1

Neural PID Control of Robot Manipulators with

Application to an Upper Limb Exoskeleton
Wen Yu, Senior Member, IEEE, Jacob Rosen

Abstract— In order to minimize steady-state error with respect control was proven in [10], where the robot dynamic was re-
to uncertainties in robot control, PID control needs a big integral written in a decoupled linear system and a bounded nonlinear
gain, or a neural compensator is added to the classical PD control system. In [11], asymptotic stability of linear PID was proven,
with a large derivative gain. Both of them deteriorate transient
performances of the robot control. In this paper, we extend the however conditions for linear PID gains are not explicit.
popular neural PD control into neural PID control. This novel Model-based compensation with PD control is an alter-
control is a natural combination of industrial linear PID control native method for PID control [2], such as adaptive grav-
and neural compensation. The main contributions of this paper ity compensation [12], Lyapunov-based compensation [14],
are semiglobal asymptotic stability of the neural PID control and desired gravity compensation [11], and PD+ with position
local asymptotic stability of the neural PID control with a velocity
observer are proven with standard weights training algorithms. measurement [13]. They all needed structure information of
These conditions give explicit selection methods for the gains of the robot gravity. Some nonlinear PD controllers can also
the .linear PID control. A experimental study on an upper limb achieve asymptotic stability, for example PD control with time-
exoskeleton with this neural PID control is addressed. varying gains [15], PD control with nonlinear gains [16], and
PD control with sliding mode compensation [8]. But these
controllers are complex, many good properties of the linear
I. I NTRODUCTION PID control do not exist.
Proportional-integral-derivative (PID) control is widely used Intelligent compensation for PD control does not need
in industrial robot manipulators [1]. In the absence of robot mathematical model, it is a model-free compensator. It can be
knowledge, a PID controller may be the best controller, classified into fuzzy compensator [17], fuzzy PID [18], neural
because it is model-free, and its parameters can be adjusted compensator [19] and fuzzy-neural compensator [20][21]. The
easily and separately [2]. However, an integrator in a PID basic idea behind these controllers is to use a filtered tracking
controller reduces the bandwidth of the closed-loop system. error in the Lyapunov-based analysis [3]. By proper weight
In order to remove steady-state error caused by uncertainties tuning algorithms, which are similar with robust adaptive
and noise, the integrator gain has to be increased. This leads control methods [22], the derivative of the Lyapunov function
to worse transient performance, even destroys the stability. is negative, as long as the filtered tracking error is outside
Therefore, many robot manipulators use pure proportional- of the ball with radius   here  is the upper bound of
derivative (PD) control or PD control with a small integral all unknown uncertainties,  is the derivative gain in PD
gain [3]. control. These neural PD controllers are uniformly ultimate
It is known that a PD controller can guarantee stability boundedness (UUB), and tracking errors go to smaller with
(bounded) of a robot manipulator in regulation case. However, increasing the gain   The cost of large  is the transient
asymptotic stability is not achieved when the manipulator performance becomes slow. Only when  → ∞ the tracking
dynamics contain gravitational torques vector and friction. error converges to zero [19].
From control viewpoint, this steady-state error can be removed It is well known that the simplest method to decrease
by introducing an integral component to the PD control. It is the tracking error is to add an integral action, i.e., change
PID control. Besides the transient performance and stability the neural PD control into neural PID control. A natural
problems of the integrator, theory analysis is also difficult for question: why do we not add an integrator instead of increasing
industrial linear PID control. In order to ensure asymptotic derivative gain in the neural PD control?
stability of the PID control, a popular method is to modify There are two different approaches to combine PID con-
the linear PID into nonlinear one. For example, the position trol with the intelligent control, such as neural control. The
error was modified into nonlinear form in [4]; The integral first one is neural networks are formed into PID structure
term was saturated by a nonlinear function in [5]; The input [23][24][25]. By proper updating laws, the parameters of PID
was saturated in [6]; An extra integral term in the filtered controllers are changed such that the closed-loop systems are
position was added in [7]; The variable structure control and stable. They are not real industrial PID controllers, because the
neural control were combined with the classic PID control PID gains (weights of the neural networks) are time-varying.
in [8] and [9]. Only a few researchers worked on the linear The second method is intelligent techniques are used to tune
PID. The stability (not asymptotic stability) of the linear PID the parameters of PID controllers, such as fuzzy tuning [26],
neural tuning [27][28], and expert tuning [29]. The controllers
Wen Yu is with the Departamento de Control Automatico, CINVESTAV- are still industrial linear PID, however the stability of closed-
IPN (National Polytechnic Institute), Mexico City, 07360, Mexico. Jacob
Rosen is with Department of Computer Engineering, University of California loop system is not guaranteed. The neural PID control of this
- Santa Cruz, Santa Cruz, California, USA paper overcomes the above disadvantages. It is an industrial
 2

³ ·´
linear PID controller adding a neural compensator. The main approximation theory, the smooth function    can be
obstacle of this neural PID is theoretical difficult in analyzing approximated by a multilayer neural network with one hidden
the stability. Even for linear PID, it is not easy to prove layer in any desired accuracy provided proper weights and
asymptotic stability [11]. Without theoretical guarantee for this hidden neurons
neural PID control, industrial applications cannot be carried ³ ·´ h i h i ³ ´
ˆ   =  c (̂  · )  () =  ∗ ( ∗  · )+  ·
out safely. From the best of our knowledge, theory analysis
for this neural PID control is still not published. (5)
where c ∈ ×  ̂ ∈ ×   is hidden node number, ̂
In this paper, the well known neural PD control of robot
manipulators is extended to the neural PID control. The is the weight in hidden layer. In order to simplify the theory
semiglobal asymptotic stability of this novel neural control is analysis, we first use linear-in-the-parameter net (4), then we
proven. Explicit conditions for choosing PID gains are given. will show that the multilayer neural network (5) can also be
When the measurement of velocities it is not available, local used for the neural control of robot manipulators. The robot
asymptotic stability is also proven with a velocity observer. dynamics (1) have the following standard properties [2] which
Unlike the other neural controllers of robot manipulators, our will be used to prove stability.
neural PID does not need big derivative and integral gains to P1. The inertia matrix  () is symmetric positive definite,
assure asymptotic stability. We apply this new neural control to and
a 7-DOF exoskeleton robot in University of California - Santa 0   { ()} ≤ k k ≤  { ()} ≤    0 (6)
Cruz (UCSC). Experimental results show that this neural PID
control has many advantages over classical PD/PID control, where  { } and  { } are the maximum and minimum
the neural PD control, and the other neural PID control. eigenvalues of the matrix 
P2. For the Centrifugal and Coriolis matrix  ( ̇)  there
II. S EMIGLOBAL ASYMPTOTIC STABILITY OF NEURAL PID exists a number   0 such that
CONTROL k ( ̇) ̇k ≤  k̇k2    0 (7)
Many industrial rigid robots (without exible links and high-
and ̇ () − 2 ( ̇) is skew symmetric, i.e.
frequency joint dynamics) can be expressed in the Lagrangian h i
form  ̇ () − 2 ( ̇)  = 0 (8)
··
³ ·´ ·
 ()  +     +  () +  (̇) =  (1) also

̇ () =  ( ̇) +  ( ̇) (9)
where  ∈  represents the link positions.  () is the ³ ·´
inertia matrix,  ( ̇) = { } represents centrifugal force, P3. The neural approximation error    is Lipschitz
 () is a vector
³ · ´of gravity torques,  (̇) is friction. All terms ·
over  and 
 ()       () and  (̇) are unknown.  ∈  is
control input. The friction  (̇) is represented by the Coulomb k () −  ()k ≤  k − k (10)
friction model
From (4) we know
 (̇) =  1 ̇ +  2 tanh ( 3 ̇) (2) ·
³ ·´
 () +  (̇) =  ∗ ( ) +    (11)
where  3 is a large positive constant, such that tanh ( 3 ̇)
can approximate  (̇)   1 and  2 are positive coeffi- Because  () and  (̇) satisfy Lipschitz condition, P3 is
cients. In this paper we use a simple model for the friction as established.
in [3] and [11], In order to simplify calculation we use the R simple model
 (̇) =  1 ̇ (3) for the friction as in (3), the lower bound of  ()  can be
estimated as
When  () and  (̇) are unknown, we may use a neural Z  ³ ´ Z  Z  Z 
·
network to approximate them as     =  ()  +  (̇)  −  ∗ ()
³ ·´ 0 0 0 0
(12)
   =  () +  (̇) where  ( ) is the potential energy of the robot,  =  () 
³ ·´ ·
³ ·´ ·

ˆ   =  c ( )    =  ∗ ( ) +  () Since (·) is a Gaussian function,  ∗ ()  0 By  ( )  0
(4) Z  ³ ´
· 1√
where  ∗³ is ´unknown constant weight,  c is estimated       1  −  1 0 −  ∗
· 0 2
weight,    is the neural approximation error,  is a R √
neural activation function, here we use Gaussian function such where 0 () = 12  erf (). Since the workspace of a ma-
·
that ( ) ≥ 0. nipulator (the entire set of points reachable by the manipulator)
Since the joint velocity ̇ is not always available, we may is known,R min { } can be estimated. We define the lower

use a velocity observer which will be discussed in Section bound of 0  ()  as
III to approximate it. This linear-in-the-parameter net is the 1√
simplest neural network. According to the universal function  =  1 min { } −  1 0 −  ∗ (13)
2
 3

Given a desired constant position   ∈   the objective of here ̃ c  In matrix form, the closed-loop system
= ∗ − 
robot control is to design the input torque  in (1) such that is
⎡ ⎤
the regulation error ⎡ ⎤  ̃
 ⎢ ⎥
̃ =   −  (14)  ⎢ ̃ ⎥ ⎢ Ã −̇ ´ ! ⎥
³
⎣ · ⎦=⎢ ⎥ ·
·  ⎣  ̇ + ̃ () +    ⎦
̃ → 0 and ̃ → 0 when initial conditions are in arbitrary large ̃ ̈  +  −1
− ̃ +  ̇ − 
domain of attraction. (22)
∙ ¸
The classical industrial PID law is ·
Z  The equilibrium of (22) is  ̃ ̃ = [ ∗  0 0]  Since at
·
 =  ̃ +  ̃ ( )  +  ̃ (15) ·
0
equilibrium
£ ¡ ¢ ¤point  =  and

¡  ¢= 0 ¡the ¢equilibrium is
   0 0  We simplify     0 as    
where    and  are proportional, integral and derivative In order to move the equilibrium to origin, we define
gains of the PID controller, respectively.
° ³ ´° ¡ ¢
° · °
When the unknown dynamic °   ° in (4) is big, in ˜ =  −    (23)
order to assure asymptotic stability, the integral gain  has to The final closed-loop equation becomes
be increased. This may cause big overshoot, bad stability, and ·
³ ·´
integrator windup. Model-free compensation is an alternative  () ̈ +  ( ̇) ̇ + ̃ ( ) +   
¡ ¢
solution, where  () is estimated by a neural network as in =  ̃ −  ̇ + ˜ +    (24)
(4). Normal neural PD control is [3] ·
˜ =  ̃
·
 =  ̃ +  ̃ + ˆ (16) The following theorem gives the stability analysis of the
³ ·´ ·
neural PID control. From this theorem we can see how to
·
c ( )
where ˆ   =  With the filtered error  = ̃+Λ̃ choose the PID gains and how to train the weight of the
(16) becomes neural compensator in (19). Another important conclusion is
the neural PID control (19) can force the error ̃ to zero.
Theorem 1: Consider robot dynamic (1) controlled by the
 =   + ˆ (17)
neural PID control (19), the closed loop system (24) is
The control (17) avoids integrator problems in (15). Unlike semiglobally
∙ asymptotically
¸ stable at the equilibrium  =
¡ ¢ · 
industrial PID control, they cannot reach asymptotic stability.  −    ̃ ̃ = 0 provided that control gains satisfy
The stability condition of the neural PD control (16) is kk 
   is a constant [32]. In order to decrease kk,  has

 ( ) ≥ 32 
to be increased. This causes long settling time problem. The  ( )
 ( ) ≤   (25)
asymptotic stability ( → 0) requires  → ∞  ()

In this paper, an integrator is added into the normal neural  ( ) ≥  +  ( )

q
PD control (16), it has a similar form as the industrial PID in  () ( )
where  =   satisfies (10), and the weight
(15), 3
Z  of the neural networks (4) is tuned by
·
 =  ̃ +  ̃ +   ̃ ( )  + ˆ (18) ·
·
0 ̂ = − ( ) (̇ + ̃) (26)
·
Because in regulation case ̇  = 0 ̃ = −̇ the PID control where  is positive design constant, it satisfies
q
law can be expressed via the following equations 1
3  ( )  ( ) 3
·
c ( ) ≥≥ ¡ ¢ (27)
 =  ̃ −  ̇ +  + 
(19)  ( )  −1  ( )
˙ =  ̃  (0) = 0 Proof: See Appendix.
Remark 2: From above stability analysis, we see that the
We require the PID control part of (19) is decoupled, i.e. gain matrices of the neural PID control (19) can be chosen
   and  are positive definite diagonal matrices. The directly from the conditions (25). The tuning procedure of the
closed-loop system of the robot (1) is PID parameters is more simple than [6][4][11][16][10]. No
³ ·´ modeling information is needed. The upper or lower bounds
 () ̈ +  ( ̇) ̇ + ˜   of PID gains need the maximum eigenvalue of  in (25), it
=  ̃ −  ̇ +  (20) can be estimated without calculating  For a robot with only
˙ =  ̃ revolute joints [2]
µ ¶
where ˜ =  − ˆ  ( ) ≤   ≥  max | | (28)


where  stands the -th element of   ∈ ×  A 

c () = ̃ () +  ()
˜ =  ∗ () +  () −  (21) can be selected such that it is much bigger than all elements.
 4

Remark 3: The main difference between our neural PID conditional integration algorithm. The the integral term is
control with the other neural PD controllers is the stability limited to a selected value:
condition is changed, we require the regulation error ∙ Z  ¸
·
 =  ̃ +  ̃ +   ̃ ( )  max + ˆ (36)
k̃k  1  ( )  ( ) (29) 0
½
The other neural PD controllers need  if kk  max
where  [ max ] =  max is
max if kk ≥ max
2 a prescribed value to the integral term when the controller
k̃k  (30)
 saturates. This approach is also called prelodding [36]. Now
where 1 and 2 are positive constants. Obviously, if the the linear PID controller becomes nonlinear PID. The semi-
initial condition is not worse and satisfies (29), (29) is always global asymptotic stability has been analyzed by [6]. When
satisfied, and k̃k will decrease to zero. But (30) cannot be max is the maximum torque of all joint actuators, max =
satisfied when k̃k becomes small, so  has to be increased.  max (|max
 |)  max
 = max (| |)   ≤ 1 A necessary
Remark 4: If the unknown  () is estimated by the multi- condition is
layer neural network (5). The modeling error (21) becomes max ≥ 3̄ k ()k ≤ ̄
h ·i ³ ·´ h i
˜ =  − ˆ =  ∗ ( ∗   ) +    −  c (̂  · ) where  () is the gravity torque of the robot (1), ̄ is the
h ·i h ·i h ·i ³ ·´ upper bound of  ()   is a design factor in the case of not
= ̃ (̂   ) −  ∗ (̂   ) +  ∗ ( ∗   ) +    all PID terms are subjected to saturation. For the controller
h ·i h ·i ³ ·´
= ̃ (̂   ) +  ∗  0 ̃   + 1 +    (36),  can selected as  = 14 
h h ·i h · ii h i ³ ´ Following the process from (4) to (13), the neural PID with
= ̃ (̂   ) + 0 ̃   +  c 0 ̃  · + 1  ·
anti-windup controller (36) requires
³ ·´ (31)
°max ≥ 3̄ ³ · ´°
where 1 () = 1 +     1 is Taylor approximation ° ∗ · °
error. The closed-loop equation (24) becomes ° ( ) +    ° (37)
° ° ° ³ ´°
n h ·i h · io ° · ° ° · °
≤ ° ∗ ( )° + °   ° ≤ ̄
 () ̈ +  ( ̇) ̇ + ̃ (̂   ) +  0 ̃  
h i ³ ´
+c  0 ̃  · + 1  · where ̄ is the ³ upper
´ bound of the neural estimator,  
∗
· ·
¡ ¢ (32) ( ) and    are defined in (4).
=  ̃ −  ̇ + ˜ +   
· We can see that the first additional condition for the neural
˜ =  ̃ PID with anti-windup is the neural estimator must be bounded.
While the linear neural PID only requires the neural estimation
If the Lyapunov function in (54) is changed as
error satisfy Lipschitz condition (10).
1 ³ ´
Since max (or max ) is a physical requirement for the
 =  +  e  −1 e (33) 
2 actuator, it is not a design parameter. In order to satisfy the
then the derivative of (33) is condition (37), we should force ̄ as small as possible. A
· h ·i good structure
° ° of the neural estimator may make the term
  = ̇ − ̇  ̃ (   ) ° ∗ · °
à ! ° ( )° smaller, such as multilayer neural network in
h h ·i h · ii · (34) Remark 3. There are several methods can be used to find a
+̇ ̃ (̂   ) +  ̃   +  e  e
 0  −1
good neural network, such as the genetic algorithm [37] and
pruning [38]. Besides structure optimization, initial condition
If the training rule (26) is changed as for the gradient training algorithm (26) also affects ̄ Since
· n h i h io the initial conditions for ̂ and ̂ in (35) do not effect the
c = − (̂  · ) +  0 ̃  · (̇ + ̃)
 stability property, we design an off-line method to find a better
· (35) value for ̂ (0) and ̂ (0). If we let ̂ (0) = 0   (0) = 0 
b = − 
c 0  (̇ + ̃) the algorithm (35) can make the identification error conver-
Theorem 1 is also established. gent, i.e., ̂ () and ̂ () will make the identification error
One common problem of the linear PID control (18) is smaller than that of 0 and 0  ̂ (0) and ̂ (0) are selected
integral windup, where the rate of integration is larger than by following steps:
the actual speed of the system. The integrator’s output may 1) Start from any initial value for ̂ (0) = 0  ̂ (0) =
exceed the saturation limit of the actuator. The actuator will 0 
then operate at its limit no matter what the process outputs. 2) Do training with (35) until 0
This means that the system runs with an open loop instead 3) If the k̃ (0 )k  k̃ (0)k  let ̂ (0 ) and ̂ (0 ) as a
0
of a constant feedback loop. The solutions of anti-windup new ̂ (0) and ̂ (0) , i.e., ̂ (0) = ̂ (0 )  ̂ (0) =
schemes are mainly classified into two types [34]: conditional ̂ (0 )  go to 2 to repeat the training process.
integration and back-calculation. It has been shown that none 4) If the k̃ (0 )k ≥ k̃ (0)k, stop this off-line identifica-
of existed methods is able to provide good performance over tion, now ̂ (0 ) and ̂ (0 ) are the final value for
a wide range of processes [35]. In this paper we use the ̂ (0) and ̂ (0).
 5

III. N EURAL PID CONTROL WITH UNMEASURABLE where  is positive design constant, provided that the PID
VELOCITIES control gains of (40) satisfy
1
The neural PID control (19) uses the joint velocities ̇ In  (∙  ) − 2  ( ) ¡ ¢ ¸
contrast to the high precision of the position measurements  ( ) +  −1 
≥ 1
2
¡ −1 ¢
by the optical encoders, the measurement of velocities by + 1+2 
2  + 2  ( ) + 2   
−1
 + 2  (  )+ 2  ( )+() () ()
1 1
tachometers may be quite mediocre in accuracy, specifically  ( ) ≥ 2 ( −1 −)−1
for certain intervals of velocity. The common idea in the design  ( ) ≤ 3  ( )
of PID controllers, which requires velocity measurements, (44)
has been to propose state observers to estimate the velocity. where  satisfies (10),  ( ) is the condition number of 
The simplest observer may be the first-order and zero-relative and the weight of neural networks is tuned by
position filter [2] · £ ¤
̂ = − ( ) ̃ +  +  −1 (̇ + ) (45)
 
 () =  ()   = 1··· (38)
 +  then the closed loop system (41) is locally asymptotically
stable at the equilibrium
where  () is an estimation of ̇   and  are the ele- ∙ ¸
¡ ¢ · 
ments of diagonal matrices  and ,  =  { },  =  =  −    ̃ ̃ = 0 (46)
 { }    0   0 The transfer function (38) can be
realized by in the domain of attraction
½ ∙ ¸
̇ = − ( + )  ( ) 1 1
ė (39) k̃k ≤  ( − ) −  () + kk (47)
 =  +   2 
Proof: See Appendix.
The linear PID control (19) becomes Remark 6: The conditions (43) and (44) decide how to
choose the PID gains. The first condition of (44) is
c ()
 =  ̃ −   +  +  1
˙ =  ̃  (0) = 0  (∙) ≥  ¡ ( ) + Ω¢ ¸
(40)    + 1+2
−1
̇ = − ( + ) Ω= 1 ¡ 2  ¢ + 1
 + 2  ( ) + 12  −1 
 2  ( )
 =  + 
(48)
where    and  are positive definite diagonal matrices, the third conditions of (44) is  ( ) ≥ 3  ( )  they
 and  in (38) are positive constants. are compatible. When  is not big, these conditions can
be established. The second condition of (44) and the third
The closed-loop system of the robot (1) is
condition of (43) are not directly compatible. We first let
⎡ ⎤  as small as possible, and  as big as possible. So 
⎡ ⎤  ̃
 ⎣
 ⎢ − +  ̇ ⎥ can not be big. These requirements are reasonable for our
 ⎦=⎢ ⎣
∙ ¸ ⎥
⎦ real control. If we select  =  +  form the third
 ̇ − ( ̇) ̇ − ̃ () −¡ ()
¢
 −1 condition of¡ (43),  ≥¢ 12  The second condition of (44)
+ ̃ −   + ˜ +   
(41) requires   −1 −h   12  there existsi1   ≥ 12 and a
h i
The equilibrium of (41) is  ˜  ̇ = [0 0 0]  small  such that   ( + )−1 −   12 . After and
The following theorem gives the asymptotic stability of  are decided, we use the second condition of (44) to select
the neural PID control with the velocity observer (38). This  
theorem also provides a training algorithm for neural weights,
and explicit selection method of PID gains. IV. E XPERIMENTAL RESULTS OF THE NEURAL PID
Since the velocities are not available, the input of the neural CONTROL FOR AN EXOSKELETON ROBOT
networks becomes Recently a research group in UCSC has successfully con-
³ ·´ structed a 7-DOF exoskeleton robot, see Figure 1. In this
ˆ   =  c ( ) paper, we apply our neural PID control in this exoskeleton. The
³ ·´ (42)
or ˆ   =  c (̂ [ ] computer control platform of the UCSC 7-DOF exoskeleton
robot is a PC104 with an Intel [email protected] GHz processor
and 512 Mb RAM. The motors for the first four joints are
Theorem 5: Consider robot dynamic (1) controlled by
mounted in the base such that large mass of the motors can
neural PID controller (40), if  and  of the velocity observer
be removed. Torque transmission from the motors to the joints
(38) satisfy
is achieved using a cable system. The other three small motors
 () are mounted in link five.
≤  ( ) 
 ()
2 () () Fortunately, this upper limb exoskeleton is fixed on the
 () ≤ 14  ( ) 2 ( )
(43) human arm, the behavior of the exoskeleton is the same as the
 ()
 ( − ) ≥ 12  () human arm, see Figure 2. It is composed of a 3-DOF shoulder
 6

a7

a4
d1
a3 d5
d3

Fig. 3. The adimittance control of the UCSC exoskeleton

Fig. 1. The UCSC 7-DOF exoskeleton robot.

shoulder

Force Sensor-1

J1-J3 Force Sensor-3

wrist

hand
elbow J5-J7

J4 Force Sensor-2

Fig. 2. Human arm vs. exoskeleton

(J1-J3), a 1-DOF elbow (J4) and a 3-DOFon wrist (J5-J7).

Fig. 4. The end-effector of the exoskkeleton
J1-J3 are responsible for shoulder exion-extension, adduc-
tion and internal–external rotation, J4 create elbow exion-
extension, J5-J7 are responsible for wrist exion–extension,
pronation-supination and radial–ulnar deviation. We regard J1,
Desired joint angles qd
J2 and J3 in Figure 5 as three spherical joints of the human Impedance
Linear PID
shoulder, see Figure 2. Also J5, J6 and J7 in Figure 5 are q~1 algorithm

considered as three spherical joints of the human wrist. + Kd

d
dt +
The exoskeleton’s height was adjusted for each user in a
− Kp exoskeleton
seated position, see Figure 3. The users left hand is an enable q~7 +
button which released the brakes on the device and engaged the Ki 
motor. The objective of the admittance control is to move the Force
Neural compensator
end-effector of the exoskeleton robot from an initial position ~ sensors
W f1
into six slotted holes, see Figure 4. q1
V
The reference signals are generated by admittance control in Human
task space. These references are sent to joint space. The robot q7 arm/hand

in joint space can be regarded as free motion without human q1

constraints. The whole control system is shown in Figure 5. q7 ~
f7 v ≈ q b
The objective of neural PID control is make the transient s +a
performance faster and less overshoot, such that human feel
comfortable. In Figure 2, Force Sensor-1 and Force Sensor-2
are used to detect the human feeling, while Force Sensor-3 is
Fig. 5. The neural PID conrol of the exoskeleton.
used to generate the control command.
The real-time control program operated in Windows XP
 7

with Matlab 7.1, Windows Real-Time Target and C++ . All

of the controllers employed a sampling frequency of 1. Angle (rad) J-1
The properties of the exoskeleton with respect to base frame 0.3
are shown in Table 1.
Table 1. Parameters of the exoskeleton
0.2
Joint Mass (Kg) Center of Mass (m) Link Length (m) Joint Offset (m) Reference Linear PID-1
Linear PID-2
1 3.4 0.3 0.7 0.3
2 1.7 0.05 0.1 0 0.1 Neural PD
Neural PID
3 0.7 0.1 0.2 0.1
4 1.2 .02 0.05 0
5 1.8 .02 0.05 0.1 0
Time (second)
6 0.2 0.04 0.1 0
7 0.5 0.02 0.05 0.1 1.05 1.1 1.15 1.2 1.25 1.3

The two theorems in this paper give sufficient conditions

Fig. 6. Comparison of seveal PID controllers for Joint 1.
for the minimal values of proportional and derivative gains
and maximal values of integral gains. We use the parameters
in Table 1 and (28) to estimate the upper and the lower bounds
of the eigenvalues of the inertia matrix  ()  and  in (10). Angle (rad) J-2
We select  ( )  3  ( )  1  = 10 We choose
 = 4  ( )
 ( ) such that  ( ) ≤ 3  ( ) is satisfied.

0.2
 = 008  is chosen as  = (30)  = 12 7
 so  =
PID tuning with Neural Net
(1758) The joint velocities are estimated by the standard
filters Neural Net in PID form

ė  18
 () =  () =  () (49)
+  + 30 0.1 PID+neural compensator

The PID gains are chosen as

 =  [150 150 100 150 100 100 100]
 =  [2 1 2 2 02 01 01] (50) Time (second)
 =  [330 330 300 320 320 300 300] 0.8 0.9 1 1.1

such that the conditions of Theorem 2 are satisfied. The

initial elements of the weight matrix  ∈ 7×7 are selected Fig. 7. Comparison of three neural PID controllers for Joint 2.
randomly from −1 to 1 The active function in (26) is Gaussian
function
n o
 = exp − ( −  )2 100   = 1 · · · 7 (51) the control result is shown in Figure 6, marked "Neural PD".
The response becomes very slow.
where  is selected randomly from 0 to 2 The weights are
Now we use Joint-2 to compare our neural PID control with
updated by (45) with  = 10
the other two types of neural PID control. The control results
The control results of Joint-1 with neural PID control is of these three neural PID controllers are shown in Figure
shown in Figure 6, marked "Neural PID". We compare our 7. Here we use a three-layer neural network with 3 nodes
neural PID control with the other popular robot controllers. which have integral, proportional and derivative properties.
First, we use the linear PID (15), the PID gains are the same A backprogration-like training algorithm is used to ensure
as (50), the control result is shown in Figure 6, marked "Linear closed-loop stability [23], it is marked "Neural Net in PID
PID-2". Because the steady-state error is so big, the integral form". Then we use a one-hidden layer neural network to tune
gains are increased as the linear PID gains as in [?], it is marked "PID tuning via
 =  [50 20 30 30 10 10 10] (52) neural net". It can be found that the "Neural net in PID form"
can assure stability, but the transient performance is not good.
The control result is shown in Figure 6, marked "Linear The "PID tuning via neural net" is acceptable except its slow
PID-1", the transient performance is poor. There sill exists response.
regulation error. Further increasing  causes the closed-loop Finally, we use the other joints, Joint 3 to Joint 7, to compare
system unstable. Then we use a neural compensator to replace our neural PID with the other popular robot controllers. The
the integrator, it is normal neural PD control (16). In order to results are shown in Figure 8-12.
decrease steady-state error, the derivative gains are increased Clearly, neural PID control can successfully compensate
as the uncertainties such as friction, gravity and the other un-
 =  [970 900 970 970 970 800 800] (53) certainties of the robot. Because the linear PID controller has
 8

0.3
Angle (rad) J-6
Angle (rad) J-3
0.3

0.2 Linear PID-1

Linear PID-2 Neural PID

0.2

Neural PID 0.1

0.1

0
Time (second)
0
Time (second)
1.9 1.95 2 2.05
1 1.05 1.1 1.15

Fig. 11. Linear PID via neural PID for Joint 6.

Fig. 8. Linear PID via neural PID for Joint 3.

Angle (rad) J-7

0.3
Angle (rad) J-4
0.3
Neural PD
0.2
Neural PID
0.2 Neural PID

0.1
Linear PID-2
0.1
Time (second)
0
2.1 2.2 2.3

0
Time (second) Fig. 12. Neural PD via neural PID for Joint 7.
1.7 1.8 1.9 2

Fig. 9. Linear PID-2 via neural PID for Joint 4. no compensator, it has to increase its integral gain to cancel
the uncertainties. The neural PD control does not apply an
integrator, its derivative gain is big.
The structure of neural compensator is very important. The
Angle (rad) J-5
number of hidden nodes  in (5) constitutes a structural prob-
lem for neural systems. It is well known that increasing the
0.3
dimension of the hidden layer can cause the ”overlap” problem
Neural PD and add to the computational burden. The best dimension to
0.2 use is still an open problem for the neural control research
Neural PID community. In this application we did not use hidden layer,
and the control results are satisfied. The learning gain  in
0.1 (45) will in uence the learning speed, so a very large gain can
cause unstable learning, while a very small gain produce slow
learning process.
0
Time (second)
V. C ONCLUSIONS
1.05 1.1 1.15 1.2 1.25 1.3
The neural PID proposed in this paper solves the problems
of large integral and derivative gains in the linear PID control
Fig. 10. Neural PD via neural PID for Joint 5. and the neural PD control. It keeps good properties of the
industrial PID control and neural compensator. Semiglobal
 9

asymptotic stability of this neural PID control is proven. When This means if  is sufficiently large ³ or´ is sufficiently
the joint velocities of robot manipulators are not available, small, (60) is established, and  ̇ ̃ ˜ is globally pos-
local asymptotic stability is assured with filtered positions. R  ³ ·´
  ·
 0    
The stability conditions give explicit methods to select PID itive definite. Using  0     =

  =
gains. We also apply our neural PID to the UCSC 7-DOF ³ ·´ ¡ ¢ £ ¡ ¢¤ · ¡ ¢
exoskeleton robot. Theory analysis and experimental study ̇       
   = 0 and  
̃     = ̃    
show the validity of the neural PID control. the derivative of  is
· · ³ · ´
VI. ACKNOWLEDGMENTS ̇ = ̇   ̈ + 12 ̇   ̇ + ̃  ̃ +    ̇
à !
· ¡ ¢ ·
The authors would like to thanks Levi Makaio Miller for 
+̃   +     f  −1 f
his help to complete the experiments.
· · µ  ¶
· · ·
˜  −1 ˜ ˜ ˜  
+   + ̃  + ̃  −  ̃  ̇ + ̃  ̇ + ̃  ̈
VII. A PPENDIX
·
Proof: [Proof of Theorem 1]We construct a Lyapunov +̃   ̃
function as (62)
R  ³ ·´ ¡ ¢ Using (8), the first three terms of (62) become
 = 12 ̇   ̇ + 12 ̃   ̃ + 0     −  + ̃    
¡ ¢ ¡ ¢ ¡ ¢ ·
+ 32    −1    + 2 ˜ −1 ˜³ −̇   () − ̇   ̇ + ̇  ˜ + ̇     + ̇  ̃ ( ) (63)
´
+̃  ˜ − ̃   ̇ + 2 ̃   ̃ + 12   f  
−1 f
 And
(54) ̇ ≤ − [ ( ) −  ( ) −  k̃k] k̇k2
where  is defined in (13) such that  (0) = 0  is a design (64)
− [ ( ) −  ( ) −  ] k̃k2
positive constant. We first prove  is a Lyapunov function,
 ≥ 0 The term 12 ̃   ̃ is separated into three parts, and If
P4  ( )
 = =1  k̃k ≤ (65)

¡ ¢ ¡ ¢ ¡ ¢
1 = 16 ̃   ̃ + ̃     + 32    −1    and
 ( ) ≥ (1 + )  ( )
2 = 16 ̃   ̃ + ̃  ˜ + 2 ˜ −1 ˜ (66)
 ( ) ≥ 1  ( ) + 
3 = 16 ̃   ̃ − ̃   ̇ + 12 ̇   ̇ ³ ´
R
4 = 0  ()  −  + 2 ̃   ̃ + 12   f   −1 

f ≥0 then ̇ ≤ 0 ¡k̃k decreases.
¢ Then (66) is established. Using
(55) (60) and  −1 =  1( )  (66) is (25).
It is easy to find ̇ is negative semi-definite. Define a ball Σ of radius   0
∙ ¸ ∙ 1 ¸∙ ¸ centered at the origin of the state space, which satisfies these
1 ¡̃ ¢   ¡̃ ¢ condition ½ ¾
1 = 3 (56)
2    3−1    ( )
Σ = ̃ : k̃k ≤ = (67)
Since  ≥ 0 1 is a semi positive definite matrix, 1 ≥ 
0When  ≥   −13  ( )  ̇ is negative semi-definite on the ball Σ There exists a ball
(  )  

Ãr s ! Σ of radius   0 centered at the origin of the state space on

1 1 3 ° ° 2 which ̇ ≤ 0 The origin of the closed-loop equation (24)
° ˜°
2 ≥  ( ) k̃k − ° ° ≥ 0 (57) is a stable equilibrium. Since the closed-loop equation is
2 3  ( )
autonomous, we use La Salle’s theorem. Define Ω as
Because n h i o
Ω =  () = ̃ ̇ ˜ ∈ 3 : ̇ = 0
   ≤ kk kk ≤ kk kk kk ≤ | ()| kk kk n o (68)
(58) = ˜ ∈  : ̃ = 0 ∈   ̇ = 0 ∈ 
√1
3  () ( )
when  ≤  ()  From (62), ̇ = 0 if and only if ̃ = ̇ = 0. For a solution
à r !2  () to belong to Ω for all  ≥ 0, it is necessary and sufficient
1 p 1 that ̃ = ̇ = 0 for all  ≥ 0. Therefore it must also hold that
3 ≥  ( ) k̇k −  ( ) k̃k ≥ 0 (59)
2 3 ̈ = 0 for all  ≥ 0. We conclude that from the closed-loop
system (24), if  () ∈ Ω for all  ≥ 0, then
Obviously, if ³ ·´ ¡ ¢ ¡ ¢
r    =     0 = ˜ +     0
1 ¡ ¢ 32 1
 −1  ( ) 
2
( ) ≥  ( ) (60) · (69)
3 ˜ = 0
there exists h i
q implies that ˜ = 0 for all  ≥ 0 So  () = ̃ ̇ ˜ = 0 ∈
1
3  ( )  ( ) 3 3 is the only initial condition in Ω for which  () ∈ Ω for
≥≥ ¡ ¢ (61)
 ( )  −1  ( ) all  ≥ 0.
 10

Finally, we conclude from all this that the origin of the This means if  is sufficiently large or  is sufficiently
closed-loop system¡ (24)
¢ is locally asymptotically stable. Be- small, (60) is established, and  is globally positive definite.
cause 1 ≤  −1  ( )  the upper bound for k̃k can Now we compute its derivative. The derivative of  is
be
 ( ) ̇ ≤ −̇   ̇ −    −1   − ̃   ̃ +  k̃k2
k̃k ≤  ( )  ( ) (70) +̇   ̇ +  k̃ − k k̇k
2

1 2 1 2
It establishes the semiglobal stability of our controller, in the +̃    −1
¡  ̃ + ̃  −1

 ̃ + ¢   + 2  kk + 2  k̃k
 
sense that the domain of attraction can be arbitrarily enlarged +̃ " Ã−  −    + ̇   !#
·
with a suitable choice of the gains. Namely, increasing  the f −1 f
+    + ( )̇  + ( )̃  + ( ) 
basin of attraction will grow.
Proof: [Proof of Theorem 2]We construct a Lyapunov (78)
function as Because ̇ = − +  ̇ and  =  { }  the last term is
R ¡ ¢
 = 12 ̇   ̇ + 12 ̃   ̃ + 0  ()  −  + ̃     zero if we apply the updating law (45). Using (58), (78) is
¡ ¢ ¡ ¢ ½ ¾
+ 32    −1    + 2 ˜ −1 ˜  ( −  ) −  k̃ − k
¡ ¢ ̇ ≤ −̇ 
1 ̇
−̃   ̇ + ̃  ³ + −1  ˜´+ 12   −1   −   ̇ ¡
− 2()  ( )
¢
µ 1 1 ¶
+  −1 ˜ + 12   f   −1 f
   −1  −  − 2  − 2  ( )
− 
(71) µ ¡ − ()
2  ( ¢ ) ¶
where the definition of  is the same as Theorem 1.  is   −  − −1  − 
−̃  ̃
a design positive constant. We first prove  is a Lyapunov − 12  − 2  ( )
function,  ≥ 0 The term 12 ̃   ̃ is separated into three (79)
P Using  ()  () ≥  () ≥  ()  ()   can be
parts, and  = 6=1 
¡ ¢ ¡ ¢ ¡ ¢ "" or " ", the last condition of (77) can be replaced by
1 = 16 ̃   ̃ + ̃     + 32    −1    ° °
°̃ − 1 °
h i
2 = 16 ̃   ̃ + ̃  ˜ + 2 ˜ −1 ˜
≤ 1   ( − )  ( ) − 2() 1
 ( )  ()
3 = 16 ̃   − ̃¢  ̇ + 14 ̇   ̇
¡  ̃−1 ¡ ¢ (72)
4 = 4  ¡  ¢  +   −1 ˜ + ˜ −1  ˜
1 
It is the attraction area (47).
1  −1  1 
5 = 4     −   ³ ̇ + 4 ̇  ̇´ Using  () +  () ≥  ( + ) ≥  () +  () 
R
6 =  ()  −  +   1 f   −1 f
 ≥0 the second condition of (77) is
0 2 £¡ ¢ ¤ ¡ ¢
Here 1 and 2 are the same as (55), i.e.   −1 −   ≥   −1 −   ( )
(80)
 ≥ 12  + 2
1
 ( ) + ()
2  ( )
 ( ) ≤  ( ) (73)
3 It is the condition for  in (44). Also
√1
 ( ) () ¡ ¢ 1 + 2 
For 3  if  ≤ 6  ()  ( ) ≥  ( )+ −1  +  +  ( )
Ãr !2 2 2
r
1 1 1 It is the condition for  in (44). The condition for  in (44)
3 ≥  ( ) k̇k −  ( ) k̃k ≥ 0 (74) is obtained from (73). The rest part of the proof is the same
2 2 3
¡ ¢ ¡ ¢ as Theorem 1.
Because  () ≤   −1  () and   −1 =
 ()  it is easy to find that,
1
R EFERENCES
¡ −1 ¢ p
if   ≤  ( −1  )  ((−1 )) or  ()
2 () ≤ [1] Y.Jin, Decentralized Adaptive Fuzzy Control of Robot Manipulators,
 ()

IEEE Transactions on Systems, Man, and Cybernetics, Part B, Vol.28,
 ( )  () No.1, 47-57, 1998
⎛ ¡ −1 ¢ ¡ ¢ ° ° ⎞ [2] M.W.Spong and M.Vidyasagar, Robot Dynamics and Control, John
° °
1⎝
1
2    kk2 − 2 −1 kk °˜° Wiley & Sons Inc., Canada, 1989.
4 ≥ ° ° ⎠≥0 [3] F.L.Lewis, D.M.Dawson, C.T.Abdallah, Robot Manipulator Control:
2 ¡¡ ¢¢ ° °2 Theory and Practice, 2nd Edition, Marcel Dekker Inc, New York, NY
+2 −1  °˜° 10016, 2004
p (75) [4] S. Arimoto, Fundamental problems of robot control: Part I, Innovations
If  ( ) ≤ 12  (( −1  ))  ( ) or  () ≤ in the realm of robot servo-loops, Robotica, vol.13, No.1, 19–27, 1995.
1  () [5] D.Sun, S.Hu, X.Shao, and C.Liu, Global Stability of a Saturated
4  ( ) 2 ( ) Nonlinear PID Controller for Robot Manipulators, IEEE Transactions
∙ ¸ on Cntrol Systems Technology, Vol.17, No.4, 892-899, 2009
1 1  −1  1  [6] J.Alvarez-Ramirez, R.Kelly, I.Cervantes, Semiglobal stability of satu-
5 =     + 2  ̇ + ̇  ̇ ≥ 0 (76) rated linear PID control for robot manipulators, Automatica, vol.39,989-
2 2 2 995, 2003
Because 6 ≥ 0 obviously, there exist   and  such that [7] R.Ortega, A.Loria, R.Kelly, A Semiglobally Stable Output Feedback
PI2 D Regulator for Robot Manipulators, IEEE Transactions on Auto-
1  ( ) () matic Control, Vol.40, No.0, 1432-1436, 1985
2 ≤ 6 2 () [8] V.Parra-Vega, S.Arimoto, Y.-H.Liu,G.Hirzinger, P.Akella, Dynamic Slid-
 ()  ()
2 () ≤  ( )  () (77) ing PID Control for Tracking of Robot Manipulators: Theory and
Experiments, IEEE Transactions on Robotics and Automation, Vol.19,
 () ≤ 14  ( ) 2 ( )
No.6, 967-976, 2003
 ()
 11

[9] T.Dierks, S.Jagannathan, Neural Network Output Feedback Control of [35] Karl J. Astrom , T. Hagglund, PID Controllers: Theory, Design and
Robot Formations, IEEE Transactions on Systems, Man, and Cybernet- Tuning, ISA Press, Research Triangle Park, North Carolina, 1995
ics, Part B, Vol.40, No.4, 383-399, 2010 [36] F. Greg Shinskey, Process-control Systems: Application, Design, Adjust-
[10] P.Rocco, Stability of PID control for industrial robot arms, IEEE Trans- ment, McGraw-Hill, New York. USA. 1996
actions on Robotics and Automation, VOL.12, NO. 4 , 606-614,1996. [37] J.Arifovica , R.Gencay, Using genetic algorithms to select architecture
[11] R.Kelly, V.Santibáñez, L.Perez, Control of Robot Manipulators in Joint of a feedforward arti cial neural network, Physica A, Vol.289, 574-594,
Space, Springer-Verlag London, 2005. 2001
[12] P.Tomei, Adaptive PD Controller for Robot Manipulators, IEEE Trans- [38] K.Suzuki, I.Horina and N.Sugie, A Simple Neural Network Pruning
actions on Robotics and Automation, VoL. 7, No. 4 , 565-570,1991. Algorithm with Application to Filter Synthesis, Neural Processing
[13] B.Paden, R.Panja, Globally asymptotically stable PD+ controller for Letters, Vol.13, 43-53, 2001.
robot manipulators, International Journal of Control, Vol. 47, No. 6,
1697–1712 , 1988
[14] K.Dupree, C-H.Liang, G.Hu, W. E.Dixon, Adaptive Lyapunov-Based
Control of a Robot and Mass–Spring System Undergoing, IEEE Trans-
actions on Systems, Man, and Cybernetics, Part B, Vol.38, No.4, 1050-
1061, 2008
[15] Z.Qu, D.M. Dawson, S. Y. Lim, J.F. Dorsey, A New Class of Robust
Control Laws for Tracking of Robots, International Journal of Robotics
Research, Vol. 13, No.4, 355–363, 1994
[16] E.V. L. Nunes, L.Hsu, F.Lizarralde, Arbitrarily small damping allows
global output feedback tracking of a class of Euler-Lagrange systems,
2008 American Control Conference,, Seattle, USA, 378-382, 2008
[17] H-X. Li, L.Zhang, K-Y.Cai, G.Chen, An Improved Robust Fuzzy- Wen Yu (M’97–SM’04) received the
PID Controller With Optimal Fuzzy Reasoning, IEEE Transactions on
Systems, Man, and Cybernetics, Part B, Vol.35, No.6, 1283-1294, 2005 B.S. degree from Tsinghua University, Beijing, China in
[18] E.Harinath,G.Mann, Design and Tuning of Standard Additive Model 1990 and the M.S. and Ph.D. degrees, both in Electrical
Based Fuzzy PID Controllers for Multivariable Process Systems, IEEE Engineering, from Northeastern University, Shenyang, China,
Transactions on Systems, Man, and Cybernetics, Part B, Vol.38, No.8,
667-674, 2008 in 1992 and 1995, respectively. From 1995 to 1996, he served
[19] F.L.Lewis, Nonlinear Network Structures for Feedback Control, Asian as a Lecturer in the Department of Automatic Control at
Journal of Control, vol. 1, no. 4, pp.205–228,1999. Northeastern University, Shenyang, China. In 1996, he joined
[20] C-S.Chen, Dynamic Structure Neural-Fuzzy Networks for Robust Adap-
tive Control of Robot Manipulators, IEEE Trans. Industrial Electronics,
CINVESTAV-IPN, Mexico, where he is currently a professor
VOL. 55, NO. 9, 3402-3414, 2008 in the Departamento de Control Automatico. He held research
[21] M.J.Er and Y.Gao, Robust Adaptive Control of Robot Manipulators positions with the Instituto Mexicano del Petroleo from 2002
Using Generalized Fuzzy Neural Networks, IEEE Trans. Industrial
Electronics, VOL. 50, NO. 3, 620-628, 2003
to 2003. He was a senior visiting research fellow at Queen’s
[22] P.A.Ioannou and J.Sun, Robust Adaptive Control, Prentice-Hall, Inc, University Belfast from 2006 to 2007, a visiting associate
Upper Saddle River: NJ, 1996. professor at University of California Santa Cruz from 2009 to
[23] S. Cong and Y. Liang, PID-Like Neural Network Nonlinear Adaptive 2010. He also holds a visiting professorship at Northeastern
Control for Uncertain Multivariable Motion Control Systems, IEEE
Trans. Ind. Electron., vol. 56, no. 10, pp. 3872-3879, 2009. University in China from 2006 until now. Dr.Wen Yu serves
[24] G.M. Scott, J. W.Shavlik, W. H. Ray, Refining PID Controllers Using as an associate editor of Neurocomputing, and Journal of
Neural Networks, Neural Computation, Vol. 4, No. 5, 746-757, 1992 Intelligent and Fuzzy Systems. He is a member of Mexican
[25] H. J. Uang and C. C. Lien, Mixed H2/H∞ PID tracking control design
for uncertain spacecraft systems using a cerebellar model articulation Academy of Science.
controller, IEE Proceedings - Control Theory and Applications,, vol.
153, no. 1, pp. 1–13, 2006. Jacob Rosen ****.
[26] G. K. I. Mann, B-G. Hu, R.G. Gosine, Two-Level Tuning of Fuzzy PID
Controllers, IEEE Transactions on Systems, Man, and Cybernetics, Part
B, Vol.31, No.2, 263-269, 2001
[27] S-J. Ho, L-S.Shu, S-Y.Ho, Optimizing Fuzzy Neural Networks for
Tuning PID Controllers Using an Orthogonal Simulated Annealing
Algorithm OSA, IEEE Trans. Fuzzy Syst., Vol.14, NO. 3, 421-434,
2006
[28] D-Li Yu, T. K. Chang, D-W.Yu, Fault Tolerant Control of Multivariable
Processes Using Auto-Tuning PID Controller, IEEE Transactions on
Systems, Man, and Cybernetics, Part B, Vol.35, No.1, 32-43, 2005
[29] F. Karray, W. Gueaieb, S. Al-Sharhan, The Hierarchical Expert Tuning
of PID Controllers Using Tools of Soft Computing, IEEE Transactions
on Systems, Man, and Cybernetics, Part B, Vol.35, No.6, 1283-1294,
2005
[30] G.D’Emilia, A.Marra, E.Natale, Use of neural networks for quick and
accurate auto-tuning of PID controller, Journal Robotics and Computer-
Integrated Manufacturing, Volume 23 Issue 2, 170-179, 2007
[31] L.B.Gutierrez and F.L.Lewis, Implementation of a neural net tracking
controller for a single exible link: comparisonwith PD and PID
controllers, IEEE Trans. Industrial Electronics, VOL. 45, NO. 2, 307-
318, 1998.No.1, 131-147, 2002.
[32] F.L.Lewis, K.Liu, and A.Yesildirek, Neural net robot controller with
guaranteed tracking performance, IEEE Trans. Neural Networks, vol. 6,
no. 3, pp. 703–715, 1995.
[33] J. J.Slotine, W.Li, Adaptive manipulator control: A case study, IEEE
Transactions on Automatic Control, Vol.33, No.11, 995–1003, 1988
[34] A. Visioli, Modified anti-windup scheme for PID controllers, IEE Proc.-
Control Theory Appl., Vol.161, No 1., 49-54, 2003