 | ||z1 ≤ B||}
where B = 2(V3∗ (0) + C/κ) with κ and C given
in (53) and (54).
The proof of Theorem 2 is straightforward using the method
of proof in Theorem 1, and hence is omitted.

IV. S IMULATION S TUDY


In this section, we will verify the effectiveness of the pro-
posed adaptive neural control (14) and (35) for the 2-DOF
robotic manipulator with input saturation. For a robotic sys-
tem defined by (4) with two rotary degree of freedom, it can
be expressed as
Fig. 1. Mechanical model of the robotic manipulator.
ẋ1 = x2
ẋ2 = D−1 [S(τ ) − τe − G(x1 ) − C(x1 , x2 )x2 ]
y = x1 (55) trajectory is described by
   
x 0.2 − 0.2 cos(π t)
where as x1 = [q1 , q2 ]T and x2 = [q̇1 , q̇2 ]T . Considering the xd (t) = d1 = . (64)
xd2 0.2 + 0.2 sin(π t)
2-DOF robotic manipulator in a vertical plane, simulations are
conducted to verify the effectiveness of the proposed control. The radius of circle is 0.2 and the center is located at x =
Let mi and li be the mass and length of link i, lci be the [0.2 0.2]T . The obstacle is located as x1 = 0.2. The initial
distance from joint i − 1 to the center of mass of link i, as condition of the robot is
indicated in the figure, and Ii be the moment of inertia of link
i about an axis coming out of the page passing through the q(0) = [0.25 − 0.05]T q̇(0) = [0.0 0.0]T . (65)
center of mass of link i.
The unknown saturations are defined as [−17.0 17.0] and
The inertia matrix D(x1 ), centripetal and Coriolis torques
[−15.0 15.0] for q1 and q2 , respectively. The parameters of
C(x1 , x2 ), and gravitational force G(x1 ) are defined as
the target impedance are chosen as
 
p1 + p2 + 2p3 cos q2 p2 + p3 cos q2
D(x1 ) = (56) Dd = diag[1.0] Cd = diag[10.0] Gd = diag[60.0]. (66)
p2 + p3 cos q2 p2
 
−p3 q̇2 sin q2 −p3 (q̇1 + q̇2 ) sin q2 The gains of the control law are chosen as
C(x1 , x2 ) = (57)
p3 q̇1 sin q2 0
  Kp = diag[10.0] K1 = diag[6.0]. (67)
p g cos q1 + p5 g cos(q1 + q2 )
G(x1 ) = 4 (58)
p5 g cos(q1 + q2 ) The simulation objective is to examine whether the control
scheme we designed for the system (4) is able to follow the
where desired trajectory (64). We have considered three cases for the
simulation.
p1 = m1 lc1
2
+ m2 l12 + I1 (59)
p2 = 2
m2 lc2 + I2 (60)
A. Case 1: PD Control
p3 = m2 l1 lc2 (61)
The PD control law is designed as: τ (t) = Kp (z2 + ζ ) +
p4 = m1 lc2 + m2 l1 (62) Kr sgn(z2 ). When the adaptive rules given by (15)–(17) are not
p5 = m2 lc2 . (63) activated, the control performance is studied at first. It is easy
to see that the designed control action is effectively a simple
Parameters of the robotic system are defined in Table I. PD control. Fig. 2 shows the position tracking of the robot
The commanded trajectory chosen for simulation is as and Fig. 3 gives the tracking errors. The control torques are
shown in Fig. 1. We require the end effector to move along given in Fig. 4. It can be observed from these results that,
the solid part of the circle xd in free space, and to move slide the PD control has significant tracking errors, especially the
along the wall when in contact with the wall. The commanded error of x1 .
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 SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

Fig. 2. Position tracking with PD control. Fig. 5. Position tracking with state feedback control.

Fig. 3. Tracking errors with PD control. Fig. 6. Tracking errors with state feedback control.

not have complete information of the robotic system, the pro-


posed control (14) can be implemented to approximate the
system. 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], and the value of the center is
set as −1 or 1. The RBFNNs with ZD = [x1 , x2 , α1 , α̇1 ],
ZC = [x1 , x2 , α1 , α̇1 ], and ZG = [x1 , x2 , α1 , α̇1 ], the number
of nodes lD , lC , and lG for SD , SC , and SG is 28 . The gain
matrix is defined as D = 100I256×256 , C = 100I256×256 ,
and G = 100I256×256 . σD = 0.01, σC = 0.01, and σG = 0.01
and the variance for all NNs is set to be 25. RBFNN centers
for all NNs are evenly distributed in the domain of [−1, 1] ×
[−1, 1] × [−1, 1] × [−1, 1] × [−1, 1] × [−1, 1] × [−1, 1] ×
[−1, 1]. Initial weights of the NNs are all zero. The same con-
Fig. 4. Control signals with PD control.
trol gains as that in case 1 are used for the adaptive impedance
control.
The position tracking and the errors of the manipulator are
plotted in Figs. 5 and 6, respectively, and the control signals
B. Case 2: Adaptive Neural Impedance Control With are given in Fig. 7. It is obvious that the tracking errors are
State Feedback (14) much smaller than the nonadaptive case, thus, demonstrating
The new design is added to the simulation, which is involved the effectiveness of the adaptive NN impedance control in
in the state feedback control. For the situation when we do handling the unknown dynamic system with saturation effect.
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 NEURAL IMPEDANCE CONTROL OF ROBOTIC MANIPULATOR WITH INPUT SATURATION 9

Fig. 7. Control signals with state feedback control. Fig. 9. Tracking errors with output feedback control.

Fig. 10. Control signals with output feedback control.


Fig. 8. Position tracking with output feedback control.

has a slightly better tracking performance than the output feed-


C. Case 3: Adaptive Neural Impedance Control With back scheme. Nevertheless, some output information may not
Output Feedback (35) be measurable in the state feedback scheme, which means
With all parameters being the same as previously defined the output feedback scheme is more practical than the state
in case 2, the new parameters are added to the simulation: feedback scheme.
 = 0.001, λ1 = [3; 3], and initial terms π1 = 0, π2 = 0,
π̇1 = 0, and π̇2 = 0. V. C ONCLUSION
Figs. 8 and 9 show the position tracking and the tracking In this paper, impedance control based on NNs has been
errors of the robotic manipulator, respectively. Fig. 10 gives the developed to solve the interaction problem for a robotic manip-
control signals. There exists the tracking errors of the output ulator. Both full state and output feedback controllers have
feedback control, but it is much smaller than the nonadaptive been considered with an auxiliary signal to handle the satu-
case. We state that the impedance control based on output ration effect of a robotic manipulator. Stability of the closed
feedback can realize the system objective and track the desired loop system has been proved by Lyapunov’s direct method.
trajectory successfully, where the system error is converging Simulation results have been provided to show that the pro-
to a small value close to zero. posed control is able to track a desired impedance model with
Based on the above simulation cases, it is obvious that a good performance.
the adaptive NN impedance control is effective in handling
the unknown nonlinear dynamic system with saturation effect.
When the end effector is required to move in free space, the ACKNOWLEDGMENT
robotic manipulator has a good tracking performance. When The authors would like to thank the Editor-in-Chief, the
the robotic manipulator is in contact with the obstacle, the Associate Editor, and the anonymous reviewers for their con-
adaptive NN impedance control has a better performance than structive comments which helped to improve the quality and
PD control evidently. Meanwhile, the state feedback control presentation of this paper.
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 SYSTEMS, MAN, AND CYBERNETICS: SYSTEMS

A PPENDIX [19] Z. Li, S. S. Ge, and A. Ming, “Adaptive robust motion/force control
of holonomic-constrained nonholonomic mobile manipulators,” IEEE
Proof for Theorem 1: Multiplying (27) by yields eκt Trans. Syst., Man, Cybern. B, Cybern., vol. 37, no. 3, pp. 607–616,
d  κt  Jun. 2007.
V2 e ≤ Ceκt . [20] Z. Liu, F. Wang, Y. Zhang, X. Chen, and C. L. P. Chen, “Adaptive
dt fuzzy output-feedback controller design for nonlinear systems via back-
stepping and small-gain approach,” IEEE Trans. Cybern., vol. 44, no. 10,
Integrating the above inequality, we obtain pp. 1714–1725, Oct. 2014.
 
C C C [21] Y.-J. Liu, L. Tang, S. Tong, and C. L. P. Chen, “Adaptive NN con-
Vn ≤ V2 (0) − e−κt + ≤ V2 (0) + . troller design for a class of nonlinear MIMO discrete-time systems,”
κ κ κ IEEE Trans. Neural Netw. Learn. Syst., vol. 26, no. 5, pp. 1007–1018,
May 2015.
For z1 , we have [22] Z. Li, S. S. Ge, and S. Liu, “Contact-force distribution optimization
1 T C and control for quadruped robots using both gradient and adaptive
z z1 ≤ V2 (0) + . neural networks,” IEEE Trans. Neural Netw. Learn. Syst., vol. 25, no. 8,
2 1 κ pp. 1460–1473, Aug. 2014.
Then, we can obtain [23] S.-L. Dai, C. Wang, and M. Wang, “Dynamic learning from adaptive
√ neural network control of a class of nonaffine nonlinear systems,” IEEE
||z1 || ≤ B Trans. Neural Netw. Learn. Syst., vol. 25, no. 1, pp. 111–123, Jan. 2014.
[24] C. L. P. Chen, Y.-J. Liu, and G.-X. Wen, “Fuzzy neural network-based
where B = 2(V2∗ (0) + C/κ). adaptive control for a class of uncertain nonlinear stochastic systems,”
IEEE Trans. Cybern., vol. 44, no. 5, pp. 583–593, May 2014.
[25] F. L. Lewis, S. Jagannathan, and A. Yeildirek, Neural Network
R EFERENCES Control of Robot Manipulators and Nonlinear Systems. London, U.K.:
Taylor and Francis, 1999.
[1] S. S. Ge, T. H. Lee, and C. J. Harris, Adaptive Neural Network Control [26] M. Chen, S. S. Ge, and B. Ren, “Robust attitude control of helicopters
of Robotic Manipulators. London, U.K.: World Sci., 1998. with actuator dynamics using neural networks,” IET Control Theory
[2] N. Hogan, “Impedance control: An approach to manipulation. Part I: Appl., vol. 4, no. 12, pp. 2837–2854, Dec. 2010.
Theory; Part II: Implementation; Part III: Applications,” J. Dyn. Syst. [27] M. Gaeta, V. Loia, F. Orciuoli, and P. Ritrovato, “S-WOLF: Semantic
Meas. Control, vol. 107, no. 1, pp. 8–16, 1985. workplace learning framework,” IEEE Trans. Syst., Man, Cybern., Syst.,
[3] C. C. de Wit, G. Bastin, and B. Siciliano, Theory of Robot Control. vol. 45, no. 1, pp. 56–72, Jan. 2015.
London, U.K.: Springer, 1996. [28] Z. Li and C.-Y. Su, “Neural-adaptive control of single-master–multiple-
[4] N. Hogan, “On the stability of manipulators performing contact tasks,” slaves teleoperation for coordinated multiple mobile manipulators with
IEEE J. Robot. Autom., vol. 4, no. 6, pp. 677–686, Dec. 1988. time-varying communication delays and input uncertainties,” IEEE
[5] W.-Z. Gao and R. R. Selmic, “Neural network control of a class of Trans. Neural Netw. Learn. Syst., vol. 24, no. 9, pp. 1400–1413,
nonlinear systems with actuator saturation,” IEEE Trans. Neural Netw., Sep. 2013.
vol. 17, no. 1, pp. 147–156, Jan. 2006.
[29] S.-L. Dai, C. Wang, and F. Luo, “Identification and learning control of
[6] C. Yang et al., “Human-like adaptation of force and impedance in
ocean surface ship using neural networks,” IEEE Trans. Ind. Informat.,
stable and unstable interactions,” IEEE Trans. Robot., vol. 27, no. 5,
vol. 8, no. 4, pp. 801–810, Nov. 2012.
pp. 918–930, Oct. 2011.
[30] Z. Liu, C. Chen, Y. Zhang, and C. L. P. Chen, “Adaptive neural control
[7] E. Burdet, R. Osu, D. W. Franklin, T. E. Milner, and M. Kawato, “The
for dual-arm coordination of humanoid robot with unknown nonlin-
central nervous system stabilizes unstable dynamics by learning optimal
earities in output mechanism,” IEEE Trans. Cybern., vol. 45, no. 3,
impedance,” Nature, vol. 414, no. 6862, pp. 446–449, 2001.
pp. 521–532, Mar. 2015.
[8] H. Kazerooni, P. K. Houpt, and T. B. Sheridan, “Robust compliant
motion for manipulators. Part I: The fundamental concepts of compliant [31] L. Wang et al., “Energy-efficient SVM learning control system for biped
motion; Part II: Design method,” IEEE J. Robot. Autom., vol. 2, no. 2, walking robots,” IEEE Trans. Neural Netw. Learn. Syst., vol. 24, no. 5,
pp. 83–105, Jun. 1986. pp. 831–837, May 2013.
[9] C.-C. Cheah and D.-W. Wang, “Learning impedance control for robotic [32] S.-C. Tong, Y.-M. Li, and H.-G. Zhang, “Adaptive neural network decen-
manipulators,” IEEE Trans. Robot. Autom., vol. 14, no. 3, pp. 452–465, tralized backstepping output-feedback control for nonlinear large-scale
Jun. 1998. systems with time delays,” IEEE Trans. Neural Netw., vol. 22, no. 7,
[10] W.-S. Lu and Q.-H. Meng, “Impedance control with adaptation for pp. 1073–1086, Jul. 2011.
robotic manipulations,” IEEE Trans. Robot. Autom., vol. 7, no. 3, [33] Y.-J. Liu, S.-C. Tong, D. Wang, T.-S. Li, and C. L. P. Chen, “Adaptive
pp. 408–415, Jun. 1991. neural output feedback controller design with reduced-order observer
[11] Y. H. Kim and F. L. Lewis, “Reinforcement adaptive learning neural-net- for a class of uncertain nonlinear SISO systems,” IEEE Trans. Neural
based friction compensation control for high speed and precision,” IEEE Netw., vol. 22, no. 8, pp. 1328–1334, Aug. 2011.
Trans. Control Syst. Technol., vol. 8, no. 1, pp. 118–126, Jan. 2000. [34] Y.-J. Liu, C. L. P. Chen, G.-X. Wen, and S. C. Tong, “Adaptive neural
[12] B. Xu, “Robust adaptive neural control of flexible hypersonic flight output feedback tracking control for a class of uncertain discrete-
vehicle with dead-zone input nonlinearity,” Nonlin. Dyn., vol. 80, no. 3, time nonlinear systems,” IEEE Trans. Neural Netw., vol. 22, no. 7,
pp. 1509–1520, 2015. pp. 1162–1167, Jul. 2011.
[13] B. Xu, Z. Shi, C. Yang, and F. Sun, “Composite neural dynamic surface [35] R. Cui, B. Ren, and S. S. Ge, “Synchronised tracking control of multi-
control of a class of uncertain nonlinear systems in strict-feedback form,” agent system with high order dynamics,” IET Control Theory Appl.,
IEEE Trans. Cybern., vol. 44, no. 12, pp. 2626–2634, Dec. 2014. vol. 6, no. 5, pp. 603–614, Mar. 2012.
[14] L. Liu, D. Liang, and C. Liu, “Nonlinear state-observer control for [36] R. Cui, J. Guo, and Z. Mao, “Adaptive backstepping control of wheeled
projective synchronization of a fractional-order hyperchaotic system,” inverted pendulums models,” Nonlin. Dyn., vol. 79, no. 1, pp. 501–511,
Nonlin. Dyn., vol. 69, no. 4, pp. 1929–1939, 2012. 2015.
[15] W. He, S. S. Ge, Y. Li, E. Chew, and Y. S. Ng, “Neural network control [37] Y.-J. Liu, S. Tong, and C. L. P. Chen, “Adaptive fuzzy control via
of a rehabilitation robot by state and output feedback,” J. Intell. Robot. observer design for uncertain nonlinear systems with unmodeled dynam-
Syst., to be published. ics,” IEEE Trans. Fuzzy Syst., vol. 21, no. 2, pp. 275–288, Apr. 2013.
[16] R. Colbaugh, H. Seraji, and K. Glass, “Direct adaptive impedance con- [38] V. Santibañez, K. Camarillo, J. Moreno-Valenzuela, and R. Campa,
trol of robot manipulators,” J. Robot. Syst., vol. 10, no. 2, pp. 217–248, “A practical PID regulator with bounded torques for robot manipulators,”
1993. Int. J. Control Autom. Syst., vol. 8, no. 3, pp. 544–555, 2010.
[17] C. Yang, Z. Li, and J. Li, “Trajectory planning and optimized adaptive [39] J. Huang, C. Wen, W. Wang, and Z.-P. Jiang, “Adaptive stabilization and
control for a class of wheeled inverted pendulum vehicle models,” IEEE tracking control of a nonholonomic mobile robot with input saturation
Trans. Cybern., vol. 43, no. 1, pp. 24–36, Feb. 2013. and disturbance,” Syst. Control Lett., vol. 62, no. 3, pp. 234–241, 2013.
[18] S. S. Ge, C. C. Hang, L.-C. Woon, and X.-Q. Chen, “Impedance control [40] J.-X. Xu, Y. Tan, and T.-H. Lee, “Iterative learning control design
of robot manipulators using adaptive neural networks,” Int. J. Intell. based on composite energy function with input saturation,” Automatica,
Control Syst., vol. 2, no. 3, pp. 433–452, 1998. vol. 40, no. 8, pp. 1371–1377, 2004.
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 NEURAL IMPEDANCE CONTROL OF ROBOTIC MANIPULATOR WITH INPUT SATURATION 11

[41] M. Chen, B. Jiang, J. Zou, and X. Feng, “Robust adaptive tracking Wei He (S’09–M’12) received the B.Eng. degree
control of the underwater robot with input nonlinearity using neural from the College of Automation Science and
networks,” Int. J. Comput. Intell. Syst., vol. 3, no. 5, pp. 646–655, 2010. Engineering, South China University of Technology,
[42] Y.-J. Liu and S.-C. Tong, “Adaptive NN tracking control of uncertain Guangzhou, China, in 2006, and the Ph.D. degree
nonlinear discrete-time systems with nonaffine dead-zone input,” IEEE from the National University of Singapore (NUS),
Trans. Cybern., vol. 45, no. 3, pp. 497–505, Mar. 2015. Singapore, in 2011.
[43] Z. Liu, G. Lai, Y. Zhang, X. Chen, and C. L. P. Chen, “Adaptive neu- He was a Research Fellow with the Department
ral control for a class of nonlinear time-varying delay systems with of Electrical and Computer Engineering, NUS, from
unknown hysteresis,” IEEE Trans. Neural Netw. Learn. Syst., vol. 25, 2011 to 2012. He is currently with the Center for
no. 12, pp. 2129–2140, Dec. 2014. Robotics and the School of Automation Engineering,
[44] W. He, Y. Chen, and Z. Yin, “Adaptive neural network control of University of Electronic Science and Technology of
an uncertain robot with full-state constraints,” IEEE Trans. Cybern., China, Chengdu, China. His current research interests include robotics, dis-
to be published. tributed parameter systems, and intelligent control systems.
[45] M. Chen, S. S. Ge, and B. Ren, “Adaptive tracking control of uncertain Dr. He has served as an Editor for the Journal of Intelligent and Robotic
MIMO nonlinear systems with input constraints,” Automatica, vol. 47, Systems (Springer).
no. 3, pp. 452–465, 2011.
[46] S.-C. Tong and Y.-M. Li, “Adaptive fuzzy output feedback tracking back-
stepping control of strict-feedback nonlinear systems with unknown dead
zones,” IEEE Trans. Fuzzy Syst., vol. 20, no. 1, pp. 168–180, Feb. 2012.
[47] Z. Guo, J. Wang, and Z. Yan, “Global exponential synchronization of
two memristor-based recurrent neural networks with time delays via
static or dynamic coupling,” IEEE Trans. Syst., Man, Cybern., Syst.,
vol. 45, no. 2, pp. 235–249, Feb. 2015.
[48] C. L. P. Chen, G.-X. Wen, Y.-J. Liu, and F.-Y. Wang, “Adaptive consen-
sus control for a class of nonlinear multiagent time-delay systems using
neural networks,” IEEE Trans. Neural Netw. Learn. Syst., vol. 25, no. 6,
pp. 1217–1226, Jun. 2014.
[49] C. Wen, J. Zhou, Z.-T. Liu, and H.-Y. Su, “Robust adaptive con- Yiting Dong (S’14) received the B.Eng. degree from
trol of uncertain nonlinear systems in the presence of input saturation the School of Automation Engineering, University
and external disturbance,” IEEE Trans. Autom. Control, vol. 56, no. 7, of Electronic Science and Technology of China,
pp. 1672–1678, Jul. 2011. Chengdu, China, in 2014, where he is currently pur-
[50] W. He and S. S. Ge, “Vibration control of a flexible beam with output suing the M.Eng. degree.
constraint,” IEEE Trans. Ind. Electron., to be published. His current research interests include robotics,
[51] W. He, S. Zhang, and S. S. Ge, “Adaptive control of a flexible crane neural network control, and adaptive control.
system with the boundary output constraint,” IEEE Trans. Ind. Electron.,
vol. 61, no. 8, pp. 4126–4133, Aug. 2014.
[52] W. He and S. S. Ge, “Vibration control of a flexible string with
both boundary input and output constraints,” IEEE Trans. Control Syst.
Technol., to be published.
[53] W. He, C. Sun, and S. S. Ge, “Top tension control of a flexible marine
riser by using integral-barrier Lyapunov function,” IEEE/ASME Trans.
Mechatronics, vol. 20, no. 2, pp. 497–505, Apr. 2015.
[54] 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.
[55] S. S. Ge, C. C. Hang, T. H. Lee, and T. Zhang, Stable Adaptive Neural
Network Control. Boston, MA, USA: Kluwer Academic, 2001.
[56] R. Horn and C. Johnson, Matrix Analysis. Cambridge, U.K.: Cambridge
Univ. Press, 1990.
[57] W. He, S. S. Ge, B. V. E. How, Y. S. Choo, and K.-S. Hong,
Changyin Sun received the M.S. and Ph.D. degrees
“Robust adaptive boundary control of a flexible marine riser with vessel
in electrical engineering from Southeast University,
dynamics,” Automatica, vol. 47, no. 4, pp. 722–732, 2011.
Nanjing, China, in 2001 and 2003, respectively.
[58] W. He, S. Zhang, and S. S. Ge, “Robust adaptive control of a
He is a Distinguished Professor with the School of
thruster assisted position mooring system,” Automatica, vol. 50, no. 7,
Automation, Southeast University and the University
pp. 1843–1851, 2014.
of Science and Technology Beijing, Beijing, China.
[59] Y. Li, S. S. Ge, and C. Yang, “Learning impedance control for phys-
His current research interests include intelligent con-
ical robot–environment interaction,” Int. J. Control, vol. 85, no. 2,
trol, flight control, pattern recognition, and optimal
pp. 182–193, 2012.
theory.
[60] S. Behtash, “Robust output tracking for nonlinear system,”
Int. J. Control, vol. 51, no. 6, pp. 931–933, 1990.

You might also like