2010 IEEE/ASME International Conference on

Advanced Intelligent Mechatronics

Montréal, Canada, July 6-9, 2010

Adaptive Nonlinear Controller Design For Differential-Drive Mobile

Platforms
Sousso Kelouwani, Member, IEEE, Christian Ouellette and Paul Cohen

Abstract— Most differential drive platforms are equipped of the casters when driving backward; however, no controller
with two independent motors and casters. Possible misalign- structure was proposed. In [2], a fuzzy logic based controller
ment between the platform direction and caster orientations was proposed to improve a differential platform motion on
and the unbalanced motor dynamics are some of the factors
that make it difficult to properly drive the platforms. We present an inclined surface. Here again, no straightforward platform
an adaptive nonlinear controller system that reduces the effect mobile controller was derived and the stability of the over-
of disturbances on the platform motion. The controller system all control strategy was not provided. In [9], a simulated
has two main parts: (1) a kinematic controller based upon Lya- model of a mobile platform was proposed and a fuzzy
punov’s stability theory and (2) a dynamic compensator based logic control was derived. This controller aimed to keep the
upon the reference model approach in order to allow robust
trajectory tracking. Both controllers have been implemented platform stationary when the front casters are lifted, but no
and tested on two different types of powered wheelchairs. The evidence was provided about the stability scheme. Moreover,
results suggest that the proposed nonlinear controllers provide no experimentation was provided to assess the controller
better trajectory tracking performance than well-known PID operations. Astolfi [11] has proposed an exponentially stable
controllers. wheeled mobile robot dynamic controller by assuming that
I. INTRODUCTION the motor dynamics is fully compensated. However, in real
applications, this assumption is not always true. We pro-
The propulsion of a differential drive mobile platform
pose in this paper an adaptive and straightforward design
is usually done by two independent motors. The control
approach of nonlinear kinematic and dynamic controllers
of the platform configuration (position and orientation) is
that reduce significantly the effect of disturbances on the
challenging because it has non-holonomic constraints and
platform motion. This approach is related to the general
the motors are behaving under nonlinear dynamics. Usually,
stabilizing algorithm proposed in [12]. Furthermore, exten-
front and rear casters are added to the platform for stability
sive experiments on two powered wheelchairs with different
purpose. During the motion, these casters’ orientations can
drives, a centrally driven and a rear drive, are presented.
cause disturbances, as well as the type of navigation surface,
The rest of the paper is organised into 5 sections. Section
the varying load, the location of the center of gravity and
2 presents the differential drive platform dynamic model.
the center of gyration estimates [1], [2]. Several approaches
In section 3, the problem is formulated, whereas section 4
have been used to reduce the impact of perturbations in
discusses the adaptive nonlinear controller design approach.
general. But few of them have proposed a straightforward
Experimentation and conclusion are presented in section 5
design approach to limit the effect of such disturbances.
and section 6, respectively.
Recently, Dan [3] has proposed a review of the most used
control systems for the differential drive: speed control [4], II. D IFFERENTIAL -D RIVE P LATFORM DYNAMIC M ODEL
traction control [5], [6], active suspension control and sta-
bility control. She outlined the lack of a straightforward and Based upon Newton-Euler equation the following dynamic
robust design approach. Among speed controllers commonly model is obtained:
used, the Proportional-Integral-Derivative (PID) controller
M(Q)Q̈ +C(Q, Q̇)Q̇ + G(Q) = B(Q)Γ (1)
[7], [9] is the most implemented. Knowledge of the kinematic
model of the platform is required to set the PID coefficients with the following non-holonomic constraint:
[10]. However, this knowledge is not always accurate and
operation conditions (slope, navigation surface, front and J(Q)Q̇ = 0 (2)
rear caster orientations, etc.) may change during the motion.
where Q = [X,Y, θ ] , (X,Y ) and θ are, respectively, the
To allow the PID control system to adapt to such changing
configuration vector, the coordinates of the center of gyration
conditions, the PID coefficients must be continuously mod-
and the orientation of the platform as shown in Figure 1.
ified. However, stability of these adaptive schemes is not
M(Q) is a (3 × 3) matrix of the system mass and inertia.
provided [4]. In [1], the author proposed a dynamic model
C(Q, Q̇) is a (3 × 1) vector of the Coriolis and centripetal
of a differential drive platform that considers the orientation
forces, G(Q) is a (3 × 1) vector of the gravitational forces
The authors are with the Perception and Robotic and torques. Γ = [τl , τr ] is a (2 × 1) vector of the motors
Laboratory, École Polytechnique de Montréal, Québec, input torques (left and right). B(Q) is the (3 × 2) input
Canada [email protected],
[email protected], transformation matrix. J(Q) is the Jacobian matrix of the
[email protected] system. Since the platform is moving on a 2D surface,

978-1-4244-8030-2/10/$26.00 ©2010 IEEE 1238

G(Q) = 0 and C(Q, Q̇)Q̇ = 0 and Equation 1 may be rewritten This condition is very difficult to obtain in practice for the
as: following reasons: (i) since the two motors are not identical,
a perfect torque synchronisation is not easy to obtain; (ii)
M(Q)Q̈ = B(Q)Γ (3) friction conditions for each wheel are different and difficult
to assess; (iii) estimations of the total mass, the inertia, the
with:
⎡ ⎤ ⎡ ⎤ center of gravity and the center of gyration locations are
cos(θ ) cos(θ )
m 0 0 difficult to obtain accurately. The problem is to design control
⎢ r r ⎥
M(Q) = ⎣ 0 m 0 ⎦ , B(Q) = ⎣ sin(θ ) sin(θ )
⎦ (4) laws that reduce significantly the effect of these disturbances
r r
0 0 i l
− rl and uncertainties on the platform motion.
r
Considering the platform as symmetrical, l measures the dis- IV. A DAPTIVE C ONTROLLER D ESIGN
tance between the powered wheels and the center of gyration A. Representations of the Navigation Environment and Con-
and r is the wheel radius. m and i represent respectively the trol System
(platform, user) mass and the system inertia. The navigation environment is represented in Figure 1. At
The linear speed v and the angular ω speed, expressed in each time, the mobile platform configuration is represented in
the platform coordinate frame are the two most commonly the reference coordinate frame by Q = [X,Y, θ ] . At the same
used control signals. For design purposes, it is convenient to time the reference configuration induced by the user’s control
separate the platform kinematic and dynamic models. If we signals vr and ωr is Qr = [Xr ,Yr , θr ] . To analyse the nonlinear
assume that: control problem, a configuration error must be defined. In this
τr + τl τr − τl paper, we adopted the formulation presented in [13], [15].
τL = , τA = l , (5)
r r The configuration error qe and its derivative q˙e are expressed
then Expression 3 is rewritten as followed: in the platform reference system by:
⎡ ⎤ ⎡ ⎤⎡ ⎤
τL τA xe cos(θ ) sin(θ ) 0 Xr − X
v̇ = , ω̇ = (6)
m i qe = ⎣ ye ⎦ = ⎣ −sin(θ ) cos(θ ) 0 ⎦ ⎣ Xr −Y ⎦
and θe 0 0 1 θr − θ
Ẋ = vcos(θ ) ⎡ ⎤ ⎡ ⎤ (8)
Ẏ = vsin(θ ) (7) x˙e vr cos(θe ) − v + ye w
θ̇ = ω q˙e = ⎣ y˙e ⎦ = ⎣ vr sin(θe ) − xe w ⎦ (9)
θ˙e wr − w
The control law must be designed so that qe → 0 for the
given input control signals [v, ω] despite the presence of
perturbations. To achieve this goal, we adopt the following
nonlinear block diagram of the overall control system.

Fig. 2. Platform Control System

The platform user provides input control signals [vr , ωr ] .

Fig. 1. Representation of the Navigation Environment: (R, x, y) is the Using these signals, the Configuration Generator module
mobile platform coordinate frame. (O, X,Y ) is the reference coordinate
frame. Qr = [Xr ,Yr , θr ] is the mobile configuration on the reference path provides Qr to the Nonlinear Kinematic Controller. This
that should be tracked at any given time. Q = [X,Y, θ ] is the current mobile
 module will generate the linear vt and angular ωt target
platform configuration. The orientation of the platform is: θ = tan−1 ẎẊ . velocities based upon the current configuration estimation Q
and the reference configuration Qr . In order to let the control
system remain robust against perturbations, the Dynamic
III. P ROBLEM F ORMULATION
Controller generates the required accelerations [v̇, ω̇] based
Consider a two-wheel drive system with an independent upon the platform velocities [v, ω] and the target velocities
motor on each wheel. Assume that the center of gravity is [vt , ωt ] .
the same as the center of gyration for the mobile platform.
Furthermore, assume that v = 0, ω = 0 to allow the platform B. Reference Configuration Generator
to follow a straight line. According to Equation6, mv̇ = τL Given [vr , ωr ] , the Reference Configuration Generator
and 0 = τA . From Equation5, we must have τr = τl if τA = 0. uses dead reckoning in order to compute Qr [14].

1239
C. Nonlinear Kinematic Controller where:
t
The goal of the Nonlinear Kinematic Controller is to K1 (θe ) = α θe dt , α ∈ [0, 1] (18)
generate [vt , ωt ] by assuming that the platform dynamic is 0

completely modelled. Several approaches have been used t

to design adaptive nonlinear controllers in particular H∞ , K2 (xe ) = β xe dt , β ∈ [0, 1] (19)
0
adaptive and Lyapunov based approaches [3]. We select
t
Lyapunov methods since they are extensively used for a K3 (ye ) = δ ye dt , δ ∈ [0, 1] (20)
large class of nonlinear systems and the controller design 0
is straightforward. where Λ is a real value that can be equal to 0 or 1.
Given the configuration error represented by Equation8, Case 1:Λ = 1
the following Lyapunov stabilization function candidate is Substituting vt and ωt in Equation11 by vtr and ωtr and
selected [8]: calculating the new expression of V̇ (qe ) yield:
1 1
V (qe ) = xe2 + y2e + (1 − cos(θe )) (10) V̇ (qe ) = −Kx xe2 − Kθ sin2 (θe ) + Ψ(qe ) (21)
2 2
Since [vt , ωt ] are the desired velocities for kinematic stabil- where
ity, according to Equation9, we have: Ψ(qe ) = −xe K1 (θe ) − xe K2 (xe ) − sin(θe )K1 (θe )
⎡ ⎤ ⎡ ⎤ (22)
x˙e vr cos(θe ) − vt + ye wt −sin(θe )K3 (ye )
⎣ y˙e ⎦ = ⎣ vr sin(θe ) − xe wt ⎦ (11)
θe
˙ wr − wt If Ψ(qe ) ≤ 0 and its integral is positive, then V̇ (qe ) ≤ 0 and
the system is stable. On the other hand, if Ψ(qe ) ≥ 0 and its
According to Lyapunov stability theory, the function candi-
integral is positive, two situations must be considered. The
date must satisfy V (qe ) ≥ 0, ∀qe and V (0) = 0. The system
first situation occurs when |Ψ(qe )| ≤ | − Kx xe2 − Kθ sin2 (θe )|.
(Kinematic Controller and platform dynamic) is asymptot-
In this case, the system will remain stable because V̇ (qe ) ≤
ically stable if [vt , ωt ] is selected so that V̇ (qe ) ≤ 0, ∀qe .
0. The second situation occurs when |Ψ(qe )| ≥ | − Kx xe2 −
Hence, qe = 0 is an equilibrium point. Given Expression 10,
Kθ sin2 (θe )|. In this case no stability is guarantied and in
V̇ (qe ) is represented by the following expression:
order to avoid instability, one must set Λ = 0.
V̇ (qe ) = xe vr cos(θe ) − xe vt + ye vr sin(θe ) Case 2:Λ = 0
(12)
+wr sin(θe ) − wt sin(θe ) In this case, one must set K1 (θe ) = 0, K2 (xe ) = 0 and
K3 (ye ) = 0. Expressions 15 and 16 become vtr = vt and
Let us assume that:
ωtr = ωt and the system is stable.
vt = vr cos(θe ) + Kx xe (13)
E. Nonlinear Dynamic Controller
ωt = ωr + vr ye + Kθ sin(θe ) (14) The proposed kinematic controller assumes that Equations
where Kx > 0 and Kθ > 0 are real constants. Substituting vt 6 and 7 are verified. However, the system mass m (the sum of
and ωt in Equation12 by their Expressions 13 and 14, leads the platform mass and the user mass) and the system inertia
to the following expression: i are not exactly known. Moreover, the friction conditions
between the wheels and the navigation surface, the motor
V̇ (qe ) = −Kx xe2 − Kθ sin2 (θe ) (15) impedances, the battery charge and the casters are some
of usual sources of perturbations that affect the platform
Since Expression 15 is negative for all qe and since motion. The role of the dynamic controller is to generate
V (0) = 0, Expression 10 is a valid Lyapunov function. the required accelerations so that the platform will move
Hence, the two kinematic control laws represented by at speeds vt and ωtr suggested by the nonlinear kinematic
Equations 13 and 14 guarantee the stability of the system. controller regardless of the presence of the aforementioned
disturbances. The generally adopted approach to deal with
D. Additional Robustness to the Kinematic Controller disturbances is adaptive control. In this section, we opted
for the adaptive approach presented in [16] and we propose
Let us assume bounded effects of disturbances on the plat- a dynamic controller based upon Lyapunov stability analysis
form’s dynamics. We append the Kinematic Controller with with a reference model tracking. Two dynamic controllers
the integral of qe in order to improve the convergence toward are designed for the linear velocity vtr and for the angular
an equilibrium point. The new kinematic linear velocity vtr velocity ωtr , respectively.
and angular velocity ωtr are represented by the following 1) Linear velocity Dynamic Controller: Finding the ref-
expressions: erence model vm is the first step toward the design of the
vtr = vt + Λ (K1 (θe ) + K2 (xe )) (16) Lyapunov based adaptive controller. The second step consists
of designing a stabilization control law that will allow v̇ to
ωtr = ωt + Λ (K1 (θe ) + K3 (ye )) (17) follow v̇m .

1240
 a) Reference Model: If m is known exactly and if there 2) Angular velocity Dynamic Controller: As a result of
is no disturbance, then: the design approach presented in section IV-E.1, the refer-
ence model ωm is represented by:
vm = vtr (23)
ω̇m = ω˙tr + λω (ωtr − ωm ) (33)
Taking the Laplace transform of Equation23 yields:
where λω is a constant that depends upon the system
s + λv r dynamic. The angular acceleration control law is given by:
vm (s) = vtr (s) ⇔ vm (s) = v (s) (24)
s + λv t
ω̇ = θ3 ω˙tr + θ4 λω (ωtr − ω) (34)
where λv is a constant that depends mainly on the platform
dynamic transfer function. By rewriting this equation in the where θ3 and θ4 are parameters to be adapted. The angular
time domain, we get the expression of the reference model velocity error is:
of the linear velocity:
eω = ω − ωm (35)
v̇m = v˙tr + λv (vtr − vm ) (25)
Let us assume that φ3 = (1 − θ3 ) and φ4 = (λω − θ4 ).
b) Linear velocity Reference Model Tracking: Given Equation 36 represents a Lyapunov function candidate.
the reference model 25, a control law that must allow a good 1 1 1
and stable tracking of v̇m must be adopted. In order to follow Vω (eω , θ3 , θ4 ) = (e2ω + φ32 + φ42 ) (36)
2 γω γω
v̇m , we propose the following control law:
The adaptive law is given by the following two expressions:
v̇ = θ1 v˙tr + θ2 λv (vtr − v) (26)
t
θ3 = −γω eω ω˙tr dt (37)
where θ1 and θ2 are control parameters that must be adapted 0
in order to follow v̇m . Note that perfect tracking occurs when
v = vtr and θ1 = 1. Based upon Equations 25 and 26, the t
adaptation error is represented by: θ4 = −γω eω (ωtr − ω) dt (38)
0

ev = v − vm ⇒ e˙v = v̇ − v̇m (27) V. E XPERIMENTS

Substituting v̇ and v̇m in Equation27 by their expressions 25 A. Experimental Setup

and 26 respectively, leads to the derivative of ev : Several experiments were conducted in order to test the
kinematic and dynamic controllers. The goal of the first
e˙v = −λv ev − (1 − θ1 )v˙tr − (λv − θ2 )(vtr − v) (28) experiment is to compare PID controllers and the proposed
nonlinear controllers during a straight and a curvilinear
Given e˙v , the updating laws for θ1 and θ2 must guarantee
path following. Through the second experiment, we tested
the stability of the system. According to Shankar [16], a
the robustness of the controllers. A middle drive powered
Lyapunov candidate function is:
wheelchair, model Quicky F11 from Sunrise Medical was
1 1 1 used to perform these two experiments. A 75kg load was
Vv (ev , θ1 , θ2 ) = (e2v + φ12 + φ22 ) (29) used on the wheelchair. At the beginning of each experiment,
2 γv γv
the front caster orientations were set perpendicularly to
where γv > 0 is a constant whose value is chosen in order to the direction of the motion in order to add a maximum
meet a predefined dynamic performance. φ1 = (1 − θ1 ); φ2 = disturbance.
(λv − θ2 ). Clearly, Vv (ev , θ1 , θ2 ) > 0, ∀(ev , θ1 , θ2 ) = (0, 0, 0) Since the locations of the gravity center and the gyration
and Vv (ev , θ1 , θ2 ) = 0 when ev = 0, φ1 = 0 and φ2 = 0. center of a rear drive powered wheelchair are very different,
its angular control through ω is difficult to achieve. The third
φ̇1 φ̇2 experiment aims to demonstrate that, the proposed design
V˙v (ev , θ1 , θ2 ) = −λv e2v + φ1 ( − ev v˙tr ) + φ2 (ev v + − ev vtr )
γv γv approach can be applied for the rear-wheel type powered
(30) wheelchairs as well. For this experiment, we use a 75kg load
Selecting φ̇2 = γv ev (vtr − v) and φ̇1 = γv ev v˙tr , ensures that on a Quicky 646 from Sunrise Medical.
V˙v (ev , θ1 , θ2 ) ≤ 0, ∀(ev , θ1 , θ2 ). Hence, the adaptive laws for Figure 3 shows the trajectory obtained with the Quicky F11
θ1 and θ2 are: platform when [vr , ωr ] = [0.3m/s, 0rad/s] and no controller
t
θ1 = −γv ev v˙tr dt (31) is used. Rather than describing a straight line (comparable to
0 the reference trajectory represented by the dotted black line),
the trajectory is curvilinear (gray curve), indicating that one
t
θ2 = −γv ev (vtr − v) dt (32) of the two motors is faster than the other. The deviation at
0 the end of the 10m move is more than 1.9m.

1241
 Figures 5 and 6 show the error between the input velocity and
the platform velocity. Both controllers (PID and nonlinear)
errors are bounded in a small interval. However, the angular
correction introduced by the nonlinear controllers is more
important compared to the same correction applied by the
PID, in order to maintain a good linear motion tracking
(Figure 6).

Fig. 3. Open Loop Trajectory for a Middle-Wheel Powered Wheelchair

B. Experiment 1: Straight Line Motion

The purpose of the test is to demonstrate the nonlinear
controller’s effectiveness in following a straight line trajec-
tory. The following parameter values are used: [vr , ωr ] =
[0.3m/s, 0rad/s] , Kx = 1, Kθ = 1, α = 0.01, β = 0.001, δ =
0.001. The sampling frequence is set to 10Hz. For com-
parison purposes, we designed two PID controllers: one
for the linear velocity and another one for the angular Fig. 5. Linear velocity Errors
velocity. Two Root Mean Square Error (RMSE) RSMEX
and RSMEY are selected as comparison measures. RSMEX =
∑Ni=1 (Xr (i) − X(i))2
, where N and i are respectively the
N
number of samples during the experiment and the sample
∑Ni=1 (Yr (i) −Y (i))2
index. RSMEY = .
N
Figure 4 shows that the nonlinear controllers were able to

Fig. 6. Angular velocity Errors

C. Experiment 2: Straight Line Motion Including Inclines

Fig. 4. Trajectory Comparison Robustness is examined by inclunding two incline plans
on the straigth line path as shown in Figure 7. From its rest
maintain a rectilinear motion of the platform despite the position, the platform must move straightforward by 10m. At
non identical motors (dotted gray curve in Figure 4). Ten 3.72m, it must get over a 2.5cm threshold, ascend the first
trials were performed while recording RSMEX and RSMEY . incline (length: 1, 25m, width: 1.2m, slope: 10%), must roll
The average of RSMEX and RSMEY , when the proposed on a horizontal plan (length:1, 22m, width: 1.2m) and must
controllers were use are respectively 0.10 and 0.02. On the descend a decline (length: 1, 25m, width: 1.2m, slope: 10%).
other hand, when PID controllers are used, the average values Finally, it must roll straightforward for 2.56m.
obtained for RSMEX and RSMEY are respectively 0.16 and The obtained RSMEX and RSMEY values are 0.18 and 0.2
0.35. These results suggest that the controllers presented respectively. These results show that the controllers can effi-
in this paper perform better than classical PID controllers. ciently handle perturbation commonly found in a navigation

1242
 When the nonlinear controller are used, the results are
similars to the one obtained with the middle drive platform:
RSMEX and RSMEY are repectively 0.15 and 0.11.

Fig. 7. Setup for Robustness Testing 2) Curvilinear Motion: In Figure 10 the trajectory de-
scribed when using nonlinear controllers is closer to the
reference trajectory than is the one obtained with PID con-
trollers. However, the obtained trajectory following results on
the middle wheel wheelchair are better than the one obtained
with the rear-wheel wheelchair. Recall that in Section III,
we made the assumption that the locations of the center of
gyration and the center of gravity are the same. Although
this assumption is not true when using the rear driven
wheelchair, the proposed controllers performed better than
PID controllers.

Fig. 8. Trajectories Obtained with Inclined Sections for the Middle-Wheel

Powered Wheelchair

environment with powered wheelchairs. We were unable to

complete the same test when we used PID controllers. Due to
insufficient compensation of perturbations introduced by the
2.5cm threshold and by the inclined surface, the wheelchair
was unable to remain on the first incline.
D. Experiment 3: Curvilinear Trajectory Following
The input signals of the platform were designed to describe
an ’8-shape’ trajectory. In Figure 9, the solid black curve Fig. 10. Rear Driven Wheelchair: Comparison of Controllers for Curvi-
linear Trajectory Following
represents the reference trajectory. One can see in this Figure
that the PID controllers (gray solid curve) were unable
to respect the reference trajectory, whereas the nonlinear
VI. C ONCLUSION
Lyapunov based controllers (gray dashed curve) were able
to faithfully recreate the reference trajectory. The design of adaptive nonlinear kinematic and dynamic
controllers have been presented. Both controllers are based
on Lyapunov theory in order to find stabilization functions.
By adding integrators, the convergence of the kinematic
controller is improved. We provided stability conditions for
that controller. We also proposed a stable dynamic controller
design based upon the reference model theory. Experiments
performed on a centrally driven and a rear drive wheelchairs
illustrated the effectiveness and the stability of the proposed
approach compared to PID controllers.

VII. ACKNOWLEDGEMENT
This work has been supported by the Natural Science and
Engineering Research Council of Canada (Grant No CRD
349481-06 Scholarship No BESC D3-348674-2007). The
Fig. 9. Middle-Wheel Powered Wheelchair: Comparison of Controllers for
Curvilinear Trajectory Following collaboration and support of Robovic Inc., Sunrise Medical
Canada are as well gratefully acknowledged. The authors
wish to thank Hai Nguyen, Patrice Boucher, Vincent Zalzal,
E. Experiment 4: Rear Driven Wheelchair Control Raphael Gava and Alexandre Fortin from the Perception
1) Straight Line Motion: With the rear drive wheelchair and Robotics Laboratory of Ecole Polytechnique for their
Quicky 646, we performed the straightforward motion test. contributions on implementation and testing.

1243
 R EFERENCES
[1] D. Ding, R. A. Cooper, S. Guo and T. A. Corfman, ”Analysis of Driv-
ing Backward in an Electric-Powered Wheelchair”, IEEE Transaction
on Control Systems Technology, (12)6,934-943, 2004.
[2] Murakami, H., Seki, H, Minakata, H, Tadakuma, S, ”Operationality
improvement control of electric power assisted wheelchair by fuzzy
algorithm”, 2009 IEEE International Conference on Industrial Tech-
nology, 2009.
[3] Dan Ding, Rory A. Cooper, ”Electric-powered wheelchairs- A review
of current technology and insight into feature directions”, IEEE
Control System Magazine, 22-34, 2005.
[4] L. Boquete, R. Garcia, R. Barea and M. Mazo, ”Neural Control of
the Movements of a Wheelchair”, Journal of Intelligent and Robotic
Systems, 213-226, 1999.
[5] Makki A. and Siy P., ”A robust traction controller design using
the theory of variable structure systems”, Proc. 36th Midwest Symp.
Circuits and Systems, Detriot, MI, 356-359, 1993.
[6] Venkat Krovi and Vijay Kumar, ”Optimal Traction Control In A
Wheelchair With Legs And Wheels”, Proc. 4th Ann. Applied Mecha-
nisms and Robotics Conference, AMR- 030, (30), 1995.
[7] Ruei-Xi Chen, Liang-Gee Chen and Lilin Chen, ”System Design
Consideration for Digital Wheelchair Controller”, IEEE Transaction
on Industrial Electronis, (47)4,898-907, 2000.
[8] Kanayama, Y., Kimura, Y., Miyazaki, F., Noguchi, T., ”A Stable
Tracking Control Method for an Autonomous Mobile Robot”, IEEE
International Conference on Robotics and Automation, (1), 384-389,
1990.
[9] S. Ahmad and M.O. Tokhi, ”Modelling and Control of a Wheelchair
on Two Wheels”, Second Asia International Conference on Modelling
and Simulation, 579-584, Kuala Lumpur, Malaysia, 2008.
[10] Brown, K.E. Inigo, R.M. Johnson, B.W., ”Design, Implementation, and
Testing of an Adaptable Optimal Controller or an Electric Wheelchair”
Industry Applications Society Annual Meeting, (2),2332-2339, San
Diego, CA, 1989.
[11] stolfi, A., ”A robustly stabilising adaptive controller for systems
in feedback form” Journal of Dynamic Systems, Measurement, and
Control, (121)121-126 2006.
[12] Karagiannis, D. Astolfi, A., ”A robustly stabilising adaptive controller
for systems in feedback form” American Control Conference, 3557-
3562 2006.
[13] Kanayama Y. and Miynke N. ”Theory of two-dimensional transforma-
tions”, IEEE Trans. Robot. Autom. (14), 827-834, 1998.
[14] Jin-Xia Y. Zi-Xing C. Zhuo-Hua D., ”Dead reckoning of mobile robot
in complex terrain based on proprioceptive sensors ” International
Conference on Machine Learning and Cybernetics, (4), 1930-1935,
2008.
[15] Kanayama, Y. Nilipour, A. Lelm, C.A. ”A Locomotion Control
Method for Autonomous Vehicles”, IEEE International Conference
on Robotics and Automation, (2)1315-1317, Philadelphia, PA, USA,
1988.
[16] Shankar Sastry and Marc Bodson ”Adaptive Control: Stability, Con-
vergence, and Robustness”, Prentice-Hall Advanced Reference Series
(Engineering), New Jersey, 1994.

1244