10th IFAC Symposium on Robot Control

International Federation of Automatic Control

September 5-7, 2012. Dubrovnik, Croatia

Four-state Trajectory-tracking Control Law

for Wheeled Mobile Robots
Sašo Blažič ∗
∗
University of Ljubljana, Faculty of Electrical Engineering, Tržaška
25, Ljubljana, Slovenia (e-mail: saso.blazic@fe.uni-lj.si).

Abstract: In this paper a four-state kinematic model is proposed where the transformation
between the robot posture and the system state is bijective. Some nonlinear control laws are
constructed based on the kinematic model in the Lyapunov stability analysis framework. These
control laws achieve global asymptotic stability of the system based on the non-zero reference
velocities requirements.

Keywords: Mobile robot, Kinematic model, Nonholonomic models, Lyapunov stability, Error
model

1. INTRODUCTION approaches only guarantee local stability, while others

The problem of nonholonomic systems control has at-
tracted numerous researchers in the past. A thoroughly
studied case with great practical significance is wheeled-
mobile robot with kinematic model similar to a unicycle.
The differentially driven mobile robots that are very com-
mon in practical applications also have the same kinematic
model. Although many researchers coped with the more
difficult problem of stabilizing dynamic models for differ-
ent types of mobile robots, see e.g. Jiang and Nijmeijer
(1997), the basic limitations of mobile-robot control still
come from their kinematic model, as shown in Brockett
(1983), Morin and Samson (2009), and Lizarraga (2004).
Kinematic control laws are also very important from the
practical point of view, since the wheel-velocity control
is often implemented locally on simple, micro-controller-
based hardware, while the velocity command comes from
high-level hardware that also provides the current control
objective.
Traditionally, the problem of mobile-robot control has
been approached by point stabilisation (Pourboghrat,
2002) or by redefining the problem as a tracking-control
one (Kanayama et al., 1988). There are also some
approaches that tackle both problems simultaneously
(Buccieri et al., 2009). We believe that the tracking-
control approach is somewhat more appropriate, since the
nonholonomic constraints and other control goals (obsta-
cle avoidance, minimum travel time, minimum fuel con-
sumption) are implicitly included in the path-planning
procedure (LaValle, 2006; Pozna et al., 2009). It is also
easier to extend this approach to more complex schemes
such as the control of mobile robot platoons (Klančar
et al., 2011). Many control algorithms were proposed in
the path-tracking framework, such as PID (Kanayama
et al., 1990), Lyapunov-based nonlinear controllers (Sam-
son, 1993), adaptive controllers (Jiang and Nijmeijer,
1997), model-based predictive controllers (Klančar and
Škrjanc, 2007), fuzzy controllers (Guechi et al., 2010), etc.
In some cases they are implemented on chips or other
industrial hardware (Precup and Hellendoorn, 2011). Some
also ensure global stability and global convergence under
certain assumptions.
It is very important to find a (kinematic) control law
that produces a smooth control signal. If this is not the
case, the implementation on the dynamic model becomes
impossible. Unfortunately, due to a discontinuity in the
orientation error of ±180◦, quite often there is also a dis-
continuity in the angular-velocity command. This comes
from the fact that the classical kinematic model is continu-
ous with respect to orientation (there are no jumps at ±π),
while in implementation the orientation is often mapped
to the (−π, π] interval. In this paper a novel kinematic
model is proposed that overcomes this difficulty, although
it comprises more states. A control law that achieves global
asymptotic convergence to a predesigned path under some
mild conditions is also proposed and compared to existing
control laws.
2. PROBLEM STATEMENT
Assume a two-wheeled, differentially driven, mobile robot
like the one depicted in Fig. 1, where (x, y) is the wheel-
axis-centre position and θ is the robot orientation. The
kinematic motion equations of such a mobile robot are
equivalent to those of a unicycle. Robots with such an
architecture have a nonholonomic constraint of the form:
ẋ(t)
[ − sin θ(t) cos θ(t) ] =0 (1)
ẏ(t)
resulting from the assumption that the robot cannot move
in the lateral direction. Only the first-order kinematic
model of the system will be treated in this paper:
  " #
ẋ cos θ 0  
  v
q̇ = ẏ = sin θ 0 (2)
w
θ̇ 0 1
where q T (t) = [ x(t) y(t) θ(t) ] is the vector of generalized
coordinates, while v and w are the translational and the
angular velocities, respectively, of the system in Fig. 1.
The velocities of the right and the left wheels of the robot

978-3-902823-11-3/12/$20.00 © 2012 IFAC 399 10.3182/20120905-3-HR-2030.00063

IFAC SYROCO 2012
September 5-7, 2012. Dubrovnik, Croatia
v w
w q
y model is given by:
B ėy = −wr ex + vr sin eθ − ex wb
x
Fig. 1. Two-wheeled, differentially driven, mobile robot
are vR = v + wB wB
2 and vL = v − 2 , respectively, where B
is the robot inter-wheel distance. The control design goal
is to follow the virtual robot or the reference trajectory,
defined by
qrT (t) = [ xr (t) yr (t) θr (t) ] (3)
where qr (t) is a-priori known and smooth. It is very easy to
show that the system (2) is flat (Fliess et al., 1995), with
the flat outputs being x and y. Consequently, (3) can be
produced by the uniformly continuous control inputs vr (t)
and wr (t) in the absence of initial conditions, parasitic
dynamics and external disturbances. The goal is to design
a feedback controller to achieve the tracking and the
tracking should be asymptotic under the persistency of
excitation (PE) through vr (t) or wr (t).
3. ERROR MODEL OF THE MOBILE-ROBOT
KINEMATICS
The posture error is not given in the global coordinate
system, but rather as an error in the local coordinate
system of the robot: ex gives the error in the direction
of driving, ey gives the error in the lateral direction, and
eθ gives the error in the orientation. The posture error
T
e = [ ex ey eθ ] is determined using the actual posture
T
q = [ x y θ ] and the reference posture qr = [ xr yr θr ] :
T s(t) = sin(θ(t)) and c(t) = cos(θ(t)). Their derivatives are:
" # " # ṡ(t) = cos(θ(t))θ̇(t) = c(t)w(t)
ex cos θ sin θ 0
ey = − sin θ cos θ 0 (qr − q) (4)
eθ 0 0 1
3.1 Three-state error model of the system
From (2) and (4) and assuming that the virtual robot has
a kinematic model similar to (2), the posture-error model
can be written as follows:
" # " #  " −1 ey #
ėx cos eθ 0 
vr
ėy = sin eθ 0 + 0 −ex u (5)
wr
ėθ 0 1 0 −1
The transformation (4) is theoretically imposed by the
group operation, noting that the model (2) is a system
in the Lie group SE(2) (Morin and Samson, 2009). The
approach itself was adopted by Kanayama et al. (1988),
where the authors also proposed PID control for the
stabilization of the robot at the reference posture. Later,
many authors used the error model (5) for the tracking-
control design.
Very often, e.g. in Kanayama et al. (1990), the following
control u is used to solve the tracking problem:
400
   
v vr cos eθ + vb
u= = (6)
wr + wb
where uTb = [ vb wb ] is the feedback signal to be deter-
mined later. Inserting the control (6) into (5), the resulting
ėx = wr ey − vb + ey wb
(7)
ėθ = −wb
3.2 Four-state error model of the system
The problem of using the three-state error model presented
in the previous section is that the transformation between
the robot posture and the error model is not bijective.
This can be observed from the fact that any error state
T
[ ex ey eθ + 2kπ ] for fixed ex , ey , eθ , and for arbitrary
k ∈ Z corresponds to the same robot posture. To say this
more clearly: If we take any robot posture and rotate the
robot for any multiple of 360◦, the same robot posture is
obtained (the sensors would not observe any difference be-
tween the two postures). Consequently, by just observing
the robot posture it is impossible to deduce the orientation
error. In practical control implementations the orientation
error is often mapped onto the interval (−π, π] to somehow
overcome the above mentioned bijectivity problem. The
side effect of this is that the (angular velocity) control
signal often expresses discontinuity when the orientation
error of ±π is crossed (this will be shown in the example
at the end of the paper). Discontinuous velocity control
signals are even more problematic because of the imple-
mentation on the real dynamic system.
The bijectivity between the robot posture and the states of
the system should be therefore reflected in the kinematic
model of the system and also in the error model of the
system. This can be achieved by increasing the number
of system states to 4. The variable θ(t) from the original
kinematic model (2) is exchanged by two new variables
(8)
ċ(t) = − sin(θ(t))θ̇(t) = −s(t)w(t)
The new kinematic model is then obtained:
   
ẋ c 0  
 ẏ   s 0  v
q̇ =   = 
0 c  w
(9)
ṡ
ċ 0 −s
The new error states are defined as:
ex = c(xr − x) + s(yr − y)
ey = −s(xr − x) + c(yr − y)
(10)
es = sin(θr − θ) = sr c − cr s
ecos = cos(θr − θ) = cr c + sr s
After the differentiation of Eq. (10) and some manipula-
tions, the following system is obtained:
ėx = vr ecos − v + ey w
ėy = vr es − ex w
(11)
ės = wr ecos − ecos w
ėcos = −wr es + es w
Like in (6), v = vr ecos + vb and w = wr + wb will be used
in the control law. The control goal is to drive ex , ey , and
es to 0. The variable ecos is obtained as the cosine of the
IFAC SYROCO 2012
September 5-7, 2012. Dubrovnik, Croatia

Rt
error in the orientation and should be driven to 1. This is Lemma 1. (Barbălat’s lemma). If limt→∞ 0 f (τ )dτ ex-
why a new error will be defined as ec = ecos − 1 and the ists and is finite, and f (t) is a uniformly continuous func-
final error model of the system is now: tion, then limt→∞ f (t) = 0.
ėx = wr ey − vb + ey wb
ėy = −wr ex + vr es − ex wb Lemma 2. If f, f˙ ∈ L∞ and f ∈ Lp for some p ∈ [1, ∞),
(12) then f (t) → 0 as t → ∞.
ės = −ec wb − wb
ėc = es wb In Lemma 2 the Lp norm of a function f (t) is used. It is
defined as:
4. LYAPUNOV-BASED CONTROL DESIGN Z ∞ 1/p
p
kf kp = |f (τ )| dτ (20)
A controller that achieves asymptotic stability of the 0
error model (12) will be developed based on a Lyapunov
where |·| denotes the vector (scalar) length. If the above
approach. A very straight-forward idea would be to use a
integral exists (is finite), the function f (t) is said to
Lyapunov function of the type
belong to Lp . Limiting p towards infinity provides a very
k 2
1 2

V0 = e + e2y + e + e2c (13) important class of functions L∞ – bounded functions.
2 x 2 s Theorem 3. If the control law (18) is applied to the system
however, a slightly more complex function will be proposed
(12) where k is a positive constant, a > 2 is a constant, kx
here, which also includes the function (13) as a special case.
and ks are positive bounded functions, and the reference
The following Lyapunov-function candidate is proposed to
velocities vr and wr are bounded, then the tracking errors
achieve the control goal:
ex , es , and ec converge to 0. The convergence of ey to
k 2
1

V = ex + e2y + ec

e2s + e2c (14) 0 is guaranteed, provided that at least one of the two
2 2 1+ a conditions is met:
where k > 0 and a > 2 are constants. Note that the range (1) vr is uniformly continuous and does not go to 0 as
of the function ec = cos(θr − θ) − 1 is [−2, 0], and therefore t → ∞, while ks is uniformly continuous,
a−2 ec (2) wr is uniformly continuous and does not go to 0 as
0< ≤1+ ≤1
a a t → ∞, while vr , kx , and ks are uniformly continuous.
1 a (15)
1≤ ec ≤ Proof. It follows from (19) that V̇ ≤ 0, and therefore
1+ a a−2
the Lyapunov function is non-increasing. Consequently,
Due to (15) the function V in (14) is lower-bounded by
the following can be concluded:
the function V0 in (13). Since the latter is of class K, V
fulfills the conditions for the Lyapunov function. The role ex , ey , es , ec ∈ L∞ (21)
of (1 + eac ) will be explained later on. The function V can Based on (21), it follows from (18) that the control signals
be simplified by using the following: are bounded, and from (12) that the derivatives of the
e2s + e2c = e2s + (ecos − 1)2 = 2 − 2ecos = −2ec (16) errors are bounded:
Taking into account the equations of the error model (12) vb , wb , ėx , ėy , ės , ėc ∈ L∞ (22)
and (16), the derivative of V in (14) is: where we also took into account that vr , wr , k, kx , ks , and
1
2n
V̇ = −kex vb + kvr ey es +
(−2es wb )+ 1 + eac are bounded. It follows from Eqs. (21) and (22)
2 1 + eac that ex , ey , es , and ec are uniformly continuous (note that
! the easiest way to check the uniform continuity of f (t) on
− a1 es wb (−2ec ) wb
+
2 = −kex vb + es kvr ey −
2 [0, ∞) is to see if f, f˙ ∈ L∞ ).
2 1 + eac 1 + eac
(17) In order to show the asymptotic stability of the system,
let us first calculate the following integral:
In order to make V̇ negative semi-definite, the following
Z∞
control law is proposed:
V̇ dt = V (∞) − V (0) =
vb = kx ex  n
 ec 2 ec 2 0
(18)
wb = kvr ey 1 + + ks es 1 + Z∞ Z∞  n−1
a a ec 2
=− kkx e2x dt − ks e2s 1+ dt (23)
where kx (t) and ks (t) are positive functions, while n ∈ Z. a
For practical reasons n is a small number (usually −2, −1, 0 0
0, 1 or 2 are good choices). By taking into account the where the notation V (t) is used although V does not
control law (18), the function V̇ becomes: depend on t explicitly. Since V is a positive definite
 n−1
ec 2
function, the following inequality holds:
2 2
V̇ = −kkx ex − ks es 1 + (19) Z∞ Z∞  n−1
a 2 2 ec 2
V (0) ≥ kkx ex dt + ks es 1 + dt ≥
Two very well-known lemmas will be used in the proof of a a
0 0
theorem in this section. The first one is Barbălat’s lemma Z∞ Z∞
and the other one is a derivation of Barbălat’s lemma.
Both lemmas are taken from Ioannou and Sun (1996) and ≥ kk x e2x dt + ksL e2s dt (24)
are given below for the sake of completeness. 0 0

401
IFAC SYROCO 2012
September 5-7, 2012. Dubrovnik, Croatia

where the lower bounds of the functions kx (t), ks (t), and limt→∞ wb = 0 as shown before. Since ex and wb converge
 2(n−1) to 0 while kx and ey are bounded, the last two terms in
1 + eca(t) are introduced Eq. (29) go to 0 as t goes to infinity. Consequently, the
0 < k x ≤ kx (t) t ∈ R product wr ey also goes to 0. Since wr does not go to 0, ey
0 < k s ≤ ks (t) t ∈ R has to go to 0.
 2(n−1) (25)
ec (t) It is somehow more difficult to prove the convergence
0<L≤ 1+ t∈R of ec to 0. However, it has already been shown that ec
a
converges either to 0 or to −2, and it is easy to show
and two cases can be distinguished to determine the value
that ec will converge to 0 if the initial condition of the
of L due to (15):
 error vector is in the vicinity of the origin or the design

1 n≤1 parameter a is sufficiently close to 2. The error ec converges
 2(n−1) to 0 asymptotically if one of the following conditions is
L= a−2 (26)

 n>1 satisfied:
a
• V (0) ≤ 2, or
It follows from (24) that ex , es ∈ L2 . Applying Lemma 2, • V (0) > 2 and a < 2V (0)
V (0)−2 .
the convergence of ex (t) and es (t) to 0 follows immediately.
When es is 0, ec can be either 0 or −2 since es and To prove this statement, let us calculate the value of the
(ec + 1) are the sine and the cosine, respectively, of the T
Lyapunov function V in e−2 = [ 0 0 0 −2 ] :
same argument. Due to es → 0, it follows from (12) that 

ėc → 0 and consequently the limit limt→∞ ec (t) exists and 1

2 2a
V−2 = V |e=e−2 = e  = > 2 (30)
is either 0 or −2. 2 1 + eac c  a−2
ec =−2
Until now we only established the convergence of ex (t) and It is easy to show that V (0) < V−2 if one of the above two
es (t) to 0, while ec (t) was shown to converge either to 0 or conditions is satisfied. Among all the points that share the
to −2. To show the convergence of ey (t) to 0, at least one same ec = −2, e−2 is the point with the lowest V . Since
of the conditions 1 or 2 of Theorem 3 have to be fulfilled. V is a monotonically non-increasing function, the system
Let us first analyze the condition 1. Applying Lemma 1 to can never reach any point with ec = −2. Thus, it is only
ės (t) ensures that limt→∞ ės (t) = 0 if limt→∞ es (t) exists possible that ec converges to 0.
and is finite (which has already been proven) and ės (t) is
uniformly continuous. The latter is true (see (12)) if (ec + Now let us assume that ec is in the vicinity of −2. Upon
1)wb is uniformly continuous. It has already been shown inserting wb from Eq. (18), ėc in Eq. (12) becomes:
 ec 2  ec 2n
that ec is uniformly continuous. The feedback control for
ėc = es wb = kvr ey es 1 + + ks e2s 1 + (31)
the angular velocity wb defined in (18) is uniformly con- a a
tinuous since ks and vr are uniformly continuous from the The second term in Eq. (31) is always positive. The error
assumption in condition 1 of the theorem. The statement ec will increase and thus be repelled from ec = −2 if the
limt→∞ ės (t) = 0, which is identical to product vr ey es is positive. If this is not satisfied, then ec


lim − (ec (t) + 1) wb (t) = 0 (27) will still increase if |vr ey | is small enough (the second term
t→∞ is dominant in Eq. (31)):
has therefore been proven. Since (ec (t) + 1) converges to 1 ks  ec 2(n−1)
or to −1, the following can be concluded from (18): |vr ey | < |es | 1 + ⇒ ėc > 0 (32)

2 h
2 in  k a
lim wb = lim kvr ey 1 + eac + ks es 1 + eac =0 Even if this is not the case, the analysis in the vicinity of
t→∞ t→∞ ec = −2 will show that this equilibrium point is repelling.
(28)
The errors ex and es always converge to 0 and we can
The factor (1 + eac ) can converge either to 1 or to (1 − a2 ),
say that after some time they belong to O(ε) where ε is
but is always strictly positive and bounded. Since the same
sufficiently small. It is easy to conclude from (12) that the
holds for ks , while it was shown that es converges to 0, the
derivatives of the errors then become:
whole second term of wb in Eq. (28) also converges to 0. In
ėy = O(ε)
the first term k is a finite constant, (1 + eac )2 is bounded  ec 2 (33)
and strictly positive, while vr is also bounded and does not ės = (−ec − 1)kvr ey 1 + + O(ε)
diminish as t → ∞ due to the assumption in the condition a
1. Consequently, the convergence of ey to 0 follows. Now we are interested in the derivative of vr ey es :
d
For the second case (condition 2) the Barbălat’s lemma dt (vr ey es ) = v̇r ey es + vr ėy es + vr ey ės =


ec 2
(Lemma 1) is applied to ėx in Eq. (12) after inserting the = O(ε) + O(ε2 ) + (−ec − 1)kvr2 e2y 1 + + O(ε) a
control law for vb (18): (34)
ėx = wr ey − kx ex + ey wb (29) We can see that in the vicinity of ec = −2 the dominant
In order to show that limt→∞ ėx = 0 we again have to term in Eq. (34) is always positive and ėc in (31) also
guarantee that all the signals on the right-hand side of eventually becomes positive. Even if for a short time the
Eq. (29) are uniformly continuous. We have seen that ex error ec is approaching −2, it is always repelled after
and ey are uniformly continuous, wr and kx are uniformly that. Note that this happens when es and vr ey are of
continuous from the assumption of the condition 2, while the opposite sign (if they are of the same sign ec always
wb is uniformly continuous if ks and vr are uniformly increases). From (12) it can be concluded that es will then
continuous. The latter two conditions also guarantee that move in the direction of changing its sign. But it has to

402
IFAC SYROCO 2012
September 5-7, 2012. Dubrovnik, Croatia

cross es = 0 when ec becomes −2. This is why the error

ec actually crosses ec = −2 but it is immediately repelled 0.3 ref
law 1
from it. The solution ec = −2 is therefore an unstable 0.25 law 2
equilibrium point. law 3
0.2
Since ec cannot be attracted to −2 for a longer time, it 0.15
0.3
has to eventually be driven to 0 because this is the only ref
law 1
0.1 law 2
alternative. ✷

y
0.2
law 3

0.05 0.1

5. COMPARISON OF DIFFERENT CONTROL LAWS 0 0

y
−0.05 −0.1
The proposed method has been extensively tested and
−0.1 −0.2
compared to the existing methods from the literature. The
method shows promising tracking performance. Here it will 0 0.1 0.2
−0.3
0.3 −0.3
0.4
−0.2 −0.1 0 0.1 0.2 0.3 0.4
x
be compared to two similar and very well-known control x

laws. The first one is from Kanayama et al. (1990) and the
other one is inspired by Samson’s work (Samson, 1993).
These algorithms were chosen because they were both
designed using Lyapunov stability theory. The Lyapunov
functions used therein have the same limit around the zero
error in orientation (eθ = 0). The control law proposed by
Kanayama et al. (1990) is :
vb = kx ex
(35)
wb = kvr ey + ks sin eθ
Sometimes, there is also a third factor vr in the second
term of wb . Note that the control law (35) can be easily
developed if the most straight-forward Lyapunov function
V0 from (13) is used in the controller design procedure.
Note also that the control law (35) is a special case of (18)
when a limits to infinity.
The control law due to Samson is:
vb = kx ex
sin eθ (36)
wb = kvr ey + ks eθ
eθ
In our simulation experiments the orientation error eθ in
(36) was always mapped to the interval (−π, π] to make the
comparison fair. It is easy to see that all three control laws
given in Eqs. (18), (35), and (36) that will be compared
in the paper have the same limit around eθ = 0 and it is
therefore fair to use the same control gains – in our case
we shall use k = 10, kx = 10, and ks = 10.
The form of the control laws suggests that the role of vb is
always the reduction of ex , while wb needs to cater for the
remaining two errors. A very simplified explanation is that
the first term in wb takes care of the lateral error, while
the second term is responsible for the orientation error. In
reality, the problems are much more complicated, due to
the nonholonomic nature of the system. Nevertheless, we
can still be sure of certain facts:
• The feedback vb is the same in all three control laws
that have been compared, and therefore we cannot
expect any drastic differences in performance.
• The control laws (18) and (35) are continuous with
respect to the orientation error (at the orientation
error of ±π) while the control law (36) does not share
this property. This is a problem of the latter approach
since the implementation of the discontinuous control
law on a dynamic model becomes questionable.
• Let us denote the “gain” from ey to wb by gy (t) ,
|wb (t)|/|ey (t)||eθ (t)=0 . Additionally, the superscripts
p, k, and s denote the proposed control law (18),
Fig. 2. The reference trajectory and three robot trajecto-
ries (initial pose is shown by the arrow)
Kanayama’s control law (35), and Samson’s control
law (36), respectively. Among the three control laws
compared, gyk is always the highest. It is easy to show
that if a ≥ 6, gyp is always larger than gys . If 2 < a < 6,
gys > gyp when eθ is small, and gys < gyp when eθ is large.
• Let us denote the “gain” from eθ to wb by gθ (t) ,
|wb (t)|/|eθ (t)||ey (t)=0 . The superscripts are the same
as above. If n > 0, gθp < gθk < gθs . If n = 0,
gθp = gθk < gθs . If n is negative, then gθk is the lowest.
The comparison between Samson’s control law and
the proposed one depends on the choice of design
parameters. If a ≤ 6|n|, gθp < gθs . If a > 6|n|, the
comparison also depends on the orientation error:
gθp > gθs for small eθ and gθp < gθs for large eθ .
If we try to summarise the above analysis we can see that
for n > 0, the proposed control law always has the lowest
gain gθ , while either Samson’s law or the proposed one
have the lowest gy gain. This means that we can expect
lower control effort from these two laws, while Kanayama’s
law will exert more control action.
6. EXAMPLE
Some simulations were performed to test and compare
all the approaches under the same circumstances. The
reference trajectory is the same in all the studied cases:
xr (t) = b(a + R cos(5ω0 t)) cos(ω0 t)
(37)
yr (t) = b(a + R cos(5ω0 t)) sin(ω0 t)
with a = 1, R = 0.7, b = 0.2, and ω0 = 2π/40. The
simulation run always started at t = 0 and finished at
t = 40. The control signals vb and wb were saturated to
±10. Three algorithms were compared, the proposed one
given by Eq. (18) is shown in blue colour (the control law
design parameters are a = 30 and n = −1), the one given
by Eq. (36) in red colour, and the one given by Eq. (35)
in magenta colour. The resulting trajectories are shown in
Fig. 2, the errors in Fig. 3, and feedback velocities in Fig.
4.
Some additional studies where performed to try to get
some guidelines on choosing n and a. For low orientation
errors the high values of n are good, while for large initial
orientation errors the negative values of n are better. The
explanation is very simple. When n < 0 the second term in

403
IFAC SYROCO 2012
September 5-7, 2012. Dubrovnik, Croatia

0.1 integer), but this could lead to stability problems. These

aspects are, therefore, a topic of future research.
0
ex

−0.1 REFERENCES
0 1 2 3 4 5
0.2 Brockett, R.W. (1983). Differential Geometric Control
Theory, chapter Asymptotic stability and feedback sta-
0
ey

bilization, 181–191. Birkhauser, Boston, MA.

−0.2 Buccieri, D., Perritaz, D., Mullhaupt, P., Jiang, Z.P., and
0 1 2 3 4 5
Bonvin, D. (2009). Velocity-Scheduling Control for a
1
Unicycle Mobile Robot: Theory and Experiments. IEEE
0 Transactions on Robotics, 25(2), 451–458.
es

Fliess, M., Levine, J., Martin, P., and Rouchon, P. (1995).

−1
0 1 2 3 4 5 Flatness and defect of nonlinear-systems - introductory
0 theory and examples. International journal of control,
law 1 61(6), 1327–1361.
−1
ec

law 2
law 3
Guechi, E.H., Lauber, J., Dambrine, M., Klančar, G., and
−2 Blažič, S. (2010). PDC control design for non-holonomic
0 1 2 3 4 5
t wheeled mobile robots with delayed outputs. Journal of
Intelligent & Robotic Systems, 60(3), 395–414.
Fig. 3. Four error states Ioannou, P.A. and Sun, J. (1996). Robust Adaptive Con-
trol. Prentice-Hall.
0.5 Jiang, Z.P. and Nijmeijer, H. (1997). Tracking control of
mobile robots: A case study in backstepping. Automat-
0 ica, 33(7), 1393–1399.
vb

−0.5 Kanayama, Y., Kimura, Y., Miyazaki, F., and Noguchi,

T. (1990). A stable tracking control method for an
−1 autonomous mobile robot. In Proceedings 1990 IEEE
0 2 4 6 8 10
International Conference on Robotics and Automation,
5
law 1
volume 1, 384–389. Los Alamitos, CA, USA.
0 law 2 Kanayama, Y., Nilipour, A., and Lelm, C. (1988). A
wb

law 3 locomotion control method for autonomous vehicles. In

−5
−10
0 1 2 3 4 5
t
Fig. 4. Feedback signals vb and wb
wb is the dominant one. This means that the main control
goal is to reduce the error in the orientation, while the
lateral error is not so important. Such a strategy is useful
when the error in the orientation is high and it is necessary
to reduce it quite quickly (otherwise the error in ey can also
increase due to the interconnection). When, on the other
hand, the orientation error is low, it is more important
to cope with ey , which is a problematic error due to the
nonholonomic constraints. We achieve the emphasis on the
regulation of ey when n > 0.
7. CONCLUSION
In this paper a novel kinematic model is proposed where
the transformation between the robot posture and the sys-
tem state is bijective. A novel control law is also proposed.
It is designed within the Lyapunov stability framework. It
is proven that the global asymptotic stability of the system
is achieved under some very mild conditions if the reference
velocities satisfy the condition of persistent excitation.
The proposed approach shows good results. An obvious ex-
tension of the proposed control law is that the parameters
n and a could be scheduled according to the orientation
error. The possibility of on-line adapting these parameters
also exists (n could also be a real number and not an
Proceedings of the 1988 IEEE International Conference
on Robotics and Automation, volume 2, 1315–1317.
Washington, DC, USA.
Klančar, G., Matko, D., and Blažič, S. (2011). A control
strategy for platoons of differential drive wheeled mobile
robot. Robotics and Autonomous Systems, 59(2), 57–64.
Klančar, G. and Škrjanc, I. (2007). Tracking-error model-
based predictive control for mobile robots in real time.
Robotics and Autonomous Systems, 55(6), 460–469.
LaValle, S.M. (2006). Planning Algorithms. Cambridge
University Press, Cambridge, U.K.
Lizarraga, D. (2004). Obstructions to the existence of
universal stabilizers for smooth control systems. MCSS
Mathematics of Control, Signals and Systems, 16, 255–
277.
Morin, P. and Samson, C. (2009). Control of Nonholo-
nomic Mobile Robots Based on the Transverse Function
Approach. IEEE Transactions on robotics, 25(5), 1058–
1073.
Pourboghrat, F. (2002). Exponential stabilization of
nonholonomic mobile robots. Computers & Electrical
Engineering, 28(5), 349–359.
Pozna, C., Troester, F., Precup, R.E., Tar, J.K., and
Preitl, S. (2009). On the design of an obstacle avoiding
trajectory: Method and simulation. Mathematics and
Computers in Simulation, 79(7), 2211 – 2226.
Precup, R.E. and Hellendoorn, H. (2011). A survey on
industrial applications of fuzzy control. Computers in
Industry, 62(3), 213 – 226.
Samson, C. (1993). Time-varying feedback stabilization of
car like wheeled mobile robot. International Journal of
Robotics Research, 12(1), 5564.

404