A NEW METHOD FOR TRACKING MEMORIZED

PATHS: APPLICATION TO UNICYCLE ROBOTS

F. Diaz del Rio, G. Jimenez, J. L. Sevillano, C. Amaya, A. Civit Balcells.
Robotics and Computer Technology for Rehabilitation Laboratory (Universidad de Sevilla)
Address: Avda. Reina Mercedes s/n. 41012 Sevilla. SPAIN
fax: (+34) 954556821.
e-mail: fdiaz@atc.us.es
Keywords: nonholonomic constraints, unicycle
mobile robots, path following, trajectory tracking,
differential geometry.
Abstract
When following a memorized trajectory, there are
several possibilities attending to the way in which
the actual robot state can be related with the whole
trajectory. In this case several possibilities appear.
One of them is the extension of the servosystem
approach, usually called trajectory tracking. This
is the only possibility if we need strict temporal
deterministic requirements. But if not, we can
decide, for example, that the paths point to track is
the nearest (under several conditions) to the
actual robots position. This is frequently called
path following in mobile robot literature. In this
paper we present another method suitable for non-
deterministic systems, which we may call error
adaptive tracking, because the tracking pace
adapts to the errors. Its benefits are identified.
Mainly, it conserves most of the advantages of the
path following method, while it avoids its main
obstacle: the non-uniqueness of the nearest path
point. Afterwards, we determine how to construct
this method and we apply it to the case of a
unicycle mobile robot. Then a control law that
ensures asymptotic stability is extracted for the
unicycle robot, using the second Lyapunov method
and under the considerations of the error adaptive
tracking approach. Finally, we compare the
behavior of the new method to a similar control law
under the pure trajectory tracking approach.
1 Introduction
Many researchers have studied various tracking
methods when the desired trajectory or path is
memorized or previously generated. During the last
years and mainly in the field of mobile robots,
several methods have been proposed, and in a
general sense, two main tracking methods have
been used. In a first group we find those that
consider time explicitly in the tracking [1,14,15,11]
(usually called "trajectory tracking"), and try to
approach the robot to a moving point. In a second
one, we find those that do not have timing
requirements and try to converge to a path
[9,10,7,11,18,1,20,17] (usually called "path
following"). Moreover, we can find several
excellent compendia of both methods in some
reports or books [6,4]. For this analysis, the desired
path is usually parameterized [16] and the desired
point is selected to be the closest point on the
path to the actual robots position [13]. The term
closest means the paths point that makes certain
distance criterion minimum. This implies that the
robots posture is projected into the path in order to
select the desired point.
Trajectory tracking (TT) has been well studied
because it is similar to servosystems, and it is
guaranteed that the system will converge to the
desired trajectory in a deterministic time. On the
other hand, path following (PF) is not well suited
for systems with strict timing requirements but it is
very suitable for many mobile robots since they are
not usually involved in hard real time systems.
Although TT approach seems to be the most
straightforward, it has been shown that PF is more
suitable for many situations in which time is not a
critical parameter (see section II). The main
problem with PF is that projection uniqueness has
not been guarantied yet. Thus PF is not applicable
for all the possible paths.
In the middle of TT and PF, other tracking methods
can be designed. Encarnaao and Pascoal [13]
Proceedings of the 10th Mediterranean Conference
on Control and Automation - MED2002
Lisbon, Portugal, July 9-12, 2002.


combine both approaches through a weight
parameter. In the end, this means that timing
requirement are preserved, while they are forgotten
for the period where errors were large. This method
is not very extended and there has not been
demonstrated that projection uniqueness is
preserved under it, so we might have the same
problem as with PF.
For all the reasons explained before, we first
propose in this paper a new tracking policy that
tries to overcome the difficulties of the other
methods. We have named it error adaptive tracking
(EAT), because the tracking adapts to the system
errors. Also it can be considered something like a
relaxed trajectory tracking, because the design is
similar to that of TT but it is intended for path
recovering without strict timing requirements. We
will show that it retains the exposed advantages of
PF. Moreover the uniqueness of the paths point
selection is trivially satisfied, so EAT can be
applied to all sets of paths.
The need for this new tracking method comes to us
from the development of advanced wheelchairs.
Our research group has been interested during the
last years in the improvement of electrical
wheelchairs [9,5], which incorporate advanced
features. In this sense some features have been
introduced into classical wheelchairs, and playing
back previously recorded trajectories is considered
a very helpful aid. This avoids the user the difficult
maneuvering of reverse driving, and may be very
useful in small areas like bathrooms. In our group
we have developed SIRIUS, an advanced
wheelchair prototype that includes path recovery of
usual trajectories, detection and avoidance of
obstacles through simple sensors like sonars,
intelligent user interfaces with shared control, etc.
[5][9]. It is important to mention in SIRIUS we
must contemplate all the possible desired paths that
can be made by the user (usually driving his/her
joystick), including zero-radius turns. One of the
typical topologies for electric wheelchairs is the so-
called unicycle. They include driver motors at each
rear wheel, that can turn independently forward or
backward. Different speeds at each rear wheel
cause the turn of the chair.
Once the tracking method has been selected, a
convergence law must be found. As it is well
known, mobile robots are intrinsically nonlinear
systems, because of their kinematic model.
Moreover they have more state coordinates than
degrees of freedom (DOF), because of their
nonholomic constraints (except for the special case
of omnidirectional robots). Due to this,
convergence to a path may acquire a special
treatment. Besides the most frequent contributions
are those based in the Lyapunov direct method,
which we will also use it in this paper. Finally, as
we are interested in convergence to a path, we will
suppose in this work, that the desired trajectory has
no end. We do not allow the trajectory to end
because convergence to a fixed point can not be
achieved through a smooth feedback stabilization
control law (a direct result of Brocketts theorem
[3]). Due to this fact, the point stabilization
problem is very different to the paths convergence
problem, which is considered here.
In the next sections we will try to analyze and
expose the new error adaptive tracking. In section 2
we formulate the different tracking methods
according to the paths descriptor parameter
selection. In section 3 we define our robot model.
We construct the new tracking technique and
propose an asymptotically stable control law in
section 4, which is evaluated through several
simulated and real examples in section 5. Finally in
section 6 we summarize the conclusions.
2. Tracking methods formulation
In mobile robots and other robotic applications it is
usual that the reference trajectory is memorized. A
memorized, reference or desired path (or merely
path) can be described by a single descriptor
parameter, namely r, and it can be expressed as a
vector of state coordinates q
des
(r).
As a result, tracking progress can be identified with
the progress of r. Although the parameter may be
time, in the case where time dependence is not
relevant, many others are possible. For example, in
differential geometry the natural arc parameter
[16], which makes the linear speed equal to 1, is
generally preferred. Independently of the selected
parameter we can classify the tracking of the path
according to the way in which we impose or design
(when possible) the progress of parameter r.


In servosystems we track a mobile system or target
at the time it moves. In figure Fig. 1a we show this
case. The state coordinates are called q(t), the
desired coordinates q
des
(t), and the error
coordinates are defined as e
q
(t)=q(t)-q
des
(t). In
servosystems where time is critical, this is the only
possibility we have. Nonetheless, when we try to
track a memorized path, the tracking methodology
can be designed openly, as we know a priori the
whole trajectory. Of course the classical
servosystem tracking can be done just by
identifying the parameter r associated to the path
with time, that is r(t)=t. This is usually called
trajectory tracking (TT) in mobile robots.
Furthermore, TT can be extended in a more general
assumption than simple servosystems (see Fig. 1b):
let the parameter r be a generic function of time
r=r(t). Therefore the error coordinates are
e
q
(t)=q(t)-q
des
(r(t)). Then we can go through the
stored trajectory with the most appropriate scale for
r(t), for example r=at. Independently of the form of
the function r(t) the tracking process is similar,
because (using an asymptotically stable control
law) it is guaranteed that the system will converge
to a point in the desired path q
des
(r) in a
deterministic time, except for the inherent
perturbations that it may suffer.
The alternative best established in literature and
very suitable when time is not a critical parameter
is path following (PF). This is based in some
relation between actual systems state q(t) and the
whole memorized path. This relation or projection
will give us the desired point q
des
(r), i.e. the
descriptor parameter r as a function of the actual
position and the path: r=(q), where is a some
kind of projection of actual position to the path.
Then the real system must try to follow this point
instead of the one given by the other approach (see
Fig. 1c). The error coordinates are also e
q
(t)=q(t)-
q
des
(r). Of course using this approach, it is not
guaranteed that the system will reach a point of the
desired trajectory in a deterministic time. As we
detailed at the Introduction, the main problem with
PF is that projection uniqueness has not been
guarantied yet.
It has been shown that PF is more suitable for many
situations (in which time is not a critical parameter
of course). This can be understood if we consider
the following example: if perturbations force the
system to be at rest, the desired point for TT will
move unavoidably. This means that errors will
grow up to some value that may introduce
instability. On the other hand, if PF were used, the
desired point will be the same in spite of these
perturbations, because the paths shape and the real
robot state remain the same. This allows the PF
system to overcome possible unstable states.
Memorized
path
Memorized
path
e
q
(t)
Memorized
trajectory
q(t) q(t)
Actual reference
trajectory
r=(q)
e
q
(t)
e
q
(t)
q
des
(t)
Fig. 1a: Tracking in a servosystem
e
q
(t)
q
des
(r)
Fig. 1b: Trajectory Tracking
r=at; a=scale
q(t)
q
des
(r)
Fig. 1c: Path Following
r& =g(t,e
q
)
q(t)
q
des
(r)
Fig. 1d: Error adaptive
tracking

Fig. 1: Tracking methods regarding to descriptor
parameter.
Memorized
path
e
q
(t)
e
q
(t)
q
des
(r
PF
)
q
des
(r)
q
des
(t)
r tends to r
PF
robot is far from
the path
q(t)
q
des
(r
PF
)
q
des
(t)
q
des
(r)
r tends to t
Memorized
path
q(t)
robot is close to
the path

Fig. 2: Variation of descriptor parameter in combined
trajectory tracking and path following
In the middle of TT and PF, the proposal by
Encarnaao and Pascoal [13] has been called
combined trajectory tracking and path following.
This appellation comes from the combination of
both approaches (through a weight parameter).
Their aim is very desirable: it is to gain the benefits
of PF when errors are large but preserving timing
requirements. If the perturbations make the errors
very large, the desired point tends to be that chosen
by PF (that is, r tends to r
PF
). When the robot


recuperates and approaches to the path, the
desired point approaches that of selected by TT
(that is, r tends to t), as shown in Fig. 2. However,
as we see at the Introduction, we might have the
same problem as with PF.
Having in mind the difficulties and the goals
explained before, we propose another possibility,
which we have called error adaptive tracking (EAT).
Note that in a pure TT the parameter r is
determined exclusively in function of the time t.
That is, desired point selection does not
contemplate actual robots posture, because the
intention is to have a deterministic tracking. But
when the time is not so critical, we can design
the variation of r, instead of stating the parameter r
itself. And we can regard robot state, i.e. taking the
errors into account. Thus, contrasting the TTs rigid
variation of r ( r& =1 or in general r& =f(t)), we
propose r& =g(e
q
) (see Fig. 1d), where g(e
q
) is a
convenient function of the errors. Here
convenient is referred to the designer objectives,
but also to ensure that tracking is done correctly
and that the previously explained advantages of PF
are preserved. That is, the function g(e
q
) should
fulfil (see Fig. 3):
If errors are small, g(e
q
) should tend to 1, so the
tracking resembles TT, and near deterministic
following is done. The real and the reference
robot advance at the same pace.
If errors are large, the reference robot should
wait for the actual robot. That is, g(e
q
) should
be small until a good convergence is reached.
Here the tracking rate must not be far from that
of PF. In fact when errors are large in PF, we
expected that the projecting point on the desired
path varies very little, that is r& tends to zero.
Of course in this situation, no deterministic
following is expected.
The above conditions imply that the tracking will
not preserve time determinism. Otherwise, if our
design needs some aspects of time determinism, the
variation of r can be extended to include the
inaccuracy in the deterministic tracking. That is,
the difference between the descriptor parameter r
(that indicates the target of our control design
q
des
(r)) and the time t (that indicates the target of
the timing requirements q
des
(t)). To sum up, a more
general function for r& would be g(t,e
q
).
r& =g(t,e
q
) tends to 1
q
des
(r)
e
q
(t) are small
Memorized
path
r& =g(t,e
q
) should be small
e
q
(t) are
large
Memorized
path
q(t)
q
des
(r)
q(t)

Fig. 3: Rules on the variation of descriptor parameter
Of course many possible functions g(t,e
q
) can be
designed attending to the characteristics and
purposes of each system but, in this work, we
would explore mainly two:
g(t,e
q
) is only a function of errors e
q
and
|g(e
q
)|1e
q
. With these conditions the tracking
can not be deterministic and reminds that of PF.
g(t,e
q
) is a combined function of errors e
q
and
the difference between parameter r and time t.
This strategy is intended to get a deterministic
tracking at the end, although if errors are large,
determinism is not applied for a while. The
behavior of this tracking reminds that of the
combined trajectory tracking and path
following [13]. Note that this method can be
applied to a servosystem if the reference
trajectory is collected as the target moves
(except that r will never be bigger than t).
The formulation advantage of EAT is that it can be
applied to all sets of paths.
3 Definitions and robot model
Lets consider the mobile robot shown in Fig. 4
(whose dimensions are those of SIRIUS) and let
q=(X, Y, )
t
be its state coordinates, which
represent the coordinates (X, Y) of a certain point
P
o
(typically the midpoint between rear wheels) in
the basis of the fixed frame ()=(O; i, j) and the
orientation of the robot with respect to the fixed
frame. Other variables that characterize internal
states, such as the angles turned by the wheels or
the relative orientation of castor wheels, do not
represent any useful state for the tracking problem.
The unicycle robot has three state variables but
only two degrees of freedom, as a result of the
nonholonomic constraint. We assume that the


wheels are nondeformable and that they are moving
on a horizontal plane without slip in order to hold
the constraint. Let u=(v, )
t
be the pair of input
variables which are the linear velocity of point P
o

and the angular velocity of the robot, respectively.
For these vector variables the kinematic model of
the unicycle robot can be expressed by the
equations (that are non-linear in q and linear in u):
|
|
|
.
|

\
|
=
|
|
.
|

\
|
=
(
(
(

= =

Y
X
v
q u B u q B q ; ;
1 0
0 sin
0 cos
; ) ( & (1)
A reference or desired path to be followed can be
described by a single parameter, namely r, and it
can be expressed as a vector of desired state
coordinates q
des
(r)=(X
des
(r), Y
des
(r),
des
(r))
t
. To
study the tracking of a path q
des
(r) let us define
another intrinsic coordinate system ()=(q
des
(r); t,
n) linked to the path (usually called the Frenet
frame [16], see Fig. 4). t is the unitary vector
parallel to the robot orientation in the desired point
q
des
and n the normal to it. Let (e
x
, e
y
) be the
position errors of point P
o
relative to these axis and
e

the robot orientation error, so e
q
=(e
x
, e
y
, e

)
t
will
be the relative errors vector (an analogous
coordinate system was used by Kanayama et al
[14], but in that case the system was linked to the
robot itself). Let u
des
(r)=(v
des
(r),
des
(r))
t
be the
desired inputs expressed as a function of the
descriptor parameter r. Besides, applying the chain
law for the desired inputs, we can get to:
r r r r
dt
dr
dr
d
dt
d
t
des des
des des
des
& & ) ( ) ( ) (

= = = =
where (') means differentiation with respect to r.
The relation for v
des
is analogous if we consider the
length s
des
of the path (
des des
s t v & = ) ( ). To sum up we
can declare that r r t & ) ( ) (
des des
u u = .
For the chosen frame, error coordinates e
q
can be
obtained as a rotation around an axis normal to the
XOY-plane. In fact and according to Fig. 4:
( ) ( )
( ) ( )
|
|
|
.
|

\
|
= =
1 0 0
0 cos sin
0 sin cos
) ( ; ) )( (
des des
des des
des des


R q q R e
des q

Therefore, using the relative coordinates linked to
the path, and by simple calculations, the following
state equations can be found [8]:
|
|
.
|

\
|
|
|
|
.
|

\
|
+
|
|
|
.
|

\
|
|
|
|
.
|

\
|
+
|
|
|
.
|

\
|

=
|
|
|
.
|

\
|

v
e
e
e
e
e v
e
e
e
y
x
des
des
des
des
y
x
1 0
0 ) ( sin
0 ) cos(
0 0 0
0 0
0 0
0
&
&
&
(3)
O
Y
X
2R
2d
P
l
Po
t
n
r
v
des

des
()

()
SIRIUS dimensions:
d=0.25m. R=0.16m

des
e

=-
des
e
x
e
y
v

Fig. 4: Extrinsic and intrinsic robot coordinates.
This form agrees with the intuition that error
variables must grow with both real and desired
postures advancement.
4 Error adaptive tracking selection and
Lyapunov-based control law
In section II we proposed the error adaptive tracking
based on the election of r& as a function g(t,e
q
). Let
us first analyze the most elementary case, where g
has not dependence on t. We have shown that if
errors are small, then g(e
q
) should tend to 1, but if
they were large, then g(e
q
) would be small.
The conditions stated above give us a first simple
idea for g(e
q
): a function bounded by 0 and 1 e
q
,
e. g. g(e
q
)=exp(-|e
q
|). Therefore when any error
grows, g tends to zero. Although this elementary
construction may be useful for some systems, the
nonholonomic constraint of the unicycle prohibits it
(if we want to look for a Lyapunov-based control
law). For this kind of mobile robots, it can be seen
that if a Lyapunov function V is the sum of positive
definite functions of the relative errors e
q
=(e
x
, e
y
,
e

), then the derivative of V cannot be negative

definite, but only a negative semidefinite function.
This is a direct consequence of the nonholonomic
constraint: if (e
x
, e

) are null at a time, then

y
e& must
also be null because of equation (3). Therefore e
y

cannot decrease at this moment, and hence V does
not too. Therefore we can not impose V
&
to be
negative definite, but only negative semidefinite. In


other words, if the robot is very far from the path
and e
x
, e

are null at a time, then the previous

g(e
q
)=exp(-|e
q
|) will tend to zero. In this extreme
case, a Lyapunov-based control law can not be
found, because any robot movement (satisfying the
constraint) will increase e
x
or e

, while e
y
remains
constant. Even going further, in this situation we
can observe that when r& becomes lower, the
convergence will be slower. This is due to the fact
that the input v can not exceed v
des
(t)=r& v
des
(r) (to
avoid e
x
being increased). To sum up, when V
&
is
zero, the decrement of V is "connected" or
"attached" to the increment of r& . Therefore if V
&
is
null, r& should be maximum, to get a fast
convergence. Moreover if the control law will
pursuit V
&
to be the closest to -KV, K>0, then
g(e
q
)=exp(K
V
V
&
) can be a plausible solution (K
V
>0
is a scale factor). With closest we mean that V
&

can not be -KV because of the constraint, although
it was desirable to get exponential convergence.
Considering this, we can design a control law based
on the following theorem.
Theorem: For unicycle robots and the Lyapunov
function:
( )
V e e A e
x y
= + +
1
2
1
2 2 2
( ) cos( )

, then
e
q
=(e
x
,e
y
,e

)=0 is an asymptotically stable

equilibrium point, and ) ( ) ( r r t
des des
u u u & = , if:
u
des
(r) 0.
r& =g(e
q
)=exp(K
V
V
&
), K
V
>0
The following control is imposed:
e
v
=-
1
-1
a
1
(e
q
),
1
>0
e

=-
2
-1
A

-2
a
2
(e
q
),
2
>0
where a
1
(e
q
)= e
x
cos(e

)+ e
y
sin(e

);
a
2
(e
q
)= A

2
sin(e

), and v = e
v
+ v
des
(t)cos(e

)
) cos( ) (
1
) ( ) (
1
) (
2 2

e t v e
A
e sin t v e
A
t e
des y des x
des
+
+ + =

Note that v and are composed by feedforward
terms and e
v
, e

are correction terms.

Proof: Differentiating V and using state equations:
( )
( ) ) sen( ) (
) sen( )) cos( ) sen( (
2
2

e A e v
e A e e e e v V
des x des
x y
+ +
+ + + =
&

Substituting v and :
V
&
= e
v
a
1
(e
q
)+ e

a
2
(e
q
)
And applying the control law:
V
&
= -
1
-1
a
1
(e
q
)
2
+ -
2
-1
A

-2
a
2
(e
q
)
2

Then V is non-increasing, so 0 V
&
, and V
converges to some limit: 0
lim
V V . Therefore by
Barbalats lemma [19] 0 , 0

e e & . At this limit,
V
.
is negative definite on error e
x
, so 0 , 0
x x
e e & ,
e
y
e
y,lim
<, and the control law is:
v= e
v
+ v
des
(t)


) (
1
) (
, 2
t v e
A
t e
des lim y des

+ =
and e
v
, e

0
,

And the state equations tend to:
|
|
|
.
|

\
|
=
|
|
|
.
|

\
|

+ +
0
0
0
) (
0
) ( ) (
,
t
v e t t v
des
lim y des des

Therefore by the third state equation we have

des
(t), and by the control law: vv
des
(t). Then:
0=e
y,lim
v
des
(t);
0=e
y,lim

des
(t);
Squaring and adding previous equations we have:
( ) ) ( ) ( 0
2 2 2
,
t v t e
des des lim y
+ =
At this limit r& = 1, so ( ) ) ( ) ( 0
2 2 2
,
r v r e
des des lim y
+ = ,
that implies (by the first hypothesis) that e
y,lim
=0.
QED
Remark 1: Note that if g(e
q
) tends to zero in the
case 0 , 0
x x
e e & , e
y
e
y,lim
then u also tends to
zero and convergence will be very slow.
Remark 2: Actually there is not special exigencies
on g(e
q
); only it is convenient to tend to one in the
case 0 V
&
and to zero when V
&
. In fact we
will use g(e
q
)=exp(-K
V
V
&
), K
V
>0 in the next
section to achieve an adequate convergence.
The results of this proposal will be shown in the
next section. Now we will study the case where g
has also dependence on t. As mentioned before, we
concentrate on a function g(t,e
q
) that is a combined


function of the errors and the difference between
the parameter r and the time t (i.e. t-r), in order to
get a relaxed deterministic tracking (determinism
is not applied when errors are large). The
dependence of g on errors could be the same as
before, because here previous reasons apply also
when r=t. The intention for introducing the
dependence on t is to permit that r remains at rest
when the robot is far from the path (in spite of the
increase of r-t). In a second case, when robot
recuperates and approaches the path, the
difference r-t should be reduced. Therefore g(t,e
q
)
can not be bounded by 1, because in the second
case r& must approach t. Moreover at the origin (t-
r=0, e
q
=0), g must be 1. Thus we might think of a
first proposal for this more complex EAT, e.g.:
g(t,e
q
)=exp(V
&
) (1+K
tr
arctan(t-r)) (4)
where K
tr
>0 is a scale factor that indicates how fast
the convergence of r to t is. Here g(t,e
q
) is upper
bounded by
2
1

tr
K + ,and lower bounded by
2
1

tr
K .
We want to point the fact that the second factor of
g(t,e
q
) is just a control law for parameter r. The
whole function g can even be seen as a new DOF
of the system since we are designing the behavior
of r, which can be considered a new state variable.
Note that in this law, the desired evolution for the
parameter is r=t, hence the constant 1 is a
feedforward term
1
. Moreover, when the difference
t-r becomes small, K
tr
arctan(t-r) behaves as a
proportional control law. We simply choose the
arctan because it is a bounded function, and large
inputs on the new DOF would mean large robot
inputs (v and ), and because such simple control
law will be sufficient for a state variable r that
has not any real physical model. We assume that
the previous choice can not be optimal at all for
some applications and more sophisticated g can be
tried in future work, but this simple alternative is
sufficient for SIRIUS wheelchair as a first
proposal.
In addition, the proposed function g(t,e
q
) fulfils
remark 2 of the previous theorem: naturally it tends
to zero when V
&
, and in the case 0 V
&
it

1
In general if the tracking would need an evolution like
r=f(t), the feedforward term would be r& = ) (t f
&
.
does not tend to one if and only if r is much greater
than t (i.e. the robot is in advance with respect to
time). But in this last case, the robot will simply be
quiet until the time has approached r, and at that
moment the convergence will be rapid (however
note that this situation is not very usual). Finally if
K
tr
>2/, the variation of r can be null or even
negative when robot is in advance with respect to
the desired point. This situation does not make
sense when a wheelchair is required to track a path
and it will be avoided in this paper (we will choose
exactly K
tr
=2/, so the chair will wait for q
des
(t) if
necessary, i.e. r& =0).
5 Simulation results.
In this section we show the behavior of our system
in several situations. For our real system SIRIUS,
the inertia load driven by the motors is important
even when gearbox ratio is high (31:1). Due to this
it is very important to consider the case where the
motors have a response delay (i.e., the real
kinematic inputs are delayed with respect to the
commanded inputs). In spite of this, we will show
that the EAT method ensures that errors will not
grow too much when motor response is slow.
Even when asymptotic convergence is ensured,
simulation is always a good way to verify and
observe the control behavior. This behavior is
primarily significant when errors are large,
because, if they were small, simulated systems
behavior will be similar to that of an exponential
convergent system.
In this subsection, we first analyze a path
composed of three very different segments: a
straight line, a 1.5m constant radius arc (which are
the most frequent cases found in current literature)
and a zero-radius turn. The first two segments have
a length of 5m in the state coordinates space (X, Y,
A

; A

2
=0.25m
2
) and the last one is non-ended.
The generality of the EAT method is confirmed;
moreover, we present three different initial
conditions (with considerable initial errors), so the
robot has to cope with each segment. We
summarize the initial position and errors of the
three cases in the following table. The duration is
10 seconds for each experiment.


Case e
x
(m) e
y
(m) e

(rad) R(m)
1 0 -0.4 -+0.5 0
2 0 0.7 /2 4.5
3 -0.3 -0.3 -/2 10
The values for constant parameters have been tuned
to ensure a smooth convergence and achieve a
satisfactory variation of r when errors are large:
K
V
=5s
1/2
m
-1
,
1
=0.5s,
2
=1/6s.
0 1 2 3 4 5 6 7
-0.5
0
0.5
1
1.5
2
2.5
3
3.5
Y(m)
X(m)
Desired path
Real trajectories

Fig. 5: EAT behavior under different initial conditions.
Obviously when errors are small the convergence
to the path is smooth and fast. However, in spite of
the presence of extreme conditions, good
convergence of EAT method is observed. In Fig. 5
the three different initial conditions are examined;
each one begins near each segment. The
convergence of the three errors to zero is similar
for all the cases.
The solid line gives us the real robot trajectory, and
the arrows show the initial robot orientation . Note
that in the second experiment the robot begins
following the path (at the first transient) in reverse
direction, in order to reduce errors faster, while
parameter r remains nearly constant. In Fig. 6 we
can see how r& evolves for each case: r& begins
with a value near to zero, and then it recovers
until it reaches the stationary value of 1. The
moments where r& approaches 1 rapidly (around the
first second) coincides approximately with the
situation when the robot is parallel to the desired
posture (e

=0), because in such situations, V

&
is
near zero (only lateral error e
x
contributes to V
&
).
Of course the differences between r& and 1 make r
less than t. This will be avoided whit g(t,e
q
).
Otherwise, when the errors are almost zero, the real
angular and linear velocities are exactly those of
the desired path, regardless of the curvature
discontinuities that the desired trajectory has (at
every segment change). In spite of this curvature
change, the errors do not increase appreciably (as
we can see at X=5m in Fig. 5). But this situation is
naturally not real. The above commented high
inertia of SIRIUS motors will not permit
instantaneous speed changes of its wheels. In fact if
the user makes a path composed of various
segments (of course stopping between them) and
tries to play it back again, this curvature
discontinuity dilemma can appear. We then must
go further in the simulations and emulate the inertia
of the motors, so the real speeds change always
smoothly. Although the exact tracking can not be
achieved (without stopping the robot) for this real
situation, our aim is to demonstrate that EAT
behaves better than TT in these changes.
case 3 case 1
case 2
r&
(m/s)
0 2 4 6 8 10
0
0.1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
t(s)

Fig. 6: Variation of r for the three experiments.
In the following experiment we emulate the motors
inertia like a first order delay, with the equation:
dt
dv
= (v
cont
v
real
);
where >0 is the time constant, v
cont
the velocity
demanded by the control and v
real
the real one
(analogous for the angular velocity). With a time
constant of 1s, we repeat the tracking for 20s,
supposing the robot begins at point r=0, e
q
=0. The
results for lateral error e
y
and the Lyapunov
function V are represented in Fig. 7. Note that the
transients are produced later for EAT (at seconds 5,
10 for TT, but at 7.2, 16.3 for EAT; marked by
vertical lines). The lateral error is considerably
smaller for the new EAT method, and V (which
gives us an idea of the whole error) is also inferior.
The main contribution in V is e
x
, which increase is
obviously reduced in EAT for the reduction of r& .


0 5 10 15 20
-0.5
-0.4
-0.3
-0.2
-0.1
0
0.1
0.2
V
TT

e
y,TT

e
y
(m)
V(m
2
)
t(s)
V
EAT

e
y,EAT

Trajectory tracking
Error adaptive tracking

Fig. 7: Tracking with delayed response.
The same comparison can be done for large initial
errors. In Fig. 8, the experiment is prolonged for 8
seconds (constant time=0.5s). The robot starts at
r=0, e
q
=(0,0.3m,/2+/4 rad). TT method deviates
the system very faraway from the path, because the
desired posture advances continuously. On the
other hand EAT method obliges q
des
(r) to "wait" for
the robot until the errors have reduced sufficiently,
so the deviation is not so pronounced. Moreover
and due to its large errors, the real speeds
demanded by TT are bigger than those of EAT, and
bigger than usual for wheelchairs (2m/s). It is
obvious that speed limitation will degrade even
more TT response. In the end the TT method
introduces more oscillations than EAT, and it has a
transient response that separates the robot from the
desired line. This is a well-known advantage of PF
[4,18], that the new EAT retains.
-2 0 2 4 6 8
-0.2
0
0.2
Trajectory tracking
Error adaptive tracking
EAT with g(t,eq)
(m)
X(m)

Fig. 8: Effect of delayed response under large errors.
The circles indicate the posture reached after 8 seconds.
Although previous simulations are not a rigorous
demonstration that EAT is more robust than TT
against perturbations or unmodeled dynamics, we
think that there is no doubt that the adaptive
variation of r facilitates robustness. Furthermore it
is important to observe that the qualitative behavior
of EAT is similar to that of PF, i.e. r& reduces in
presence of large errors until system approaches the
path. Of course both methods are constructed in a
very different way, so it is not easy to make a
quantitative comparison.
Finally we present some results for the EAT that
include time in r& ; exactly:
g(t, e
q
)=exp(-K
V
V
&
) (1+K
tr
arctan(t-r)) (5)
whit K
V
=5s
1/2
m
-1
, K
tr
=2/ m/s. The first experiment
was depicted in previous Fig. 8. The pointed curve
represents the tracking done by EAT method with
function (5). Note that, while tracking is similar to
that of previous EAT (same time and control
constants are used), it can advance much more; in
fact it has almost reached desired point q
des
(t)
without the deviation of the pure TT.
The good qualities of this EAT are highlighted
when the robot has reached the path (i.e it is on
some q
des
(r)) but it is not upon the point q
des
(t).
Then applying an asymptotically convergent
control law like the previous one, the input u is a
scale of desired input u
des
. In effect note that in this
situation e
q
=q(t)-q
des
(r) is null and a
1
(e
q
), a
2
(e
q
) too.
Hence only the feedforward terms of v and are
not null. On the contrary, if a pure TT were
applied, the inputs drive the robot to the point
q
des
(t), which may be out of the path.
6 Conclusions.
We have presented a new method for tracking
memorized paths, which we have named error
adaptive tracking (EAT) and we have evaluated it
for unicycle mobile robots. It is based on the design
of the variation of paths descriptor parameter, so
that the tracking adapts to the system errors. Based
upon this EAT methodology, a control law that
ensures asymptotic stability is proposed for
unicycle robots.
Its benefits are identified and demonstrated through
several simulations:
It conserves most of the advantages of the path
following method, while it avoids its main
obstacle: the non-uniqueness of the selected
paths point.
It is valid for all the possible trajectories.


Its behavior and robustness under large errors
or delayed response is much better than that of
a pure trajectory tracking. Overall, the
simulations presented here reveal that EAT
facilitates robustness.
A variety of EAT that includes time in the
variation of paths parameter to preserve
deterministic (non-strict) tracking, is also
presented. It also behaves better than pure TT
and similarly to previous EAT. Future research
will investigate this method more deeply.
We hope that EAT could be implemented
successfully in a computerized wheelchair that play
back paths done by the user.
Acknowledgements
This work has been supported by the Spanish
Ministerio de Ciencia y Tecnologa through
project TIC-2000-0087-P4-03.
References
1. Altafini. C., A path-tracking criterion for an
LHD articulated vehicle. The Int. J. of Robotics
Research. 18 (5), pp 435-441. May 1999.
2. Bloch, M., M. Reyhanoglu and N. H.
McClamroch. Control and stabilization of
nonholonomic dynamic systems. IEEE Trans.
on Aut. Control, 37 (11), pp 1746-1757, 1992.
3. Brockett, R. W. "Asymptotic stability and
feedback stabilization". Differential Geometric
Control Theory, Birkhauser, pp. 181-208, 1983.
4. Canudas de Wit, C., Khennouf, C. Samson and
O. J. Sordalen. Nonlinear control design for
mobile robots. In Recent trends in mobile
robots. Edited by Y. F. Zheng. World Scientific
Series in Robotics and Autom. Systems. 1993.
5. Civit Balcells, A., F. Daz del Ro, G. Jimnez,
J. L. Sevillano. A proposal for a low cost
advanced wheelchair architecture. The 4th
AAATE Conf. Oct 1997. Thessaloniki. Greece.
6. De Luca, A., G. Oriolo, and C. Samson.
"Feedback control of a nonholonomic car-like
robot". Robot Motion Plannning and Control,
pages 171-253. Springer-Verlag, Berlin, 1998.
7. DeSantis. R. M. Path-tracking for a tractor-
trailer-like robot. The International Journal Of
Robotics Research. Vol 13. No.6 Dec. 1994.
8. Daz del Ro, F. Analysis and evaluation of
mobile robot control : application to electric
wheelchairs (in Spanish). Ph. D. Thesis.
University of Seville, (Spain), 1997.
9. Daz del Ro, F., A. Civit Balcells, G. Jimnez,
J. L. Sevillano. Path tracking in the SIRIUS
wheelchair. Same as 5.
10. Daz del Ro, F., G. Jimnez, J. L. Sevillano, S.
Vicente, A. Civit Balcells. A generalization of
path following for mobile robots. Proc. of the
IEEE Int. Conf. on Robotics and Automation,
ICRA99. Detroit. May 1999.
11. Daz del Ro, F., G. Jimnez, J. L. Sevillano, S.
Vicente, A. Civit Balcells. A path following
control for Unicycle robots. Journal of Robotic
Systems 18(7), 325-342 (2001).
12. Doh-Hyun Kim, Jun-Ho Oh. Nonlinear tracking
control of trailer systems using the Lyapunov
direct method. J. of Rob. Syst. 16(1), 1-8. 1999.
13. Encarnaao P., A. Pascoal. Combined
trajectory tracking and path following for
marine craft. Proceedings of the MED01.
Dubrovnik, Croatia. June 2001.
14. Kanayama et al. A stable tracking control
method for an autonomous mobile robot. Proc.
1990 IEEE Int. Conf. Robotics and Autom,
Cinccinatti, Ohio, pp. 384-389.
15. Lamiraux, F., S. Sekhavat, J.P. Laumond,
Motion planning and control for Hilare pulling
a trailer, IEEE Trans. on Robotics and
Automation, Vol 15 No 4. August 1999.
16. Manfredo do Carmo, P. Differential geometry
of curves and surfaces. Prentice-Hall. 1976.
17. Samson, C. Control of chained systems:
application to path following and time-varying
point-stabilization of mobile robots, IEEE
Trans. on Autom.Control, 40 (1995), pp. 64-77.
18. Sarkar, N., X. Yun, V. Kumar. Control of
mechanical systems with rolling constraints:
application to dynamic control of mobile
robots. The Int. J. Rob. Research. 13(1) 1994.
19. Sastry, S. Nonlinear systems: analysis, stability
and Control. Springer-Verlag, New York. 1999.
20. Sordalen, O.J. and C. Canudas de Wit.
Exponential control law for a mobile robot:
extension to path following. IEEE Trans. on
Robotics And Autom., V. 9, N 6, Dec. 1993.