Robotica: page 1 of 18.

© Cambridge University Press 2016

doi:10.1017/S0263574716000564

Design, analysis and modelling of a hybrid controller

for serial robotic manipulators
Dan Zhang†,∗ and Bin Wei‡
†Department of Mechanical, Engineering, York University, 4700 Keele Street, Toronto, Ontario,
M3J 1P3, Canada
‡Faculty of Engineering and Applied Science, University of Ontario Institute of Technology,
2000 Simcoe Street North, Oshawa, Ontario L1H 7K4, Canada. E-mail: Bin.Wei@uoit.ca
(Accepted July 21, 2016)

SUMMARY
When the end-effector of a robotic arm grasps different payload masses, the output of joint motion
will vary. By using a model reference adaptive control approach, the payload variation effect can
be solved. This paper describes the design for a hybrid controller for serial robotic manipulators by
combining a PID controller and a model reference adaptive controller (MRAC) in order to further
improve the accuracy and joint convergence speed performance. The convergence performance
of the PID controller, the MRAC and the PID+MRAC hybrid controller for 1-DOF, 2-DOF and
subsequently 3-DOF manipulators is compared. The comparison results show that the convergence
speed and its performance for the MRAC and the PID+ MRAC controllers is better than that of
the PID controller, and the convergence performance for the hybrid control is better than that of the
MRAC control.

KEYWORDS: Robotic manipulator, Model reference adaptive control, Hybrid control, Convergence
performance.

1. Introduction
Control of a serial manipulator can be divided into joint control and operational/task space control.
Most robotic industries use a PID controller to control each joint of robotic manipulators.1−5 The
problem of not being able to compensate the payload variations (i.e. when the end-effector of the
robotic arm grasps different payload masses, as most applications require, the output of joint motion
will vary, which decreases the end-effector positioning accuracy of the robotic arm system) results in
using adaptive control, especially model reference adaptive control (MRAC).
The MRAC method was first introduced by Whitaker et al.6 in 1958, when they considered
adaptive aircraft flight control systems, using a reference model to obtain error signals between the
actual and desired behaviour. These error signals were used to modify the controller parameters to
attain ideal behaviour in spite of uncertainties and varying system dynamics. The MRAC was later
developed further.7−17 S. Dubowsky18 was the first to apply the MRAC in the robotic manipulator.
The approach follows the method in ref. [19]. A linear, second-order, time-invariant differential
equation was used as the reference model for each degree of freedom of the manipulator arm. The
manipulator was controlled by adjusting the position and velocity feedback gains to follow the model.
A steepest-descent method was used for updating the feedback gains, after which Horowitz applied the
hyperstability method and developed an adaptive algorithm20 for a serial robotic arm for the purpose
of compensating a non-linear term in dynamic equations and decoupling the dynamic interaction
amongst the joints.
The adaptive method proposed by Horowitz in ref. [20] is different from Dubowsky’s approach.18
Two main differences are summarized as: first, in Horowitz’s method, the overall control system has
an inner loop model reference adaptive system controller and an outer loop position and velocity

* Corresponding author. E-mail: dzhang99@yorku.ca

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 2 Design, analysis and modelling of a hybrid controller for serial robotic manipulators

feedback loop, whereas the control system in Dubowsky’s method is entirely based on the MRAC;
second, in Dubowsky’s paper, the coupling amongst joints and non-linear terms in the manipulator
equations are ignored, whereas this is considered in Horowitz’s method. The drawback of Horowitz’s
method is that the matrices M and N are assumed constant, which means one cannot use the normal
Lagrange method to derive the dynamic equation of robotic manipulators to control the robotic
arms. In ref. [20], Horowitz applied the Gibbs–Appell formulation for dynamic modelling of robotic
manipulators to meet the requirement that matrices M and N of the dynamic equation are constant.
An improved version of the method was later proposed in Sadegh.21 The assumption that the inertia
matrix and non-linear term are constant during adaptation can be removed by modifying the control
law and parameter adaptation law. It is demonstrated that, by modifying the control law (i.e. make
the Coriolis and centripetal acceleration compensation controller a bilinear function of the joint and
model reference velocities, instead of a quadratic function of the joint velocities) and by modifying
the parameter adaptation law (i.e. decompose the non-linear parameters in the manipulator dynamic
equations into the product of two quantities: one constant unknown quantity, which includes the
numerical values of the masses, moments of inertia of the links, payload and link dimensions and
the other a known non-linear function of the manipulator structural dynamics), the assumption that
the inertia matrix and non-linear term are constant during adaptation is removed.
Based on the MRAC control and by combining the PID control, a PID+MRAC hybrid controller
is proposed for serial robotic manipulators. Horowitz’s MRAC controller20 assumes that M and N
are constant during adaptation. For the 1-DOF link, it can directly combine the PID and MRAC
controllers to design the PID+MRAC controller. By using Sadegh’s improved adaptive algorithm
and structure, and by combining the PID and MRAC controllers, a hybrid controller is designed for
the more than 1-DOF link (e.g. 2-DOF and 3-DOF links) case. The convergence performance of the
PID, MRAC and PID+MRAC hybrid controllers for 1-DOF, 2-DOF and 3-DOF manipulators are
compared. The results show that the convergence speed and its performance for the MRAC and PID+
MRAC controllers is better than that of the PID controller. Whereas for the MRAC and PID+ MRAC
controllers, the convergence performance for the hybrid control is better than that of the MRAC
control.
The organization of this paper is as follows: Section 2 presents the PID and MRAC control and
proposes a hybrid controller; dynamic modelling and re-parametrization of the dynamic model are
presented in Section 3; Section 4 simulates and compares the convergence performance between PID,
MRAC and hybrid control for 1-DOF, 2-DOF and subsequently 3-DOF manipulators. The conclusion
is presented in Section 4.

2. PID, MRAC and Hybrid Control

2.1. PID controller

The PID controller is illustrated in Fig. 1. The output of the plant will compare with the desired model
rp and then will have an error. This error will go through the PID control and through “error times
control actions”. The output of the PID controller will be the input to the plant model, and this circle
will continue until the error between the actual output from the plant and the desired model converges
to 0. This is the basic working principle of the PID control.

2.2. MRAC controller

For the MRAC controller, Fig. 2 shows such a system.
One can see that this system does not contain any PID control. The output from the plant will
compare with the reference model, which will produce an error. This error will be used by the adaptive
algorithm block and then produce the input elements to the plant. In the meantime, the output of
the plant will compare with the desired model rp and will produce another error. This error will go
through the integration action and then subtract the feedback processed position and velocity by the
Kp and Kd elements. This process is very similar to the PID control, but it is not a PID control. The
output from this process, times the elements from the adaptive algorithm, plus the elements from the
adaptive algorithm, will be the input to the plant. This process will continue until the error between
the output of the plant and the reference model converges to 0. The ideal system is isolated from the
plant, in the sense that the feedback values of the plant variables are not used to process the input

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 Design, analysis and modelling of a hybrid controller for serial robotic manipulators 3

PID controller

“error x control acons”

Desired Error
I Plant
rp
xp
D

Fig. 1. PID controller.

Kp, Kd
xp

KI/s PARM Plant

rp

[Fp, Fv]

Adapve
[Cp, Cv]
algorithm

Reference
Ideal system
model

Fig. 2. MRAC controller.

to the reference model. The reference model input is processed from its own output variables by
the “similar PID controller”. The ideal system is completely unaffected by the plant performance.
Sadegh’s improved MRAC is illustrated in Fig. 3.

2.3. PID+MRAC hybrid controller

By combining the PID and MRAC controllers, the PID+MRAC hybrid controller is obtained, as
shown in Fig. 4. As with the MRAC, the only difference between this hybrid PID+MRAC and
MRAC is that the output of the plant will compare with the desired model rp and will produce
an error. This error will go through the PID controller. The output of the PID controller, times the
elements from the adaptive algorithm, plus the elements from the adaptive algorithm, will be the input
to the plant20 assumes that M and N are constant during adaptation. For the 1-DOF link, because
the M and N matrices of the dynamic equation are constant (M is constant, N is 0), it can directly
combine the PID and MRAC controllers to design the PID+MRAC controller. However, for more
than 1-DOF link, it is no longer applicable because the M and N matrices of the dynamic equation are
not constant. For the PID control, we need to use the Lagrange dynamic model, but for the MRAC,
we need to use the Gibbs–Appell dynamic formulation. Since they are not compatible, we cannot

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 4 Design, analysis and modelling of a hybrid controller for serial robotic manipulators

Kp, Kd
xp

KI/s Plant
rp

[erp;
erv]
[Fp, Fv]

W Adapve erv
algorithm

Reference
Ideal system
model

Fig. 3. Improved MRAC controller.

PID controller

“error x control acons”

P
xp
rp
Error
I PARM Plant

[Fp, Fv]

Adapve
[Cp, Cv]
algorithm

Reference
Ideal system
model

Fig. 4. PID+MRAC hybrid controller for 1-DOF link.

combine the PID and MRAC in this case. On the positive side, however, Sadegh21−22 proposed an
improved MRAC that can remove the condition that the M and N matrices are constant, so that
the Lagrange dynamic equation can be used. By using Sadegh’s improved adaptive algorithm and
structure, and by combining the PID and MRAC controllers, a hybrid controller is designed for the
more than 1-DOF link (e.g. 2-DOF and 3-DOF links) case. For the 2-DOF and 3-DOF link cases, the
hybrid controller is shown in Fig. 5.

3. Dynamic Modelling and Re-parametrization

3.1. 1-DOF link case

Here, the 1-link manipulator will be used as an example, as shown in Fig. 6. In order to implement
PID control of the 1-link manipulator case, the dynamic equation has to be derived. By using the
Lagrange method,23 the dynamic equation is presented.

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 Design, analysis and modelling of a hybrid controller for serial robotic manipulators 5
PID controller

“error x control acons”

xp P

I Plant
rp

[erp;
D erv]
[Fp, Fv]

W Adapve erv
algorithm

Reference
Ideal system
model

Fig. 5. PID+MRAC hybrid controller for more than 1-DOF link.

l1

θ1

Fig. 6. 1-link manipulator.

The kinetic and potential energy of this link are as follows:

1 •
K1 = m1 (l1 θ1 )2 , P1 = m1 g(l1 sin θ1 ), (1)
2
where m1 is mass of the link 1, l1 is length of the link 1 and θ1 is joint angle of the link-1.
According to the Lagrange method,

1 •
L=K −P = m1 (l1 θ1 )2 − m1 g(l1 sin θ1 ). (2)
2
Thus, the torque applied to the joint can be determined by

d ∂L ∂L
τ1 = • − , (3)
dt ∂ θ ∂θ1
1

• ••
where ∂L
• = m1 l1 2 θ1 , d ∂L
dt •
= m1 l1 2 θ1 , ∂L
∂θ1
= −m1 gl1 cos θ1 .
∂ θ1 ∂ θ1
Rewriting Eq. (3) results in
•• •• ••
τ1 = (m1 l1 2 ) θ1 +(m1 l1 cos θ1 )g = M θ1 +0 + (m1 l1 cos θ1 )g = M θ1 +0 + Gg. (4)

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 6 Design, analysis and modelling of a hybrid controller for serial robotic manipulators

Through applying PID control, the controller output is the torque, i.e.

•
Kp e + Ki edt + Kd e = τ1 , (5)

where error e = rp − xp
We know from the 1-link manipulator M and N matrices, M = m1 l1 2 , N = 0, the output from the
manipulator (i.e. acceleration of the joint) can be determined as follows:
 ••
•
Kp e + Ki edt + Kd e = τ1 = (m1 l1 2 ) θ1 +(m1 l1 cos θ1 )g

••

−1 •
⇒ θ1 =M (Kp e + Ki edt + Kd e). (6)

After deriving the acceleration of joint 1, we need to take the integral with respect to time to obtain
the velocity of joint 1 and take another integral to obtain the position of joint 1.

•
 ••
 •
θ1 = θ1 dt, θ1 = θ1 dt. (7)

For the MRAC, similarly with the PID control, the output from the controller can be determined as
follows:
∧ ∧ •
Controller Out = τ1 = M u + V −Fp e − Fv e, (8)

where u = KI (rp − xp ) − Kp xp − Kd xv .
The manipulator dynamic equation is
••
τ1 = (m1 l1 2 ) θ1 +(m1 l1 cos θ1 )g = Ma + 0 + Gg. (9)

Thus, the output from the manipulator (i.e. acceleration of the joint) is written as

∧ ∧ • •• ∧ ∧ •
M u + V −Fp e − Fv e = τ1 = Ma + V ⇒ a = θ1 = M −1 (M u + V −Fp e − Fv e −V ). (10)

After deriving the acceleration of the joint, we need to take the time integral to obtain the velocity
of the joint and take another integral to obtain the position of the joint.

3.2. 2-DOF link case

For the 2-DOF link case, as shown in Fig. 7 by applying the Lagrange method to derive the dynamic
equation for the PID control analysis:

d ∂L ∂L d ∂L ∂L
τ1 = − , τ2 = − , (11)
dt ∂ θ• ∂θ1 dt ∂ θ• ∂θ2
1 2

1 •
K1 =
m1 (l1 θ1 )2 , P1 = m1 g(l1 sin θ1 ), (12)
2
where m1 is mass of the link-1, l1 is length of the link-1 and θ1 is joint angle of the link-1.
For link-2, we first write down the coordinates of the end of link-2, then differentiate them with
respect to time in order to obtain the kinetic energy. Denote the Cartesian coordinates of the end of
link-2 as (x2 , y2 ).

x2 = l1 cos θ1 + l2 cos(θ1 + θ2 ), y2 = l1 sin θ1 + l2 sin(θ1 + θ2 ). (13)

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 Design, analysis and modelling of a hybrid controller for serial robotic manipulators 7
l2

θ2

l1

θ1

Fig. 7. 2-link manipulator.

We differentiate with respect to time:

• • • • • • • •
x2 = −l1 θ1 sin θ1 − l2 (θ1 + θ2 ) sin(θ1 + θ2 ), y2 = l1 θ1 cos θ1 + l2 (θ1 + θ2 ) cos(θ1 + θ2 ), (14)

1 •2 •2
K2 = m2 v2 2 , and v2 2 = x2 + y2 .
2
Thus,

1 1 •2 1 • • • • •
K2 = m2 v2 2 = m2 l1 2 θ1 + m2 l2 2 (θ1 + θ2 )2 + m2 l1 l2 cos θ2 θ1 (θ1 + θ2 ), (15)
2 2 2

P2 = m2 gl1 sin θ1 + m2 gl2 sin(θ1 + θ2 ), (16)

where m2 is mass of the link-2, l2 is length of the link-2 and θ2 is joint angle of the link-2.
Now,

P = P1 + P2 = (m1 + m2 )gl1 sin θ1 + m2 gl2 sin(θ1 + θ2 ), (17)

1 •2 1 • • • • •
K = K1 + K2 = (m1 + m2 )l1 2 θ1 + m2 l2 2 (θ1 + θ2 )2 + m2 l1 l2 cos θ2 θ1 (θ1 + θ2 ). (18)
2 2
The Lagrange is obtained as

•2 • • • • •
L = K − P = 12 (m1 + m2 )l1 2 θ1 + 12 m2 l2 2 (θ1 + θ2 )2 + m2 l1 l2 cos θ2 θ1 (θ1 + θ2 ) (19)
−(m1 + m2 )gl1 sin θ1 − m2 gl2 sin(θ1 + θ2 ),

d ∂L ∂L
τ1 = • − ,
dt ∂ θ ∂θ1
1

d ∂L ∂L
τ2 = • − . (20)
dt ∂ θ ∂θ2
2

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 8 Design, analysis and modelling of a hybrid controller for serial robotic manipulators

If we put them in a matrix form, we can obtain the following:

     ••   
τ1 •• m11 m12 θ1 n11
= M θ +N = •• + , (21)
τ2 m12 m22 θ2 n21

where

m11 = (m1 + m2 )l1 2 + m2 l2 2 + 2m2 l1 l2 cos θ2 , m12 = m2 l2 2 + m2 l1 l2 cos θ2 , m22 = m2 l2 2 .

• • •2 •2
n11 = 2(−m2 l1 l2 sin θ2 ) θ1 θ2 +(−m2 l1 l2 sin θ2 )θ2 , n21 = m2 l1 l2 sin θ2 θ1 .

Now, applying the PID controller, the controller output is the torque, i.e.
  
• τ1
Kp e + Ki edt + Kd e = , (22)
τ2

where error e = rp − xp . We know the 2-link manipulator M and N matrices, the output from the
manipulator (i.e. acceleration of joints 1 and 2) can be determined as follows:

     ••     
τ1 •• m11 m12 θ1 n11 g
= M θ +N + Gg = •• + + 11 g. (23)
τ2 m12 m22 θ2 n21 g21

Thus,

   •• 

τ1 •• • θ
Kp e + Ki edt + Kd e = = M θ +N + Gg ⇒ ••1
τ2 θ2

•
= M −1 (Kp e + Ki edt + Kd e −N). (24)

After deriving the acceleration of joints 1 and 2, we take the time integral to obtain the velocity of
joints 1 and 2 and take another integral to obtain the position of joints 1 and 2.
For the MRAC approach, we need to re-parameterize the dynamic equation:24
  
(m1 + m2 )l1 2 + m2 l2 2 + 2m2 l1 l2 cos θ2 m2 l2 2 + m2 l1 l2 cos θ2 u1
m2 l2 2 + m2 l1 l2 cos θ2 m2 l2 2⎡ ⎤ u2
⎡ • • •2
⎤ 1 (25)
2(−m2 l1 l2 sin θ2 ) θ1 θ2 +(−m2 l1 l2 sin θ2 )θ2 ⎢ ⎥
+ ⎣ ⎦ = W · ⎣ 2 ⎦ .
•2
m2 l1 l2 sin θ2 θ1 3

By choosing 1 = (m1 + m2 )l1 2 + m2 l2 2 , 2 = m2 l2 2 , 3 = m2 l1 l2 ,

 • • • • 
u1 u2 2u1 cos θ2 + u2 cos θ2 − 2θ 1 θ 2 sin θ2 − θ 2 θ 2 sin θ2
⇒W = • • ,
0 u1 + u2 u1 cos θ2 + θ 1 θ 1 sin θ2

⎤
⎡
1
τ = W ⎣ 2 ⎦ − Fv · erv − Fp · erp, (26)
3

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 Design, analysis and modelling of a hybrid controller for serial robotic manipulators 9
l2

θ3
θ2 l3

l1

θ1

Fig. 8. 3-link manipulator.

⎤ ⎡
••
1
M θ +N = τ = W ⎣ 2 ⎦ − Fv · erv − Fp · erp, (27)
3

⎤ ⎡
••
1
⇒ θ = (W ⎣ 2 ⎦ − Fv · erv − Fp · erp − N)/M. (28)
3

3.3. 3-DOF link case

Based on the Lagrange method, from Fig. 8 the dynamic equation is derived as follows:
⎡ •• ⎤
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
τ1 ••
m11 m12 m13 ⎢ θ1 ⎥ n11
••
⎣ τ2 ⎦ = M θ +N = ⎣ m12 m22 m23 ⎦ ⎢ ⎥ ⎣
⎣ θ2 ⎦ + n21 ,
⎦ (29)
τ3 m13 m23 m33 •• n31
θ3

where

m11 = m1 a 2 + m2 r1 2 + m3 r1 2 , m12 = r1 (m2 b + m3 r2 ) cos(θ2 − θ1 ), m13 = m3 r1 c cos(θ1 + θ3 ),

m22 = m2 b2 + m3 r2 2 , m23 = m3 r2 c cos(θ2 + θ3 ), m33 = m3 c2 ,
•2 •2
n11 = −r1 (m2 b + m3 r2 ) sin(θ2 − θ1 )θ2 − m3 r1 c sin(θ1 + θ3 )θ3 ,
•2 •2
n21 = r1 (m2 b + m3 r2 ) sin(θ2 − θ1 )θ1 − m3 r2 c sin(θ2 + θ3 )θ3 ,
•2 •2
n31 = −m3 r1 c sin(θ1 + θ3 )θ1 − m3 r2 c sin(θ2 + θ3 )θ2 ,
l1 l2 l3
a= , b = , c = , r1 = l1 , r2 = l2 , r3 = l3 ,
2 2 2

m1 , m2 , m3 are the masses of the links 1, 2 and 3, respectively, l1 , l2 , l3 are the lengths of the links 1,
2 and 3, respectively and θ1 , θ2 , θ3 are the joint angles of the links 1, 2 and 3, respectively.
Now applying the PID controller, the controller output is the torque:
⎤⎡
 τ1
•
Kp e + Ki edt + Kd e = ⎣ τ2 ⎦ , (30)
τ3

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 10 Design, analysis and modelling of a hybrid controller for serial robotic manipulators

where error e = rp − xp . Knowing the 2-link manipulator M and N matrices, the output from the
manipulator (i.e. acceleration of joints 1 and 2) can be determined as follows:
⎡ ⎤
⎡ ⎤ ⎡ ⎤ •• ⎡ ⎤ ⎡ ⎤
τ1 ••
m11 m12 m13 ⎢ θ1 ⎥ n11 g11
⎣ τ2 ⎦ = M θ +N + Gg = ⎣ m12 m22 m23 ⎦ ⎢ θ•• ⎥ + ⎣ n21 ⎦ + ⎣ g21 ⎦ g. (31)
⎣ 2⎦
τ3 m13 m23 m33 •• n31 g31
θ3

 τ1 ••
• •• θ1
So Kp e + Ki edt + Kd e = [ τ2 ] = M θ +N + Gg ⇒ [ •• ] = M −1 (Kp e
τ3 θ2

•
+Ki edt + Kd e −N).

For the MRAC approach, by re-parameterizing the dynamic equation, we obtain

⎡ ⎤⎡ ⎤
m1 a 2 + m2 r1 2 + m3 r1 2 r1 (m2 b + m3 r2 ) cos(θ2 − θ1 ) m3 r1 c cos(θ1 + θ3 ) u1
⎢ ⎥⎢ ⎥
⎣ r1 (m2 b + m3 r2 ) cos(θ2 − θ1 ) m2 b2 + m3 r2 2 m3 r2 c cos(θ2 + θ3 ) ⎦ ⎣ u2 ⎦
m3 r1 c cos(θ1 + θ3 ) m3 r2 c cos(θ2 + θ3 )
m3 c 2 u3
⎡ ⎤
1
⎡ ⎤
•2 •2 ⎢ 2 ⎥
⎢ ⎥
⎢ −r1 (m2 b + m3 r2 ) sin(θ2 − θ1 )θ2 − m3 r1 c sin(θ1 + θ3 )θ3 ⎥ ⎢ ⎥
⎢ • • ⎥ ⎢ 3 ⎥
+⎢ ⎥ ⎢ ⎥
2 2
⎢ r1 (m2 b + m3 r2 ) sin(θ2 − θ1 )θ1 − m3 r2 c sin(θ2 + θ3 )θ3 ⎥ = W · ⎢  ⎥ . (32)
⎣ ⎦ ⎢ 4⎥
•2 •2 ⎢ ⎥
−m3 r1 c sin(θ1 + θ3 )θ1 − m3 r2 c sin(θ2 + θ3 )θ2 ⎣ 5 ⎦
6

By choosing1 = m1 a 2 + m2 r1 2 + m3 r1 2 , 2 = r1 (m2 b + m3 r2 ), 3 = m3 r1 c, 4 = m2 b2 +
m3 r2 2 , 5 = m3 r2 c, 6 = m3 c2

⎡ •2 •2
⎤
u1 u2 cos(θ2 − θ1 ) − sin(θ2 − θ1 )θ2 u3 cos(θ1 + θ3 ) − sin(θ1 + θ3 )θ3 0 0 0
⎢ ⎥
⎢ ⎥
⇒W =⎢ ⎥
•2 •2
⎢ 0 u1 cos(θ2 − θ1 ) + sin(θ2 − θ1 )θ1 0 u2 u3 cos(θ2 + θ3 ) − sin(θ2 + θ3 )θ3 0 ⎥,
⎣ ⎦
•2 •2
0 0 u1 cos(θ1 + θ3 ) − sin(θ1 + θ3 )θ1 0 u2 cos(θ2 + θ3 ) − sin(θ2 + θ3 )θ2 u3

⎤ ⎡
1
⎢ 2 ⎥
⎢ ⎥
⎢ ⎥
τ = W ⎢ 3 ⎥ − Fv · erv − Fp · erp, (33)
⎢ 4 ⎥
⎣ ⎦
5
6

⎤ ⎡
1
⎢ 2 ⎥
⎢ ⎥
•• ⎢ ⎥
M θ +N = τ = W ⎢ 3 ⎥ − Fv · erv − Fp · erp, (34)
⎢ 4 ⎥
⎣ ⎦
5
6

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 Design, analysis and modelling of a hybrid controller for serial robotic manipulators 11

Fig. 9. Joint 1 motion under PID.

Fig. 10. Joint 1 motion output under MRAC.

⎤ ⎡
1
⎢ 2 ⎥
⎢ ⎥
•• ⎢ ⎥
⇒ θ = (W ⎢ 3 ⎥ − Fv · erv − Fp · erp − N)/M. (35)
⎢ 4 ⎥
⎣ ⎦
5
6
After deriving the acceleration of the joints, one can take the time integral to obtain the velocity
of the joint and take another integral to obtain the position of the joint. Similarly goes to Eqs. (10)
and (28) for the 1-DOF and 2-DOF cases, respectively.

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 12 Design, analysis and modelling of a hybrid controller for serial robotic manipulators

Fig. 11. Joint output under PID, MRAC and hybrid control when payload is 10 kg.

Fig. 12. Joint output under PID, MRAC and hybrid control when payload is 15 kg.

4. Simulation and Comparison Between PID, MRAC and Hybrid Control

For the 1-DOF manipulator, after applying different masses, the joint motion output is shown in
Fig. 9. When the payload is 0, joint 1 motion is quite steady, but when the payload increases to 5 kg
and 15 kg, one can see that joint 1 motion is no longer the same, and also the joint output increases
and decreases.
By using the MRAC approach, we can see that under different payload masses, three lines coincide
with each other in Fig. 10 under different payload masses. Joint 1 motion is the same, and the payload
mass variation effect has been compensated. Furthermore, as shown in Fig. 11, we can see that, for
the PID control, it will take roughly 40 s to converge to 0. The MRAC control will take about 20 s to

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 Design, analysis and modelling of a hybrid controller for serial robotic manipulators 13

Fig. 13. Joints 1 and 2 output.

Fig. 14. Joints 1 and 2 output.

Fig. 15. Joints 1 and 2 output under PID, MRAC and hybrid control when payload is 1 kg.

converge to the desired position, which is half the time of the PID control. Finally, the hybrid control
takes about 10 s to converge to the desired position, which halves the time of the MRAC control.
Another difference between the MRAC and the hybrid controls is that the MRAC control gradually
converges to the desired position whereas the hybrid control first very quickly overshoots the desired

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 14 Design, analysis and modelling of a hybrid controller for serial robotic manipulators

Fig. 16. Joints 1 and 2 output under PID, MRAC and hybrid control when payload is 5 kg.

Fig. 17. Joints 1, 2 and 3 output.

position and then gradually converges to the desired position. After applying different masses, we
found that hybrid control is better than that of the PID and MRAC for all the mass cases. Here, we
list two of them (10 kg and 15 kg cases) as an illustration, as shown in Figs. 11 and 12. From the
above analysis, we can see that the convergence performance for the hybrid control is better than that
of the MRAC control, and the MRAC control is better than the PID control.

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 Design, analysis and modelling of a hybrid controller for serial robotic manipulators 15

Fig. 18. Joints 1, 2 and 3 output.

For the 2-DOF link case, after applying different payload masses, the joints motion output is
illustrated in Fig. 13. For joint 1, when the payload is 0, the motion is quite steady, but when the
payload increases to 5 kg and 15 kg, we can see that joint 1 motion is no longer the same, and also
the joint output increases and decreases. The same applies to joint 2.
Figure 14 shows the joints output under different payload masses. By using the MRAC approach,
three lines coincide with each other under different payload masses; i.e. the payload mass variation
effect has been compensated. Furthermore, the convergence speed for the hybrid controller is faster
than that of the MRAC controller, as shown in Fig. 15. The hybrid and MRAC controllers are both
better than that of the PID controller. After applying different masses, we found that hybrid control
is better than that of the PID and MRAC for all the mass cases. Here, we list two of them (1 kg and
5 kg cases) as an illustration, as shown in Figs. 15 and 16. By using the same method, the process
can be extended to multi-DOF serial manipulators.
For the 3-DOF case, after applying different payload masses, the joints motion output is illustrated
in Fig. 17. For joint 1, when the payload is 0, the motion is quite steady, but when the payload
increases to 5 kg and 15 kg, we can see that joint 1 motion is no longer the same. The same applies
to joints 2 and 3.
Figure 18 shows the joints output that under different payload masses. By using the MRAC
approach, the payload masses variation effect has been resolved. We can see that three lines coincide
with each other under different payload masses; i.e. the payload mass variation effect has been
compensated. Furthermore, the convergence speed for the hybrid controller is faster than that of the
MRAC controller, as shown in Fig. 19. The hybrid and MRAC controllers are both better than that of
the PID controller. After applying different masses, we found that hybrid control is better than that

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 16 Design, analysis and modelling of a hybrid controller for serial robotic manipulators

Fig. 19. Joints 1, 2 and 3 output under PID, MRAC and hybrid control when payload is 5 kg.

of the PID and MRAC for all the mass cases. Here, we list two of them (5 kg and 10 kg cases) as an
illustration, as shown in Figs. 19 and 20.
To sum up, a hybrid controller is proposed by combining the PID and MRAC controllers, and also
compares the convergence performance of the PID, MRAC and PID+MRAC hybrid controllers for
1-DOF, 2-DOF and 3-DOF manipulators. For the 1-DOF case, the results show that the convergence
speed and its performance for the MRAC and PID+ MRAC controllers are better than that of the PID
controller, whereas for the MRAC and PID+ MRAC controllers, the convergence performance for
the hybrid control is better than that of the MRAC control. As shown in Fig. 11, for the PID control,
it will take roughly 40 s to converge to 0. The MRAC control will take about 20 s to converge to the
desired position, which is half the time of the PID control. Finally, the hybrid control takes about
10 s to converge to the desired position, which halves the time of the MRAC control. Similarly, for
the more than 1-DOF case, the results show that the convergence speed for the hybrid controller is
faster than that of the MRAC controller. The hybrid and MRAC controllers are both better than that
of the PID controller.

5. Conclusion
This paper proposed a hybrid controller by combining the PID and MRAC controllers, and also
compares the convergence performance of the PID, MRAC and PID+MRAC hybrid controllers
for 1-DOF, 2-DOF and 3-DOF manipulators. For the 1-DOF link, it can directly combine the PID
and MRAC controllers to design the PID+MRAC controller. By using Sadegh’s improved adaptive

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 Design, analysis and modelling of a hybrid controller for serial robotic manipulators 17

Fig. 20. Joints 1, 2 and 3 output under PID, MRAC and hybrid control when payload is 10 kg.

algorithm and structure, and by combining the PID and MRAC controllers, a hybrid controller is
designed for the more than 1-DOF link (e.g. 2-DOF and 3-DOF links) case. For the 1-DOF case,
the results show that the convergence speed and its performance for the MRAC and PID+ MRAC
controllers is better than that of the PID controller, whereas for the MRAC and PID+ MRAC
controllers, the convergence performance for the hybrid control is better than that of the
MRAC control. For the PID control, it will take roughly 40 s to converge to 0. The MRAC control
will take about 20 s to converge to the desired position, which is half the time of the PID control.
Finally, the hybrid control takes about 10 s to converge to the desired position, which halves the time
of the MRAC control. For the MRAC, the joint output gradually goes towards to the desired position,
whilst for the PID+MRAC, it overshoots the desired position and then gradually comes back to the
desired position. For the more than 1-DOF case, the results show that the convergence speed for the
hybrid controller is faster than that of the MRAC controller. The hybrid and MRAC controllers are
both better than that of the PID controller. When the end-effector of a robotic arm grasps different
payload masses, the output of joint motion will vary. The joint outputs are measured here. This study
will provide a guideline for future research in the direction of new controller designs for manipulators
in terms of convergence speed and other performances. Future research will focus on learning control
design by simulating human internal control and nervous system for robotic mechanisms.

Acknowledgements
The authors would like to thank the financial support from the Natural Sciences and Engineering
Research Council of Canada (NSERC). The authors gratefully acknowledge the financial support
from Canada Research Chairs program.

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564
 18 Design, analysis and modelling of a hybrid controller for serial robotic manipulators

References
1. R. Kelly, “A tuning procedure for stable PID control of robot manipulators,” Robotica 13(2), 141–148
(1995).
2. J. A. Ramirez and I. Cervantes, “PID regulation of robot manipulators with elastic joints,” Asian J. Control
5(1), 32–38 (2003).
3. A. A. Pervozvanski and L. B. Freidovich, “Robust stabilization of robotic manipulators by PID controllers,”
Dyn. Control 9(3), 203–222 (1999).
4. I. H. Akyuz, E. Yolacan, H. M. Ertunc and Z. Bingul, “PID and State Feedback Control of a Single-
Link Flexible Joint Robot Manipulator,” IEEE International Conference on Mechatronics (ICM), Istanbul,
Turkey (2011) pp. 409–414.
5. E. M. Jafarov, M. N. Parlak and Y. Istefanopulos, “A new variable structure PID-controller design for robot
manipulators,” IEEE Trans. Control Syst. Technol. 13, 122–130 (2005).
6. H. P. Whitaker, J. Yamron and A. Kezer, “Design of Model Reference Adaptive Control Systems for
Aircraft,” (Massachusetts Institute of Technology, Instrumentation Laboratory: Jackson & Moreland,
Cambridge, Massachusetts, 1958).
7. Y. D. Landau, Adaptive Control: The Model Reference Approach (Marcel Dekker, New York, 1979).
8. S. Dubowsky and D. Desforges, “The application of model-referenced adaptive control to robotic
manipulators,” J. Dyn. Syst. Meas. Control 101, 193–200 (1979).
9. C. Cao and N. Hovakimyan, “Design and analysis of a novel L1 adaptive control architecture with guaranteed
transient performance,” IEEE Trans. Autom. Control 53(2), 586–591 (2008).
10. P. Jain and Dr. M. J. Nigam, “Design of a model reference adaptive controller using modified MIT rule for
a second order system,” Adv. Electron. Electric Eng. 3(4) 477–484 (2013).
11. N. T. Nguyen, K. Krishnakumar and J. Boskovic, “An Optimal Control Modification to Model-Reference
Adaptive Control for Fast Adaptation,” AIAA Guidance, Navigation and Control Conference and Exhibit,
Honolulu, Hawaii (Aug. 2008) pp. 1–19.
12. M. Idan, M. D. Johnson and A. J. Calise, “A hierarchical approach to adaptive control for improved flight
safety,” AIAA J. Guid. Control Dyn. 25(6), 1012–1020 (2002).
13. R. Y. Tam and S. C. Thomopoulos, “Model reference adaptive control (MRAC) of robotic manipulators
using a modified output error method (MOEM),” Robot. Syst. 10, 143–150. Microprocessor-Based and
Intelligent Systems Engineering.
14. R. Kamnik, D. Matko and T. Bajd, “Application of model reference adaptive control to industrial robot
impedance control,” J. Intell. Robot. Syst. 22(2), 153–163 (1998).
15. M. C. Tsai and M. Tomizuka, “Model Reference Adaptive Control and Repetitive Control for Robot
Manipulators,” 1989 IEEE International Conference on Robotics and Automation, Scottsdale, USA (1989)
pp. 1650–1655.
16. S. Oh, W. Lee, D. Ha and C. Shin, “A Robust Model Reference Adaptive Control of Robot System Based on
TMS320C3X Chips,” International Conference on Control, Automation and Systems, Seoul, Korea (2007)
pp. 2304–2308.
17. J. M. Skowronski, “Model reference control and identification of robotic manipulators under uncertainty,”
Math. Modelling 8, 384–388 (1987).
18. S. Dubowsky and D. Desforges, “The application of model-referenced adaptive control to robotic
manipulators,” J. Dyn. Syst. Meas. Control 101, 193–200 (1979).
19. D. Donalson and T. Leondes, “A model referenced parameter tracking technique for adaptive control
systems,” IEEE Trans. Appl. Ind. 82(68), 241–252 (1963).
20. R. Horowitz and M. Tomizuka, “An adaptive control scheme for mechanical manipulators-compensation
of nonlinearity and decoupling control,” J. Dyn. Syst. Meas. Control 108(2), 1–9 (1986).
21. N. Sadegh and R. Horowitz, “Stability Analysis of an Adaptive Controller for Robotic Manipulators,”
Proceedings of 1987 IEEE International Conference on Robotics and Automation, Raleigh, NC, USA
(1987) pp. 1223–1229.
22. N. Sadegh and R. Horowitz, “Stability and robustness analysis of a class of adaptive controllers for robotic
manipulators,” Int. J. Robot. Res. 9(3), 74–92 (1990).
23. J. J. Craig, Introduction to Robotics: Mechanics and Control, 3rd ed. (Pearson/Prentice Hall, Upper Saddle
River, NJ, 2005).
24. J. J. Craig, P. Hsu and S. S. Sastry, “Adaptive Control of Mechanical Manipulators,” IEEE International
Conference on Robotics and Automation, San Francisco, CA, USA (1986) pp. 190–195.

Downloaded from http:/www.cambridge.org/core. Ecole Polytechnique, on 20 Dec 2016 at 13:45:39, subject to the Cambridge Core terms of use, available at http:/www.cambridge.org/core/terms.
http://dx.doi.org/10.1017/S0263574716000564