J Control Theory Appl 2010 8 (2) 249–256

DOI 10.1007/s11768-010-8038-x

Adaptive RBF neural network control of

robot with actuator nonlinearities
Jinkun LIU, Yu LU
(School of Automation Science and Electrical Engineering, Beijing University of Aeronautics and Astronautics, Beijing 100191, China)

Abstract: In this paper, an adaptive neural network control scheme for robot manipulators with actuator nonlinearities
is presented. The control scheme consists of an adaptive neural network controller and an actuator nonlinearities compen-
sator. Since the actuator nonlinearities are usually included in the robot driving motor, a compensator using radial basis
function (RBF) network is proposed to estimate the actuator nonlinearities and eliminate their effects. Subsequently, an
adaptive neural network controller that neither requires the evaluation of inverse dynamical model nor the time-consuming
training process is given. In addition, GL matrix and its product operator are introduced to help prove the stability of the
closed control system. Considering the adaptive neural network controller and the RBF network compensator as the whole
control scheme, the closed-loop system is proved to be uniformly ultimately bounded (UUB). The whole scheme provides a
general procedure to control the robot manipulators with actuator nonlinearities. Simulation results verify the effectiveness
of the designed scheme and the theoretical discussion.
Keywords: Adaptive control; RBF neural network; Actuator nonlinearity; Robot manipulator; Deadzone

1 Introduction works with parameter adaptation law are used to emulate

Actuator nonlinearities, including deadzone, backlash, the inertia matrix, centripetal matrix and gravitational vec-
saturation, etc. [1], are quite common in the actual robot tor of robot directly. Nolinear robot function f (x) estimate
manipulators. Because of the non-analytical nature of the is not necessary. Chattering caused by switch gain can be
actuator nonlinearities, their exact nonlinear functions are decreased effectively.
unknown. These nonlinearities often have a great influence 2) The neural network used in [6] is designed by modified
on system performances. The problems are particularly ex- multilayer BP-network, which has too much parameters that
acerbated when the required accuracy is high. For exam- need to be tuned and is not suitable for real-time control. In
ple, the deadzone in the actuator may cause limit cycles this paper, the RBF neural network is adopted to solve the
or poor performance that cannot satisfy the high accuracy- problem.
tracking requirement. Adaptive control for output backlash 3) GL matrix and its product operator are introduced to
and deadzone compensator using fuzzy logic control are help prove the stability of n-link robot manipulator, and
given in [2∼4]. Compensation of hysteresis is given in [5]. the closed-loop system is proved to be uniformly ultimately
The method presented by Lewis et al. [6] is very attrac- bounded (UUB).
tive, because it is a general procedure to compensate the The compensator not only provides compensation signals
unknown actuator nonlinearities. The compensator uses two in the feedforward path, but also takes part in the controller.
neural networks, one to estimate the unknown actuator non- The remaining part of this paper will be organized as fol-
linearities and the other to provide adaptive compensation lows. In Section 2, the dynamics of the robot with actuator
in the feedforward path. However, the controller of the sys- nonlinearity (deadzone) is described. In Section 3, an NN
tem in [6] is not given exactly and it is needed to know the controller with parameter adaptation laws is given. In Sec-
upper bounded error of the nonlinear plant estimation. tion 4, the authors propose an RBF neural network dead-
Unlike many other neural networks, for example, the BP zone compensator and describe the idea underlying the ap-
neural network [7∼9], the RBF neural network has the uni- proach. In Section 5, the closed-loop system is proved to
versal approximation property and can avoid the local min- be uniformly ultimately bounded (UUB). In Section 6, the
ima problem; the network can not only reduce the param- simulations are provided to illustrate the performance of the
eters of neural netwok tuning but also make the initializa- whole tracking system. Finally, we give the conclusion in
tion of neural network much easier [10], so it is very suit- Section 7.
able for realtime control applications. In the past few years,
RBF neural network has been used for the on-line estima- 2 Dynamics of robot system and deadzone
tion of unknown functions in nonlinear adaptive control sys- description
tems, and many researchers have used the neural network 2.1 Dynamics of robot system description
for a wide range of applications in nonlinear adaptive con-
trol [11∼13]. The dynamics of an n-link robot manipulator may be de-
Our main efforts in this paper are highlighted as follows: scribed by the nonlinear system [14]:
1) Instead of BP neural network in [6], three RBF net- M (q)q̈ + Vm (q, q̇)q̇ + G(q) = τ, (1)

Received 13 March 2008; revised 13 November 2008.


c South China University of Technology and Academy of Mathematics and Systems Science, CAS and Springer-Verlag Berlin Heidelberg 2010
250 J. LIU et al. / J Control Theory Appl 2010 8 (2) 249–256

where q(t) ∈ Rn is a vector of the link positions, M (q) difficult. Here we assume that the functions h(u) and g(u)
is the symmetric positive definite inertia matrix of robots, are monotonically increasing, smooth and invertible. There-
Vm (q, q̇) represents coriolis and centrifugal forces, G(q) is fore, there exists a deadzone pre-inverse D−1 (w),
a vector that represents the gravitational forces, and τ is a ⎧ −1
⎨g (w) < 0, u < 0,
vector of the torques (or forces) acted on the joints. −1
D (w) = 0, u = 0, (3)
2.2 Deadzone nonlinearity ⎩ −1
h (w) > 0, u > 0
The deadzone usually appears as one form of the actua-
tor nonlinearities. A mathematical model for the deadzone such that
characteristics is given by: D(D−1 (w)) = w. (4)
⎧ To offset the deleterious effects of the unknown deadzone
⎨g(u) < 0, u  d− ,
D(u) = 0, nonlinearity, a neural network compensator may be placed
d− < d + , (2)
⎩ as illustrated in Fig.1, where w and u are the output of the
h(u) > 0, u  d+ . NN controller and NN compensator, respectively. We hope
Function h(u) and g(u) are smooth, nonlinear functions, that the compensator can cause the composite throughput
so this describes a very general class of deadzone. All of from w to τ to be unity. This is the reason why the deadzone
h(u), g(u), d+ and d− are unknown, so compensation is compensator using RBF networks is proposed in Section 4.

Fig. 1 Adaptive NN control robot system with actuator nonlinearities.

3 Adaptive NN controller design The GL product of {θ}T and {ξ} is an n × n matrix

3.1 GL matrix and operator defined as:
⎡ T T T ⎤
In this paper, we define GL matrix [8], denoted by “{·}”, θ11 ξ11 θ12 ξ12 · · · θ1n ξ1n
and its product operator “◦”. Since the proofs in [6] can only ⎢ θ21
T T
ξ21 θ22 T
ξ21 · · · θ2n ξ2n ⎥
⎢ ⎥
prove the stability of one-link robot manipulator, GL matrix [{θ}T ◦ {ξ}] = ⎢ . . . ⎥.
⎣ .. .. .. ⎦
and its product operator are introduced to prove the stability
T T T
of n-link robot manipulator. θn1 ξn1 θn2 ξn2 · · · θnn ξnn
Let I0 be the set of integers, and θi,j , ξij ∈ Rnij , nij ∈ The GL product of a square matrix and a GL row vector is
I0 , i = 1, · · · , n, j = 1, · · · , n. For function approxima- defined as follows. Let Γk = ΓkT = [rk1 rk2 · · · rkn ], rkj ∈
tion, θij can be taken as the weight vector, and ξij is the n
network basis function vector. The GL row vector {θk } and Rm×nj , m = nj , then we have
its transpose {θk }T are defined in the following way: j=1

{θk } = {θk1 , θk2 , · · · , θkn }, Γk ◦ {ξk } = {Γk } ◦ {ξk }

{θk }T = {θk1T T
, θk2 T
, · · · , θkn }. = {rk1 ξk1 rk2 ξk2 · · · rkn ξkn } ∈ Rm×n .
Note that the GL product should be computed first in a
The GL matrix {θ} and its transpose {θ}T are defined
mixed matrix product. For instance, in {A} ◦ {B}C, the
accordingly as
⎧ ⎫ ⎧ ⎫ matrix [{A} ◦ {B}] should be computed first, and then fol-
⎪ θ11 θ12 · · · θ1n ⎪ ⎪ {θ1 } ⎪ lowed by the multiplication of [{A} ◦ {B}] with matrix C.
⎪
⎪ ⎪
⎪ ⎪ ⎪ ⎪
⎪
⎨ θ21 θ22 · · · θ2n ⎬ ⎨ {θ2 } ⎬ 3.2 Adaptive NN controller design
{θ} = .. .. =
.. ⎪ ⎪ .. ⎪ ,
⎪
⎪ . . . ⎪ ⎪ . ⎪ Given a desired robot arm trajectory qd (t) ∈ Rn , the
⎪
⎩ ⎪
⎭ ⎪ ⎩ ⎪
⎭ tracking error is
θn1 θn2 · · · θnn {θn }
⎧ T T T ⎫
e(t) = qd (t) − q(t). (5)
⎪ θ11 θ12 · · · θ1n ⎪
⎪
⎪ T ⎪
⎪ The filtered tracking error is
⎨ θ21
T T
θ22 · · · θ2n ⎬
{θ}T = . . . . r = ė + Λe, (6)
⎪
⎪ . . .. ⎪
⎪ where Λ = ΛT > 0 is a design parameter matrix, usually
⎪ . .
⎩ ⎪
⎭
T T T
θn1 θn2 · · · θnn selected as diagonal matrix.
We define
For a given GL matrix
⎧ ⎫ ⎧ ⎫ q̇r = r + q̇, (7)
⎪ ξ11 ξ12 · · · ξ1n ⎪ ⎪ {ξ1 } ⎪
⎪
⎪ ⎪ ⎪ ⎪ q̈r = ṙ + q̈, (8)
⎨ ξ21 ξ22 · · · ξ2n ⎪
⎬ ⎪ ⎨ {ξ2 } ⎪
⎬
{ξ} = = . z = (q T q̇ T )T . (9)
⎪ . .
.. .. .
.. ⎪ ⎪ .
.. ⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪ So we have
⎩ ⎭ ⎩ ⎭
ξn1 ξn2 · · · ξnn {ξn } q̇r = q̇d + Λe, (10)
 J. LIU et al. / J Control Theory Appl 2010 8 (2) 249–256 251

q̈r = q̈d + Λė. (11) and Wi are the ideal target weights. σ and σi are both the
−(x−c)2
Based on the Gaussian RBF networks approximation radial basis function as σ(x) = e p2 .
property [8], one can emulate M (q), Vm (q, q̇) and G(q),
respectively: By defining Ŵ and Ŵi as estimates of the ideal NN
weights, two RBF networks are used to emulate D(u) and
M (q) = [{Wm }T ◦ {σM (q)}] + εM (q), (12) wN N :
Vm (q, q̇) = [{Wm }T ◦ {σV (z)}] + εV (q), (13) D̂(u) = {Ŵ }T ◦ {σ(u)}, (26)
G(q) = [{WG }T ◦ {σG (q)}] + εG (q), (14) ŵN N = {Ŵi }T ◦ {σi (w)}. (27)
where {◦} is GL matrix and “◦” is the GL product oper-
We define W̃ = W − Ŵ and W̃i = Wi − Ŵi as the
ator defined in Section 3.1. εv , εv and εG are the NN re-
weight estimation error.
construction errors. WM , WV and WG are the ideal target
weights. σM , σV and σG are all the radial basis function as Theorem 1 Given the RBF network compensator (23)
−(x−c)2 and (27), the throughput of the compensator plus the dead-
σ(x) = e p2 . Then we can get zone is given by

M (q)q̈ + Vm (q, q̇)q̇ + G(q) τ = w − [{Ŵ }T ◦ {σ (u)}] ◦ [{Ŵi }T ◦ {σi (w)}]
= [{WM }T ◦ {σM (q)}]q̈ + [{WV }T ◦ {σV (z)}]q̇ 
+[{W̃ }T ◦ {σ (u)}] ◦ [{W̃i }T ◦ {σi (w)}]
+[{WG }T ◦ {σG (q)}] + εM (q)q̈ + εV (z)q̇ + εG (q) +d(t), (28)
= [{WM }T ◦ {σM (q)}]q̈r + [{WV }T ◦ {σV (z)}]q̇r where the modeling mismatch term is
+[{WG }T ◦ {σG (q)}] − M ṙ − Vm (q, q̇)r + E, (15)

d(t) = −[{W̃ }T ◦ {σ (u)}] ◦ [{Wi }T ◦ {σi (w)}]
where E = εM (q)q̈r + εV (z)q̇r + εG (q). −b(t) + ε(u), (29)

By defining ŴM , ŴV and ŴG as estimates of the ideal b(t) = {W }T ◦ {[σ (w + {Ŵi }T ) ◦ {σi w}]
NN weights, three RBF networks with parameter adapta-
tion law can be used to emulate M (q), Vm (q, q̇) and G(q), ◦[εi (w)]} + {W }T ◦ {R1 (W̃i , w)}
respectively: +ε(w + wN N ). (30)
M̂ (q) = {ŴM } ◦ {σM (q)}, (16) Proof From (23) and (24), one has
τk = WkT σ(uk ) + ε(uk )
V̂m (q, q̇) = {ŴV } ◦ {σV (z)}, (17)
= WkT σ(wk + ŵN N k ) + ε(wk + ŵN N k ), (31)
Ĝ(q) = {ŴG } ◦ {σG (q)}. (18)
where τk is the kth throughput of the compensator plus the
We choose the controller given by Ge et al. in [8] and [9]: deadzone.
w = M̂ (q)q̈r + V̂m (q, q̇)q̇r + Ĝ(q) + KV r + Kr sgn(r) From (4) and (25), it follows that
= {ŴM }T ◦ {σM (q)}q̈r + {ŴV }T wk = WkT σ(wk + wN N k ) + ε(wk + wN N k )
◦{σV (z)}q̇r + {ŴG }T ◦ {σG (q)} = WkT σ(wk + Wik T
σi (wk ) + εi (wk ))
+KV r + Kr sgn(r). (19) +ε(wk + wN N k )
The adaptation laws are = WkT σ(wk + Wik T
σi (wk ) + W̃ikT
σi (wk ) + εi (wk ))
⎧
⎨ ŴM k = ΓM k ◦ σM k (q)q̈r rk ,
⎪ +σ(wk + wN N k ). (32)
˙ Using the Taylor series expansion, one has
ŴV k = ΓV k ◦ σV k (z)q̇r rk , (20)
⎪
⎩ ˙
ŴGk = ΓGk ◦ σGk (q)rk . wk = WkT σ(wk + Ŵik T
σi (wk ))

+WkT σ (wk + Ŵik
T T
σi (wk ))(W̃ik σi (wk )
T
4 NN deadzone compensator design +εi (wk ))+Wk R1 (W̃ik , wk )+ε(wk +wN N k ), (33)
Based on the work of Lewis et al. [6], the deadzone in- where R1 (W̃ik , wk ) is the remainder of the first Taylor
verse can be expressed in equivalent form as polynomial. Regrouping the terms, one has
D−1 (w) = w + wN N , (21) wk = WkT σ(wk + Ŵik T
σi (wk )) + WkT σ


where wN N is the modified deadzone inverse and given by T T

⎧ −1 (wk + Ŵik σi (wk ))W̃ik σi (wk ) + bk (t), (34)
⎨ g (w) − w, w < 0, where
wN N = 0, u = 0, (22) 
⎩ −1 bk (t) = WkT σ (wk + Ŵik T
σi (wk ))εi (wk )
h (w) − w, w > 0. T
+Wk R1 (W̃ik , wk ) + ε(wk + wN N k ).
It is defined that
Combining (23) and (24) gives
u = w + ŵN N (w). (23)
wk + ε(uk )
Based on the RBF network approximation property, one
can approximate the deadzone function D(u) and wN N : = WkT σ(wk + Ŵik T
σi (wk ))


D(u) = {W }T ◦ {σ(u)} + ε(u), (24) +ε(uk ) + ŴkT σ (wk + Ŵik

T T
σi (wk ))W̃ik σi (wk )

wN N = {Wi }T ◦ {σi (w)} + εi (w), (25) +W̃kT σ (wk + Ŵik
T T
σi (wk ))W̃ik σi (wk ) + bk (t)

where ε(u) and εi (w) are the NN reconstruction error. W = WkT σ(uk ) + ε(uk ) + ŴkT σ (uk )W̃ik
T
σi (wk )
252 J. LIU et al. / J Control Theory Appl 2010 8 (2) 249–256

+W̃ T σ (uk )W̃ik

T
σi (wk ) + bk (t) desired by increasing the feedback gain KV > 0.

= WkT σ(uk ) + ε(uk ) + ŴkT σ (uk )W̃ik
T
σi (wk ) Proof Select the Lyapunov function candidate
T T T 
T 1 1 n
+W̃k σ(uk )Wik σi (wk ) − W̃k σ (uk )Ŵik σi (wk ) L = rT M r + W̃ T S −1 W̃k
2 2 k=1 k
+bk (t), (35)
1  n 1 n
which combined with (31) gives + T −1
W̃ik T W̃ik + W̃ T Γ −1 W̃M k
 2 k=1 2 k=1 M k M k
τk = wk − ŴkT σ (uk )W̃ik
T
σi (wk ) 1  n 1 n

+W̃kT σ (uk )Ŵik σi (wk ) + dk (t). + W̃VTk ΓV−1
k W̃V k + W̃ T Γ −1 W̃Gk . (42)
2 k=1 2 k=1 Gk Gk
Using the GL matrix and product operator defined in sec-
Differentiating L and using the robot property Ṁ − 2Vm
tion 3, one can get (28).
being skew-symmetric matrix yields
Theorem 2 The norm of the modeling mismatching
n
˙
term d(t) in (29) is bounded on a compact set by L̇ = rT (M ṙ + V r) + m W̃ T S −1 W̃ k k
d(t)  a1 W̃ F + a2 W̃i 2F k=1

n
T −1 ˙

n
T −1 ˙
+a3 W̃i F + a5 . (36) + W̃ik T W̃ik + W̃M k ΓM k W̃M k
k=1 k=1
The proof is in [15] and [16]. n 
n
+ W̃VTk ΓV−1 ˙ T −1 ˙
Theorem 2 gives us the upper bound of the norm of d(t), k W̃V k + W̃Gk ΓGk W̃Gk . (43)
it is an important result in the stability proof of the next sec- k=1 k=1

tion. The NN tuning algorithm is chosen as Using (40) and (41), one has
⎧ 
⎪ ˙ 
T L̇ = rT [−KV r + [{Ŵ }T ◦ {σ (u)}]
⎨ Ŵk = −Sσ (uk )Ŵik σi (wk )rk − K1 SrŴk , 
˙ 
Ŵik = T σi (wk )rk ŴkT σ (uk ) − K1 T rŴik (37) ◦ [{W̃i }T ◦ {σi (w)}] − [{Ŵ }T ◦ {σ (u)}]
⎪
⎩
−K2 T rŴi F Ŵik ◦ [{W̃i }T ◦ {σi (w)}] − d(t)
with any positive constant matrices S = S T , T = T T and +{W̃V }T ◦ {σV (q)}q̈r + {W̃M }T ◦ {σM (q)}q̇r
K1 , K2 > 0. Ŵk and Ŵik are the weights of the kth com- +{W˜G }T ◦ {σG (q)} + E − Kr sgn(r)]
pensator and 1  k  n. n
˙ +  n
˙
+ W̃ T S −1 W̃
k kW̃ T T −1 W̃ ik ik
k=1 k=1
5 Adaptive NN controller with deadzone n
−1 ˙ 
n
˙
compensator + T
W̃M k ΓM k W̃M k + W̃VTk ΓV−1
k W̃V k
k=1 k=1
Based on the dynamics of robot manipulator equation (1) n
T −1 ˙
and the definitions in equation (5)∼(8) in Section 3, there + W̃Gk ΓGk W̃Gk . (44)
exists k=1

M ṙ = −Vm r − τ + f, (38) By noting that

⎧
where f = M q̈r + Vm q̇r + G. ⎪
⎪ rT [{W̃M }T ◦ {σM (q)}]q̈r
⎪
⎪ 
From the adaptive NN controller (19), it can be written as ⎪
⎪
n
⎪
⎪ = {W̃M k }T ◦ {σM k (q)}q̈r rk ,
w = fˆ + Kv r, ⎪
⎪
(39) ⎨ T k=1
r {W̃V }T ◦ {σV (z)}q̇r
where (45)
⎪ =  {W̃V k }T ◦ {σV k (q)}q̇r rk ,
⎪ n
fˆ = {ŴM }T ◦ {σM (q)}q̈r + {ŴV }T ◦ {σV (z)}q̇r ⎪
⎪
⎪
⎪
+{ŴG }T ◦ {σG (q)} + Kr sgn(r). ⎪
⎪
k=1

⎪
⎪
n
⎩ rT {W̃G }T ◦ {σG (q)} = W̃Gk σGk (q)rk .
Combining (28) and (38) gives k=1

T T
M ṙ = Vm r − KV r + [{Ŵ } ◦ {σ (u)}] ◦ [{W̃i } And using the adaptation laws of controller (20), one has

L̇ = rT [−KV r + [{Ŵ }T ◦ {σ (u)}]

◦{σi (w)}] − [{W̃ }T ◦ {σ (u)}] ◦ [{Ŵi }T

◦{σi (w)}] − d(t) + f − fˆ. (40) ◦ [{W̃i }T ◦ {σi (w)}] − [{Ŵ }T ◦ {σ (u)}]
Noting that ◦ [{W̃i }T ◦ {σi (w)}] − d(t) + E − Kr sgn(r)
n
˙ +  n
f − fˆ = {W̃M }T ◦ {σM (q)}q̈r + {W̃V }T + W̃ T S −1 W̃ k
˙
W̃ T T −1 W̃ ik
k ik
◦ {σV (z)}q̇r + {W̃G }T ◦ {σG (q)} k=1 k=1

+E − Kr sgn(r). (41) = r [−KV r − Kr sgn(r) + E] − rT [{W̃ }T

T
 
Theorem 3 Given the system in (1), assuming that the ◦ {σ (u)}] ◦ [{Ŵi }T ◦ {σi (w)}]+rT [{Ŵ }T ◦ {σ (u)}]
ideal weights W and Wi bounded as W F  WM and n
Wi F  WiM , select the tracking control law (19) plus the ◦ [{W̃i }T ◦ {σi (w)}] + W̃kT S −1 W̃˙ k
k=1
deadzone compensator (23). Let NN weights update laws be 
n T
provided by (20) and (37). Then the tracking error r(t) is + W˜ik T −1 W˜˙ik − rT d(t). (46)
uniformly ultimately bounded, with bounds given by (56). k=1
Moreover, the tracking error r(t) may be kept as small as Applying the compensator NN tuning rules (37), one may
 J. LIU et al. / J Control Theory Appl 2010 8 (2) 249–256 253

write −a2 W̃i F − a3 W̃i F − a5 }, (51)

L̇ = rT [−KV r − Kr sgn(r) + E] 2
L̇  −r{KVmin r + K1 W̃ F − (K1 W + a1 )W̃ F

−rT [{W̃ }T ◦ {σ (u)}] ◦ [{Ŵi }T ◦ {σi (w)}] +K2 W̃i 3F + (K1 − 2K2 WiM − a2 )W̃i 2F

2
+rT [{Ŵ }T ◦ {σ (u)}] ◦ [{W̃i }T ◦ {σi (w)}] −(K1 WiM + K2 WiM + a3 )W̃i F − a5 },

n 
T L̇  −r{KVmin r + K1 [W̃ F
+ W̃k (σk (u)Ŵik σik (w)rk + K1 rŴk )
k=1 1 a1 2 1 a1 2
n − (WM + )] − K1 (WM + )
T  2 K1 4 K1
+ W̃ik (−σik (w)rk ŴkT σk (u) + K1 rŴik
k=1 +g(W̃i F ) − a5 }, (52)
+K2 rŴi F Ŵik ) − rT d(t). (47) where the function g(x) is defined as
By noting that
⎧ T 
g(x) = K2 x3 + (K1 − 2K2 WiM − a2 )x2
⎪
⎪ r [{W̃ }T ◦ {σ (u)}] ◦ [{Ŵi }T ◦ {σi (w)}] −(K1 WiM + K2 WiM2
⎪
⎪ 
+ a3 )x. (53)
⎪
⎪
n 
T
⎨= W̃k σk (u)Ŵik σik (w)rk , Let the constant C be defined as
k=1 (48)

⎪
⎪ rT [{Ŵ }T ◦ {σ (u)}] ◦ [{W̃i }T ◦ {σi (w)}] C = inf{g(x), x  0}. (54)
⎪
⎪ 
⎪
⎪
n 
⎩= W̃ T σik (w)rk Ŵ T σ (u).
ik k k Defining h(x) = g(x) + C, then
k=1
Therefore, L̇  −r{KVmin r + K1 [W̃ F

n 1 a1 2 1 a1 2
L̇ = rT [−KV r − Kr sgn(r) + E] + W̃kT K1 rŴk − (WM + )] − K1 (WM + )
k=1 2 K1 4 K1

n
T +h(W̃i F ) − C − a5 }. (55)
+ W̃ik (K1 rŴik +K2 rŴi F Ŵik )−rT d(t)
k=1
Therefore, L̇ is guaranteed to be negative as long as
= rT [−KV r − Kr sgn(r) + E] + K1 rtr[W̃ T Ŵ ] 1 a1 2
4 K1 (WM + K1 ) + C + a5
+rtr[W̃iT K1 Ŵi + W̃iT K2 Ŵi F Ŵi ] − rT d(t) r  . (56)
KVmin
= rT [−KV r − Kr sgn(r) + E] + K1 rtr[W̃ T
Remark 1 The right-hand side of (56) can be taken as
(W − W̃ ] + rtr[W̃iT K1 (Wi − W̃ )
a practical bound on the tracking error in the sense that r(t)
+W̃iT K2 Ŵi F (Wi − W̃i )] − rT d(t). (49) will never stray far above it. Note that the tracking error
By choosing Kr  E, r(t) may be kept as small as desired by increasing the feed-
back gain KV > 0. Therefore, considering the controller
L̇  −rT KV r + K1 rtr[W̃ T (W − Ŵ )]
and the compensator as an integral control part of the robot,
+rtr[W̃iT K1 (Wi − W̃i ) + W̃iT K2 Ŵi F the closed-loop system is proved to be uniformly ultimately
(Wi − W̃i )] − rT d(t). (50) bounded (UUB).
Using the inequality tr[x̃T (x − x̃)]  x̃F xF − x̃2F Remark 2 Three RBF networks are used to emulate the
and (36), one may write inertia matrix, centripetal matrix and gravitational vector of
the robot directly. Nonlinear robot function f (x) estimate
L̇  −KVmin r2 + K1 rW̃ F (WM − W̃ F ) in [6] is not necessary. Chattering caused by switch gain
+K1 rW̃i F (WiM − W̃i F ) can be decreased effectively.
+K2 rW̃i F Wi − W̃i F (WiM − W̃i F ) Remark 3 In this paper, the closed-loop system of n-
link robot manipulator is proved to be uniformly ultimately
+r(a1 W̃ F + a2 W̃i 2F + a3 W̃i F + a5 ), bounded (UUB) by the help of the GL matrix and operator.
L̇  −KVmin r2 + K1 rW̃ F (WM − W̃ F )
+K1 rW̃i F (WiM −W̃i F )+K2 rW̃i F Wi 6 Simulation tests
−W̃i F WiM − K2 rW̃i 2F Wi − W̃i F For the simulation studies we consider a two-link robot
manipulator as Fig.2, whose dynamics are described by
+r(a1 W̃ F + a2 W̃i 2F + a3 W̃i F + a5 ),
M (q)q̈ + Vm (q, q̇)q̇ + G(q) = τ, (57)
L̇  −KVmin r2 + K1 rW̃ F (WM − W̃ F )
+K1 rW̃i F (WiM − W̃i F ) where
 
2
+K2 rW̃i F WiM + 2K2 rW̃i 2F WiM M (q) =
p1 + p2 + 2p3 cos q2 p2 + p3 cos q2
,
−K2 rW̃i 3F + r(a1 W̃ F + a2 W̃i F p2 + p3 cos q2 p2
 
+a3 W̃i F + a5 ), −p3 q̇2 sin q2 −p3 (q˙1 + q˙2 ) sin q2
Vm (q, q̇) = ,
L̇  −r{KVmin r − K1 W̃ F (WM − W̃ F ) p3 q˙1 sin q2 0
 
2 p cos q1 + p5 cos(q1 + q2 )
−K1 W̃i F (WiM − W̃i F ) − K2 W̃i F WiM G(q) = 4 .
2 3
−2K2 W̃i F WiM + K2 W̃i F − a1 W̃ F p5 cos(q1 + q2 )
254 J. LIU et al. / J Control Theory Appl 2010 8 (2) 249–256

Different from the modified multilayer BP-network pro-

posed in [6], RBF networks have fewer parameters and be-
come much easier to realize.
6.2 PD controller without compensation
Let us first investigate the control performance when we
only use the PD control without deadzone compensation. In
this case, the position errors of the robot two links and the
corresponding control signals τ are shown in Figs.3 and 4.

Fig. 2 Planar two-link manipulator.

We get equation (57) from [8], and the meaning of p1 ∼
p5 are described as follows:
2
p1 = m1 lc1 + m2 l12 + I1 ,
2
p2 = m2 lc2 + I2 ,
p3 = m2 l1 lc2 ,
p4 = m1 lc2 + m2 l1 ,
p5 = m2 lc2 ,
where mi and li are the mass and length of link i, lci is the Fig. 3 Position tracking errors with PD control.
distance from joint (i − 1) to the centre of mass of link i,
and Ii is the moment of inertia of link i.
The true parameters of the robot used for simulation are:
p = [p1 , p2 , p3 , p4 , p5 ]
= [2.9, 0.76, 0.87, 29.8, 8.5]kg·m2 .
Suppose that the robot initially all rests at zero rad. The
deadzone is assumed to have linear functions outside the
deadband. We select
d+ = 12, d− = −10,
h(u) = u − d+ , g(u) = u + d− .
The desired trajectory is chosen as follows: Fig. 4 Control signals τ .
 
2 sin(0.2πt) It can be seen that the simple PD control has large track-
qd (t) = .
cos(0.2πt) ing error and cannot handle the actuator nonlinearities such
The parameters of adaptive NN controller are chosen as as deadzone.
Λ = ΛT = diag[5.0], KV = 20, Kr = 1.0 (58) 6.3 Adaptive controller without compensator
⎧ In this case, the adaptive controller is adopted but no com-
⎨ ΓM k = diag[0.05], pensator is added. The position errors of the robotic two
Γ = diag[0.01], (59) links and the corresponding control signals τ are shown in
⎩ Vk
ΓGk = diag[10]. Figs.5 and 6.
The centers of the RBF networks are uniformly dis- It can be seen that the tracking errors are much smaller
tributed based on the range of the qd , q̇d . The weights of than that of the only PD case, but they cannot satisfy the
the RBF controller are all initialized at zero. high accuracy-tracking requirement, since only the adaptive
controller cannot eliminate the effects of actuator nonlinear-
6.1 The comparison of RBF network and BP-network ities.
The neural network used for compensation in [6] is a
modified multilayer BP-network. There were too many pa-
rameters to be initialized properly without difficulty and no
principles for selecting parameters. However, using RBF
networks only needs several parameters.
From (51) we know that small K1 and K2 can help guar-
antee the stability of the closed-loop system. Therefore, we
choose K1 = K2 = 0.0001. We select S = T = 200.
It is enough to choose 5 hidden-layer nodes for each robot
link. The centers of the RBF networks are chosen as uni-
formly distributed based on the range of controller output.
The weights Ŵ are all initialized at 1.0 and weights Ŵi are Fig. 5 Position tracking errors only using the adaptive controller.
all initialized at zero.
 J. LIU et al. / J Control Theory Appl 2010 8 (2) 249–256 255

Fig. 6 Control signals τ . Fig. 10 Manipulator input u.

6.4 Adaptive controller with compensation It can be seen that the tracking errors are smaller than
those of the above two cases. Since the RBF networks com-
In this case, the adaptive controller is adopted and the
pensator eliminates the effects of the actuator nonlinearity,
deadzone compensator is also added. The position errors of
the deadzone has almost no influence on the control signal
the robot is two links and the corresponding control signals
τ as illustrated in Fig.7. Therefore, the robot manipulator
τ are shown in Figs.7∼10.
with actuator nonlinearities can achieve the high accuracy-
tracking purpose. In addition, an interesting feature can be
seen from the Fig.8, the RBF compensator signals ŴN N
not only provide the compensation in the feedforward path,
but also take part in the function of the adaptive controller.
Therefore, combining the NN compensator with the adap-
tive NN controller as an integral control part of the robot
makes the whole tracking system more robust and accurate.

7 Conclusions
In this paper, an adaptive neural network control scheme
based on RBF neural network for a robot manipulator with
deadzone is proposed. The control scheme consists of a new
Fig. 7 Position tracking errors when using the adaptive controller and the adaptive RBF neural network controller and an improved
compensator. RBF compensator. Unlike the NN control proposed in [6],
three RBF networks with parameter adaptive law are used
to emulate the inertia matrix, centripetal matrix and grav-
itational vector of robot directly, nolinear robot function
f (x) estimate is not necessary, and chattering caused by
switch gain can be decreased effectively. Compared to the
multilayer BP-network in [6], RBF-network used in the im-
proved control system has fewer parameters to be tuned and
is much easier to realize in real time control engineering.
For example, for BP neural network in [6], 20 hidden-layer
nodes with sigmoid function and 4 jump function are cho-
sen; in this paper, we only use 5 hidden layer nodes. GL
matrix and operator are introduced to help prove the sta-
bility of the n-link robot manipulator system. The closed-
Fig. 8 Control signals τ . loop system is proved to be uniformly ultimately bounded
(UUB). The whole scheme can not only be used for dead-
zone compensation, but also for general actuator nonlineari-
ties compensation, like backlash, saturation, etc. Simulation
results verify the high accuracy-tracking performance of the
designed scheme and the theoretical discussion.
In addition, the similar analysis and Matlab programs de-
sign were given in the book [16].

References
[1] B. Friedl. Advanced Control System Design[M]. Englewood Cliffs:
Prentice-Hall, 1996.
[2] G. Tao, P. V. Kokotovic. Adaptive control of systems with unknown
output backlash[J]. IEEE Transactions on Automatic Control, 1995,
Fig. 9 Compensator signals ŴN N . 40(2): 326 – 330.
256 J. LIU et al. / J Control Theory Appl 2010 8 (2) 249–256

[3] G. Tao, P. V. Kokotovic. Continuous-time adaptive control of systems [13] X. Wu, X. Zhu, G. Cao, et al. Predictive control of SOFC based on
with unknown backlash[J]. IEEE Transactions on Automatic Control, a GA-RBF neural network model[J].Journal of Power Sources, 2008,
1995, 40(6): 1083 – 1087. 179(1): 232 – 239.
[4] F. L. Lewis, W. K. Tim, L. Wang, et al. Deadzone compensation in [14] J. E. Slotine, W. Li. Applied Nonlinear Control[M]. Englewood Cliffs:
motion control systems using adaptive fuzzy logic control[J]. IEEE Prentice-Hall, 1991.
Transactions on Control Systems and Technology, 1999, 7(6): 731 –
742. [15] Y. Lu, J. Liu, F. Sun. Actuator nonlinearities compensation using rbf
[5] G. Tao, P. V. Kokotovic. Adaptive Control of Systems with Actuator neural networks in robot control system[C]//IMACS Multiconference
and Sensor Nonlinearities[M]. New York: John Wiley & Sons, 1996. on Computational Engineering in Systems Applications. Beijing:
Tsinghua University Press, 2006: 231 – 238.
[6] R. R. Selmic, F. L. Lewis. Deadzone compensation in motion control
systems using neural networks[J]. IEEE Transactions on Automatic [16] J. Liu. Robot Control System Design and Matlab Simulation[M].
Control, 2000, 45(4): 602 – 612. Beijing: Tsinghua University Press, 2008.
[7] M. Bianchini, P. Frasconi, M. Gori. Learning without local minima
Jinkun LIU was born in Heilongjiang Province,
in radial basis function networks[J]. IEEE Trans. Neural Networks,
1965. He received his Ph.D. degree in Department
1995, 6(3): 749 – 755.
of Automatic Control from Northeastern Univer-
[8] S. S. Ge, T. H. Lee, C. J. Harris. Adaptive Neural Network Control sity, China, in 1997. After a two-year post-doctor in
of Robotic Manipulators[M]. Singapore: World Scientific Publishing Zhejiang University from 1997 to 1999, he joined
Co., 1998. School of Automation Science and Electrical Engi-
[9] S. S. Ge, C. C. Hang, L. C. Woon. Adaptive neural network control of neering, Beijing University of Aeronautics and As-
robot manipulators in task space[J]. IEEE Trans. Ind. Electron, 1997, tronautics, and he is currently a professor. His re-
44(6): 746 – 752. search concerns motion control, intelligent control,
[10] S. S. Ge, C. C. Hang. Direct adaptive neural network control of and robust control. E-mail: [email protected].
robots[J]. International Journal of System Science, 1996, 27(6): 533
Yu LU received the B.E. and M.E. degrees in Elec-
– 542.
trical Engineering from the Beijing University of
[11] Q. Zhu, S. Fei, T. Zhang, et al. Adaptive RBF neural-networks control Aeronautics and Astronautics, China. Since 2008, he
for a class of time-delay nonlinear systems [J]. Neurocomputing, has been pursuing the Ph.D. degree at the National
2008, 71(16/18): 3617 – 3624. University of Singapore. His current research inter-
[12] H. Peng, J. Wu, G. Inoussa, et al. Nonlinear system modeling ests are in the fields of computer networks, human-
and predictive control using the RBF nets-based quasi-linear ARX computer interaction and applications of control the-
model[J]. Control Engineering Practice, 2009, 17(1): 59 – 66. ory.