Escolar Documentos
Profissional Documentos
Cultura Documentos
61
Dynamic Model-Based Adaptive Posture Controller for
Robotic Wheelchairs
Celso De La Cruz Wanderley Cardoso Celeste Teodiano Freire Bastos
*
Department of Electrical Engineering, Universidade Federal do Esprito Santo, Vitoria 29075-910, Brazil
Received 27 Mar 2010; Accepted 24 Oct 2010; doi: 10.5405/jmbe.761
Abstract
A novel adaptive posture control for robotic wheelchairs was designed using feedback linearization techniques and
Lyapunov theory. It is novel in that it uses a dynamic model of the mobile robot in the posture-control law. A switching
control that changes the parameter-updating law is used to improve the adaptive control to avoid the effects of
parameter drift. Experimental results show good performance of this adaptive posture control.
Keywords: Robotic wheelchair, Dynamic model, Adaptive control, Posture control
1. Introduction
A robotic wheelchair is useful for handicapped people
who are not able to use conventional wheelchairs. One of the
features of robotic wheelchairs is that their users can select a
desired target using a human-machine interface based on brain
signals, head movements, eye blinks or eye movements; then
the robotic wheelchair goes automatically to the desired target.
To accomplish this, a trajectory tracking control and/or a
posture control are necessary.
The research of [1] provides an integrated solution to
motion planning and control of robotic wheelchairs with human
inputs from three sources: at the highest level, the human
operator selects the destination; at the intermediate level, the
user interacts with the controller to avoid obstacles; and at the
lowest level, the human operator directly provides velocity
commands using a joystick.
In other research [2], driving by breath-expulsion,
voice-command guidance, and guidance by head movements
and electro-oculographic signals (EOG) have been tested using
a robotic wheelchair. Other contributions of [2] are in the
sensory system and the control system. The sensory system
uses ultrasonic and infrared sensors as well as a laser emitter
and CCD camera detector to obtain information on the
surroundings that allows the wheelchair to be located. The
control system uses an optimal-adaptive control law along with
optimal-fuzzy trajectory tracking.
A third study [3] describes a prototype of a robotic
wheelchair with manual, semiautonomous, and autonomous
* Corresponding author: Teodiano Freire Bastos
Tel: +55-27-40092077; Fax: +55-27-40092644
E-mail: tfbastos@ele.ufes.br
modes. The choice of mode usually depends on parameters
such as single-switch or proportional human-machine interface
sensors, modeled or non-modeled environment, etc.
Because robotic wheelchairs carry a heavy load, it is
important to consider the dynamic model in the control-law
design. In the literature, control-law design can be found for
mobile robots based on the approach of feedback linearization
[4,5]. The control law designed using that approach has good
performance, but it depends on the dynamic model parameters.
When the model parameters are unknown, it is necessary to
have an adaptive control to update such parameters. Some
authors designed the adaptive control law based on
backstepping techniques such as [6]. Others used a
feedback-linearization approach to design the control law based
on linear systems theories, then designed the adaptive law
using Lyapunov theories, such as [4]. In [7] the kinematics is
separated from the dynamics. Thus, a kinematic control law is
used at a high level to control the position of the mobile robot.
In addition to that, a dynamic control law is used at a low level
to control the velocity of the mobile robot.
This work develops a novel adaptive posture-control law
for robotic wheelchairs. The posture controller is novel in that
it uses the dynamic model in the control law. The posture
controllers that can be found in literature are generally based on
the kinematic model [8,9]. The feedback-linearization
techniques and Lyapunov theories are used in the adaptive
control laws design. The controller design is made for a
defined class of model output functions. Then, according to the
control objective (posture control or only heading control), the
output function is defined. The switching scheme of [10] is
used in the adaptive control law. This scheme switches between
an adaptive control and a non-adaptive control. The
non-adaptive control keeps the model parameters constant, and
it is turned on when the control error is small. Thus it avoids
J. Med. Biol. Eng., Vol. 32 No. 1 2012
62
the parameter drift problem, which is caused by measurement
errors, noise or disturbances. Several experimental results are
presented and analyzed in order to show this control systems
performance.
2. Adaptive controller design
In this section, an adaptive posture-control law and an
adaptive heading-control law are developed for a robotic
wheelchair (see Fig. 1), taking into account constant but
unknown model parameters. In Fig. 2, the block diagram of the
control system is shown. The posture controller cannot be used
when the position error is very small, because it uses polar
coordinates of the vehicle. Therefore the heading controller is
used when the position error is small.
Figure 1. Schematic of the robotic wheelchair.
Figure 2. Block diagram of the control system.
First, the dynamic model of the mobile robot is presented.
Second, the adaptive control law is designed for a general
output function. Finally, an output function related to the
posture control and other output function related to the heading
control are proposed.
2.1 Dynamic model
The robotic wheelchair schematic is shown in Fig. 1,
where the point (x; y) is the interest point that defines the
vehicle position, G is the center of mass, B is the wheel
baseline center, u and are the linear and angular velocities,
is the heading of the vehicle, d, b and a are distances. Let us
consider the following dynamic model of the robotic
wheelchair [5]:
(
(
(
(
(
(
(
(
+
(
(
(
(
(
(
(
(
=
(
(
(
(
(
(
0
2
0
1
0
2
0
6
0
2
0
5
0
1
0
4 2
0
1
0
3
0
0
0
cos sin
sin cos
u
e
u
e
u
u
e
u
u
u
u
e
u
u
e
e
e
e
ref
ref
u
u
u
a u
a u
u
y
x
(1)
where
0
i
u is the i-th model parameter and u
ref
and
ref
are the
linear and angular reference velocities. Generally these
reference velocities are control signals in most commercially
available autonomous vehicles. Thus they are used as control
signals in the robotic wheelchair to maintain compatibility with
other autonomous vehicles. The equation (1) can be split in a
kinematic part (first three lines of the equation) and a dynamic
part (last two lines of the equation).
The equations of the model parameters are:
( ) ( ) 2 2 2
2 0
1
|
|
.
|
\
|
+ + =
PT DT e
a
a
rk rk I mr
k
R
u
( ) ( ) ( ) 2 2 2
2 2 2 0
2
|
|
.
|
\
|
+ + + =
PR DR z e
a
a
rdk rdk mb I r d I
k
R
u
( ) 2
0
3
=
PT
a
a
k mbr
k
R
u
( ) 1
0
4
+
|
|
.
|
\
|
+ =
PT e
a
b a
a
a
rk B
R
k k
k
R
u
(2)
( )
0
5
=
PR
a
a
dk mbr
k
R
u
( ) 1 2
0
6
+
|
|
.
|
\
|
+ =
PR e
a
b a
a
a
rk d B
R
k k
k
R
u
where m is the mass of the wheelchair including the user; I
z
is
the moment of inertia about the vertical axis located in G of the
wheelchair including the user; r is the right and left wheel
radius; k
b
is equal to the voltage constant multiplied by the gear
ratio; R
a
is the electric resistance constant; k
a
is the torque
constant multiplied by the gear ratio; I
e
and B
e
are the moment
of inertia and the viscous friction coefficient of the combined
motor rotor, gearbox, and wheel; k
PT
and k
DT
are the
proportional and derivative gains of the PD low-level linear
velocity control of the robotic wheelchair; and k
PR
and k
DR
are
the proportional and derivative gains of the PD low-level
angular velocity control of the robotic wheelchair.
As is usual in adaptive controller design [11], initially the
disturbance vector of the dynamic model is not considered, but
it is considered in a second stage. Thus, a robustness study of
the system in presence of measurement errors, noise and/or
disturbances is made in the next section.
The linear parameterization of dynamic part of (1) is [5]:
(
=
(
ref
ref
u
u
u u
e
u
e e e
e
0
2
0 0 0
0 0 0
(3)
Rearranging, it follows that
ref
C M v v = + (4)
Dynamic Adaptive Controller for Robotic Wheelchair
63
where
, ,
0
0
0
2
0
1
(
=
(
=
e
v
u
u u
M (5)
,
(
=
ref
ref
ref
u
e
v
(6)
(
+
+
=
e u e u
u e u
0
6
0
5
0
4
2 0
3
u
u
C
(7)
2.2 Closed loop equation
The output function of the model (1) is a general function
with the following structure:
( ) , , y x h h =
(8)
where h is a column vector of two elements.
Assumption 1: The temporal derivative of h has the expression
( )v , , y x N h =
(9)
where exists N
-1
and N
, it results that
N N h
+ = (10)
Rearranging the vector C the following expression can be
obtained:
( )
0
, T C e u
a
= (11)
where
(
=
e e
e
u
u
T
a
0 0 0 0
0 0 0 0
2
(12)
Obtaining an expression for from (10), substituting in (4)
and using (11), the inverse dynamic is obtained:
( )
ref a
T N h MN v u v = +
0 1
(13)
Let us consider the following linearization control signal:
( ) u v v
1
a ref
T N v N M + =
(14)
where u
~
1 1
= + +
u v v
a
T h N M N h N M
(15)
where u u u
~
0
= and
M M M
~
=
. Rearranging (15), the
following equation can be obtained:
( ) 0
~
1
= +
u v T h N M
(16)
where
(
=
e e
e
u T
u T
T
0 0 0
0 0 0
22
2
11 (17)
( ) v N h N
T
T
=
(
1
22
11
(18)
Let us define the variable v as follows:
h K h K h v
~ ~
2 1
+ + =
d
,
h h h =
d
~
(19)
where h
d
defines the desired output function, and K
1
and K
2
are
positive definite diagonal matrices. Substituting (19) in (16)
and rearranging, the closed loop equation is obtained:
u
~
Y e A e
T k T
+ = (20)
where
(
=
h
h
e
T ~
~
,
(
=
1 2
0
K K
I
A
K
(21)
(
=
T M N
Y
1
0
(22)
and I is the identity matrix. From the definition of K
1
and K
2
,
matrix A
K
is Hurwitz.
2.3 Stability analysis
Let us consider that h
d
,
d
h
and
d
h
are bounded, and that
k
DT
>0, k
PT
>0, k
DR
>0 and k
PR
>0. Let us also consider the
following assumptions:
Assumption 2: N and N
-1
have bounded elements if h is
bounded.
Assumption 3: N
(23)
then the adaptive control law (14), (19) and (23) guarantees that
~
h 0 and
~
h
0 as t .
Proof. Let us define the following Lyapunov candidate
function:
( ) u u u
~ ~ ~
,
A
T
T
T
T T
K V + = Pe e e
(24)
where 0 > =
T
A A
K K and P = P
T
> 0 such that
Q PA P A
k
T
k
= +
,
0 > =
T
Q Q (25)
Let us take into account that A
K
is Hurwitz. Thus, if matrix
P is calculated from (25), matrix P results such that P = P
T
> 0.
By substituting (20) and (25) in the temporal derivative of V,
the following equation is obtained:
u u u
~ ~
2
~
2
A
T
T
T T
T
T
T
K Y Q V + + = Pe e e (26)
By considering that the model parameters are time-invariant
and considering the adaptive law (23), the following equation is
obtained:
T
T
A
Y K Pe
1
~
= = u u
(27)
By substituting (27) in (26), it results that
J. Med. Biol. Eng., Vol. 32 No. 1 2012
64
0 s =
T
T
T
Q V e e
(28)
Therefore V,
T
e and u
~
are bounded. Integrating (28), the
following inequalities are obtained:
( )
( ) ( )
( )
<
s
s =
}
} } }
Q
V V
dt
dt Q dt Q dt V
T
T T
T
T
min
0
2
0
2
min
0 0
0
e
e e e
(29)
with
( ) Q
min
is
bounded, because u
~
and
0
u are bounded. From (9), the
following equation can be obtained:
h N
1
= v
(30)
Then, by considering assumption 2 as true, | |
T
u v e = is
bounded. From (14), by considering assumption 2 and 3 as true,
it can be concluded that
ref
= [u
ref
ref
]
T
is bounded. Therefore,
by observing (10) e (1), it can be concluded that h
is bounded.
Finally,
T
e
is bounded.
Let us consider the following function:
2
T T
T
T
e e e f = =
(31)
The integral ( )
}
t
d f
0
t t is an increasing function and bounded,
because
n
T
L e
2
e . Then the integral ( )
}
t
d f
0
t t has a limit as
t . The time derivative of (31) is
T
T
T
e e f
2 =
(32)
From (32), one can observe that f
is bounded. Therefore,
applying the Barbalat lemma f 0 as t . This implies,
from (31), that e
T
0 as t .
2.4 Projection algorithm
There is a problem to take into account with the parameter
updating law (23). From (22), it can be observed that 0
1
= u
and 0
2
= u must be avoided to compute Y. To solve this
problem, a projection algorithm [11] can be applied. The
projection algorithm used in the present work is:
> s
< >
s s
> >
< <
=
0 and
if or
0 and
if 0
0 and
if or
0 and
if or
if
i i ui
i i li
i i ui
i i li
ui i li i
i
l
l
l
l
l l
u
u
u
u
u
u
(33)
where l
ui
and l
li
are the i-th upper and lower limit, respectively,
of the i-th estimated parameter i
u
, and i
=
=
=
2 0
1
1
1
on
on T
T
A
C if
C if Pe Y K
(35)
The variable Con switches from 1 to 2 when V
eT
goes from
V
eT
> C
V
to V
eT
C
V
, where V
eT
is function of e
T
, and C
V
is a
positive constant. The variable C
on
switches from 2 to 1 when
V
eT
goes from V
eT
< C
V
to V
eT
C
V
. Let the function V
eT
be
T
T
T
T
V Pe e
e
=
(36)
The Lyapunov function and its temporal derivative,
corresponding to the adaptive law, are:
0
,
~ ~
where
1
~ ~
1
s =
= + =
T
T
T
A
T
V
V V V V
T
Qe e
K
e
u u
(37)
In [12], the same case is analyzed, and it is concluded that V
eT
is finally bounded. This implies that T
e
is also finally
bounded.
3. Definitions of the output function h
3.1 Definition of h for the posture control
Polar coordinates are used to define the vehicles states.
The variables e
c
,
u
and
c
in Fig. 3 are the polar coordinates of
the vehicle posture calculated from the states x, y and using
a = 0.
The control objective is to attain e
c
= 0, C so the vehicle
will attain the origin of the reference coordinate <g> and the
heading of the x
g
-axis of this coordinate. The function h must
be defined such that the convergence of h to h
d
implies the
convergence of the polar coordinates to zero. The desired
output function equal to zero (h
d
= 0) is used for simplification.
Let h be as follows:
Dynamic Adaptive Controller for Robotic Wheelchair
65
(
+
+
=
u c
u c
e
h
| |
|
2
2
1
(38)
where
1
and
2
are real constants greater than zero. This h
function complies with the previously presented requisite:
when h converges to zero, the polar coordinates e
c
,
u
and
c
converge to zero. Now it is necessary to prove that this function
allows the positive verification of the assumptions made in the
adaptive controller design; this is performed in the following
paragraphs.
Figure 3. Unicycle-like mobile robot in polar coordinates; <g> is a
fixed reference coordinate.
The temporal derivatives of these coordinates, considering
the reference coordinate <g> as fixed, are
c
u
c
c
u
u
u c
e
u
e
u
u e
|
|
|
e |
|
sin
sin
cos
=
+ =
=
(39)
By deriving h on time, the variable N of equation (9) is
obtained:
( )
(
(
+
+
=
1
sin
1
2 2 cos
2
1
c
u
u u u
e
N
|
| o |
(40)
where
c
u
u u
e
|
| o
sin
=
(41)
The inverse of N is
( ) ( )
( )
( ) ( )
(
(
+
+
=
N N e
N N
N u u
c
u
u
det
2 cos
det
sin
1
det 2 det 1
1
2
1
o | |
|
(42)
where
( )
u u
N | o cos 2 det
1 2
+ =
(43)
A sufficient condition for the equation (43) be not null is
that
2 t | s
u
. The assumptions 1, 2 and 3 are true
2 t | s
u
and e
c
> 0. Then, choosing h
d
= 0, 0 =
d
h
and
0 =
d
h
, the adaptive control law guarantees that h 0 and
0 h
as t .
A further analysis can be made over the function h. When
the second element of h, i.e. h
2
=
2
c
+
u
, is converging to
zero, then the sign of
u
tends to be the opposite sign of
c
. As a
consequence, the vehicle goes around the origin of <g>
decreasing
c
(see Fig. 4). Remark. When e
c
is small, the
execution of the posture control can cause control problems
because the control system is close to a singular case, i.e. a
state that e
c
= 0. Besides, the calculation of
c
and
u
is very
sensitive to noise. Thus the controller must be turned off when
e
c
is less than a small boundary.
Figure 4. Example of the situation when h
2
is close to zero in the
posture control.
3.2 Definition of h for the heading control
In this case, the control objective is to obtain a
convergence of the heading of the vehicle to the orientation of
the x
g
-axis of the reference coordinate <g> (see Fig. 3) and
obtain a stable or bounded characteristic of the vehicle position.
Let h be as follows:
(
+
=
g
g g g g
y x
sin cos
h (44)
where
x
g,
y
g and
=
=
=
g
g g
g g
u y
u x
sin
cos
(45)
The following position error
e
p is defined to analyze the
relation between the position error and the first element of h.
2 2
g g p
y x e + = (46)
The temporal derivative of
e
p is
g g g g p
u y u x e sin 2 cos 2 + = (47)
that results in
1
2uh e
p
= (48)
J. Med. Biol. Eng., Vol. 32 No. 1 2012
66
The boundedness of u and h
1
was stated in the subsection
2.3 if assumptions 1 and 2 are true. That such assumptions are
positive is verified further in the next paragraph. Therefore,
p
e
is bounded. Furthermore, it can be obtained that the
inequality
( )
1
sup 2 h u e
p
s
, where sup (u) is the supreme of u,
implies that p
e
converges to zero when h
1
converges to zero.
This conclusion is not enough to conclude that e
p
converges to
a finite value. Unfortunately, the relation between u and h
1
involves the dynamic and parameter errors of the robot that
make it difficult to establish whether p
e
is a negative function.
However, e
p
varies more slowly as time passes. In the worst
case, e
p
may be increased considerably over a long time. In
order to avoid e
p
is greater than a considerable value, the
controller has to be turned off after a short time, but long
enough to attain a very small heading error. Considering that
p
e
is bounded, e
p
is bounded if the controller is turned off
after a finite time. If h
1
0 as t , then the scalar product
of the vector [x
g
y
g
]
T
by the unitary vector of the vehicle
heading converges to zero t . Therefore, either e
p
is
decreasing or the orientation is tending to be perpendicular to
the vector [x
g
y
g
]
T
. In the experimental results (section 4) one
can see that e
p
tends to decrease and no divergence of e
p
is
observed.
By deriving h on time, the variable N of the equation (9)
is obtained:
(
=
1 0
sin cos 1
g g g g
x y
N
(49)
The matrix N
-1
is
(
+
=
1 0
sin cos 1
1 g g g g
x y
N
(50)
The assumptions 1, 2 and 3 are true if x
g
and y
g
are
bounded; such conditions can be guaranteed if the controller is
turned off after a finite time. Therefore, choosing h
d
= 0,
0 =
d
h
and
0 =
d
h
, the adaptive control law guarantees that
h
d
0 and 0
d
h