IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO.
2, APRIL 1999
Development and Experimental Validation
of an Adaptive Extended Kalman Filter
for the Localization of Mobile Robots
Leopoldo Jetto, Sauro Longhi, Member, IEEE, and Giuseppe Venturini
Abstract—A basic requirement for an autonomous mobile robot
is its capability to elaborate the sensor measures to localize itself
with respect to a coordinate system. To this purpose, the data
provided by odometric and sonar sensors are here fused together
by means of an extended Kalman filter. The performance of the
filter is improved by an on line adjustment of the input and
measurement noise covariances obtained by a suitably defined
estimation algorithm.
Index Terms— Adaptive filtering, localization systems, sensor
fusion, wheeled mobile robots.
I. INTRODUCTION
A N ACCURATE determination of location is a funda-
mental requirement when dealing with control problems
of mobile robots. Two different kinds of localization exist:
relative and absolute. The first one is realized through the
measures provided by sensors measuring the dynamics of
variables internal to the vehicle. Typical internal sensors are
optical incremental encoders which are fixed to the axis of the
driving wheels or to the steering axis of the vehicle. At each
sampling instant the position is estimated on the basis of the
encoder increments along the sampling interval. A drawback of
this method is that the errors of each measure are summed up.
This heavily degrades the position and orientation estimates of
the vehicle, especially for long and winding trajectories [1]. In
[2] practical methods are proposed to reduce odometry errors
due to uncertainty about the effective wheelbase and unequal
wheel diameter.
Absolute localization is performed processing the data pro-
vided by a proper set of sensors measuring some parameters
of the environment in which the vehicle is operating. A set of
sonars is generally used as external sensors device. Sonars are
fixed to the vehicle and measure the distance with respect to
parts of the known environment [3]–[10]. These sensors are
also widely utilized for the guidance of autonomous vehicles
with obstacle avoidance in unknown environment [11]–[14].
The characterization of sonar measures and/or the rejection of
unreliable sonar readings have been widely investigated [6],
[15]–[17].
Manuscript received September 22, 1997; revised August 12, 1998. This
work was supported by the TIDE Project 1279 of the CEC (MOVAID) and ASI
funds. This paper was recommended for publication by Associate Editor A.
Bicchi and Editor V. Lumelsky upon evaluation of the reviewers’ comments.
The authors are with the Dipartimento di Elettronica e Automatica, Uni-
versitá di Ancona, Ancona 60131, Italy (e-mail: L.Jetto@ee.unian.it; S.
Longhi@ee.unian.it).
Publisher Item Identifier S 1042-296X(99)01210-0.
1042–296X/99$10.00  1999 IEEE
219
The main drawback of absolute measures is their depen-
dence on the characteristics of the environment. Possible
changes to environmental parameters may give rise to erro-
neous interpretation of the measures provided by the localiza-
tion algorithm.
The actual trend is to use sensors of different nature and
to properly weigh the relative data according to their reliabil-
ity. For this purpose Kalman filtering techniques represent a
powerful tool [8], [10], [18]–[20].
In this paper, the data provided by odometric and sonar
sensors are combined together through an adaptive extended
Kalman filter (AEKF) providing on line estimates of robot
position.
The use of Kalman filtering techniques requires to derive
a stochastic state-space representation of the robot model
and of the measure process. Formally this can be readily
performed by applying the kinematic model of the robot and
the knowledge of measure equipment. The extended Kalman
filter (EKF) techniques proposed in the literature [8], [10],
[18]–[20], for the estimation of robot localization are based
on some fixed values of the input and measurement noise
covariance matrices. In many practical applications an “a
priori” information of this kind is often unavailable. This is
especially true if the trajectory is not elementary and if changes
in the environment occur and/or if it has a complex structure.
On the other hand it is well known how poor estimates of noise
statistics may seriously degrade the Kalman filter performance.
The main novelty of the AEKF here proposed is its capability
of adaptively estimating such unknown statistical parameters.
In developing the algorithm, a particular attention has been
paid to prevention of filter divergence and to the simplicity of
implementation in view of on line applications. Experimental
validation shows the effectiveness of the proposed algorithm.
The paper is organized in the following way. The robot
model and the sensor device equipment are described in
Section II. The adaptive algorithm is reported in Section III.
Section IV contains a detailed description of the experimental
tests performed on the mobile base.
II. THE SENSOR EQUIPMENT
Consider an unicycle-like mobile robot with two driving
wheels, mounted on the left and right sides of the robot, with
their common axis passing through the center of the robot (see
Fig. 1). Localization of this mobile robot in a two-dimensional
(2-D) space requires knowledge of the coordinates and of
220 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO. 2, APRIL 1999

Fig. 2. Sonar displacement.

Fig. 1. Scheme of the unicycle robot.

the midpoint between the two driving wheels and of the angle
between the main axis of the robot and the -direction. The
kinematic model of the unicycle robot is described by the
following equations:
(1)
(2)
(3)
where and are, respectively, the displacement and
angular velocities of the robot. The encoders placed on the
Fig. 3. Sonar measure.
driving wheels provide a measure of the incremental angles
over a sampling period . The odometric have the following form:
measures are used to obtain an estimate of the displacement
and angular velocities and , respectively, which are (4)
assumed to be constant over the sampling period. Numerical (5)
integration of (1) and (2) based on the computed and (6)
results in an estimate of the position and orientation
increments over each sampling period of the unicycle robot. The walls and the obstacles in an indoor environment of
The above processing is generally performed by an odometric a mobile robot can be well represented by a proper set of
device connected with the low level control equipment of the planes orthogonal to the plane of the inertial coordinate
robot imposing the desired and . system. Each plane , where is the
An analysis of the accuracy of the estimation procedure number of planes which describe the indoor environment, can
implemented by an odometric equipment has been developed be represented by the triplet , , , where is the
in [1]. The encoders introduce incremental errors, which espe- normal distance of the plane from the origin , is the
cially affect the estimate of the orientation . This limits the angle between the normal line to the plane and the -direction
applicability of the odometric localization to short trajectories and is a binary variable, , which defines the
only. face of the plane reflecting the sonar beam. Using the above
The localization of a mobile robot in an indoor environment notation, the expectation for the distance measure of
can also be performed with sonar measures [3]–[8], [10]. the sonar in the position , and with orientation
In this context, the environment can be represented by its from the plane represented by , , has the
elementary geometric parts as corners, angles, cylinders and following expression (see Fig. 3):
planes [6].
(7)
In the following, the sonar readings will be easily related to
the model of the indoor environment and the configuration if the , where is the
of the mobile robot. Consider a planar distribution of beamwidth of the sonar sensor. The vector composed of
sonar sensors. Denote by the planar position and geometric parameters , , , , is
orientation of the th sonar, , referred to the denoted by .
coordinate system fixed to the mobile robot, as To simplify the position estimation algorithm without appre-
reported in Fig. 2. ciably reducing its accuracy, the sonar echoes traveling along
The position , and orientation at the sampling time the cone edges, i.e., for ,
of the th sonar referred to the inertial coordinate system and , where is an angle
JETTO et al.: DEVELOPMENT AND EXPERIMENTAL VALIDATION OF AN ADAPTIVE EXTENDED KALMAN FILTER 221

depending of the surface roughness, have been omitted. In fact, the plane with . The environment map provides
the measures along the cone edges require an a priori model the information needed to detect which is the plane in
of the environment also including the different roughness of front of the th sonar.
the walls and are less accurate of the distance measures given By definition of the measurement vector one has that the
by (7). output function has the following form:

III. ADAPTIVE ESTIMATION OF ROBOT LOCATION

The proposed AEKF for the fusion of odometric and sonar
measures, providing on line estimates of robot position and
orientation, is derived in this section.
Denote with the robot state ..
and with the robot control input. The .
kinematic model of the robot can be written in the compact
form of the following stochastic differential equation: (10)
(8) where . The number of measures may vary from
the minimum value 3 to the maximum value 3, where
where is obtained by (1)–(3) and is a is the number of sonar sensors.
Wiener process such that . Its weak Assume for . To obtain an
mean square derivative is a white noise process extended Kalman filter (EKF) with an effective state prediction
representing the model inaccuracies (parameter equation in a simple form, model (1), (2) has been linearized
inaccuracies, slippage, dragging). Assuming a constant sam- about the current state estimate and the control
pling period and denoting by , input applied until the linearization instant.
the following sampled nonlinear measure equation can be Subsequent discretization with period of the linearized
associated to (8): model results in the following EKF (where explicit dependence
(9) on has been dropped for simplicity of notation)
(11)
where is the vector containing sonar and odometer
measures and is a white sequence . (12)
The dimension of is not constant, depending on
the number of sonar sensors that are actually used at each
time instant. The measure vector is composed of
(13)
two subvectors and
, where
, , (14)
are the measures provided by
the odometric device, and ,
(15)
, , with given by (7) is
the distance measure provided by the th sonar sensor from and in (16)–(19), shown at the bottom of the page.

(16)

(17)

(18)

..
.

(19)
222 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO. 2, APRIL 1999

The form of expressed by (17) derives by the possible to apply the methods of the adaptive filtering theory
hypothesis that . This developed for the linear case.
simplification assumption has been introduced to obtain a The two above assumptions will allow us to define a simple
which is completely known up to the unknown multi- and efficient estimation algorithm based on the condition of
plicative scaling factor . Moreover, the covariance matrix consistency, at each step, between the observed innovation
is assumed to have the following diagonal form: process samples and their predicted
statistics . Imposing such a condition,
diag (20)
one stage estimates and ,
this means that no correlation is assumed between the measure- of and , respectively, are
ment errors introduced by the sensors. As is diagonal, the obtained at each step. To increase their statistical significance,
components of may be processed one by one reducing the one stage estimates and ,
the inversion of the matrix in (13) to the inversions are averaged obtaining the relative smoothed versions
of scalars [21], thus saving much computation time. The and .
sequential processing of each component From (21) it is apparent that the statistical information
must be performed in a period of time (typically the sampling carried by each depends, at the same
period) such that no significant change occurs in the state time, on the two unknown parameters and .
estimate and in its covariance matrix due to dynamics (8) [21]. This indeterminateness is here dealt with using a number
The EKF can be implemented once estimates of (say ) of innovation process samples ,
and are available. In general, a complete and reliable , to estimate and the others (say ) to
information about these matrices is not available; on the other estimate . In the light of the nearly stationary
hand it is well known how poor knowledge of noise statistics assumption, the two integers and are chosen such that
may seriously degrade the Kalman filter performance. This .
problem is here dealt with introducing an adaptive adjustment Assume , let and two coprime integers
mechanism of and values in the EKF equations. such that and let and two integers such
that ; then, the innovation process sequence is
A. Adaptive Estimation of and subdivided into intervals composed of samples.
A considerable amount of research has been carried out Each interval contains sequences of samples used to
in the adaptive Kalman filtering area (see [21]–[23] and estimate (the faster varying parameter), the ensembles
references therein), but in practice it is often necessary to of samples are separated by sequences of one sample
redesign the adaptive filtering scheme according to the partic- used to estimate (the more slowly
ular characteristics of the problem faced. For example, in view varying parameter), the last samples of each interval
of real time applications, a particular attention has been here are used to estimate . This scheme minimizes the interval
devoted to simplicity of implementation and to prevention of of time over which either one step estimate is not updated. A
filter divergence, moreover, the particular structure of the input symmetric situation holds if .
noise covariance matrix , which is completely known When the one step estimate
save that for a multiplicative scalar, has been suitably taken is updated, the other single stage estimate
into account. is kept constant, so
The following nearly stationarity assumption is made: the that the symbol does not necessarily imply
parameters , , and are nearly that this estimate has been computed using the last observed
constant over 2 and 2 samples, respectively. innovation process sample .
Define , where Because of the particular form (17) of and of
and are the th component of the sequential scalar processing of measures, one stage
and , respectively. For analogy estimates of the unknown
with the linear case, residuals are can be determined maximizing the probability of observing
called the innovation process samples and are assumed to be the corresponding th component of the predicted residual
well described by a white sequence , where [21]. Namely, each is
can be expressed as determined by the operation
max prob
The maximizing is obtained by imposing the con-
dition of consistency between residuals and their predicted
(21)
statistics, namely .
This simplifying assumption is as more valid as discretiza- Using (21) and replacing with one has
tion and linearization of (8) is more accurate and makes it (22), shown at the bottom of the page.

(22)
JETTO et al.: DEVELOPMENT AND EXPERIMENTAL VALIDATION OF AN ADAPTIVE EXTENDED KALMAN FILTER 223

To obtain a unique estimate of and to increase the information, a large value of is useful to prevent
statistical significance of estimators (22), which are based divergence.
on only one component , the following smoothed
estimate is computed: B. Sonar Readings Selection
It is known that more sensor data does not always mean bet-
(23) ter estimates. The grounds may be an inadequate interpretation
of erroneous data due to sensor failures or the inconsistency of
the underlying filter equations, this second reason is discussed
where denotes the number of one-stage estimates
in [24]. In the present case, sonar measurements may not only
yielding the smoothed estimate.
be affected by white Gaussian noise but also by crosstalk
In a recursive form the proposed estimate of is
interferences and multiple echoes [5], [25]. Techniques for the
treatment of these systematic errors based on the construction
of a navigation map and/or on a time pulse modulation have
been proposed in [16] and [17], respectively. A simple method
(24) is proposed here to deal with the undesired interferences
produced by the presence of unknown obstacles. A sort of
Analogously, the operation prefiltering is performed neglecting those sonar measures that
are not deemed to be reliable. A simple and efficient way to
max prob perform these preliminary measure selection is to compare the
actual sonar readings with their expected values. Measures are
and (21) give the following one stage estimate of discharged if the difference exceeds a threshold. The value
of the threshold has to be chosen on the basis of two oppo-
site considerations: too large values would accept undesired
interferences as reliable sonar readings, while too low values
would produce an undesired loss of useful information. For
this reason, it is preferable not to choose an unique fixed
(25) threshold, but to adapt the preliminary choice on line, on the
basis of the reliability of the current predicted sonar measure.
the smoothed version is This is here done in the following way: at each step, the
residual ,
, represents the difference between
(26)
the actual sonar measure and its expected value
which is computed on the basis of
where denotes the number of one-stage estimates the estimated robot location and on the a priori knowledge
yielding the smoothed estimate. of the environment. As , the
In a recursive form the proposed estimate of current value is accepted if
becomes . Namely, the variable threshold is chosen as
two times the standard deviation of the innovation process.
(27) This adaptive threshold realizes a reasonable tradeoff between
the aforementioned opposite considerations. If the
The proposed adaptive estimation algorithm is given by (24)
is reliable, so is , namely the expected
and (27) and is able to prevent filter divergence. In fact, as
sonar reading, so that a strict test must be performed by
long as the innovation samples are
choosing a low threshold. This is automatically obtained taking
sufficiently small and consistent with their statistics, the filter
into account equation that, by (21), reliability of
operates satisfactorily and the noise model is kept small (or
correspond to a low . The opposite occurs if
null) by (22). If a sudden increase of the absolute value of
has a large error covariance matrix.
the innovation process samples is observed, (22) provides an
The structure of the proposed localization algorithm is
increased , and hence an augmented filter
reported in Fig. 4.
gain, thus preventing filter divergence.
Parameters and of estimators (24) and (27) are chosen
on the basis of two antagonist considerations: low values IV. EXPERIMENTAL RESULTS
would produce noise estimators which are not statistically The experimental tests have been performed on the LabMate
significant, large values would produce estimators which are mobile base in an indoor environment with different geometry.
scarcely sensitive to possible rapid fluctuations of the true This mobile robot is realized with two driving wheels, as
and . During filter initialization, reported in Fig. 1, and the odometric data are the incremental
the starting values and of measures that at each sampling interval are provided by
and , respectively, must be chosen on the basis of the the encoders attached to the right and left wheels of the
“a priori” available information. In the case of a lack of such robot. These measures are directly acquired by the low level
224 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO. 2, APRIL 1999
Fig. 4. Scheme of the localization algorithm.
Fig. 5. LabMate mobile base with the proximity system.
controller of the mobile base. The sonar measures have been
realized by the standard proximity system of the LabMate base
composed by a set of nine Polaroid sonar sensors. In Fig. 5, a
picture of LabMate system with the sonar sensors displacement
is reported.
A preliminary reduction of crosstalk interferences has been
realized by a proper distribution on the orientations of the
sonar sensors. The sonars on the lateral sides of the mobile
base have been mounted with a difference on the orientation of
15 , as reported in Fig. 6. A significant reduction of the wrong
readings produced by the presence of unknown obstacles has
been also realized selecting of the sonar measures by the
procedure described in Section III-B.
The control algorithm is based on a proper discrete-time
implementation of the control algorithm proposed in [26] and
based on the kinematic inversion approach. The continuous-
time control algorithm is explicitly represented by the follow-
ing equations:
(28)
(29)
Fig. 6. Location of the nine Polaroid sonar sensors on the LabMate mobile
base.
where , , , and are the gain values and
(30)
(31)
and , , and represent the desired planned
trajectory. The AEKF has been implemented on a personal
computer with Windows 3.1 system by the developing en-
vironment described in [27]. In this development system,
the planned trajectory has been computed considering the
nonholonomic and environment constraints according to the
algorithm proposed in [28]. The system is connected directly
with the robot low level controller and the proximity system
by a standard RS232 serial port. The sampling period
0.3 s has been chosen.
In the developed experiments, the proposed localization
algorithm has been tested with relatively long trajectories
on a indoor environment represented by a suitable set of
planes orthogonal to the plane of the inertial system.
The first experiment performed is illustrated in Fig. 7. The
planned trajectory from the start configuration to the goal
configuration (the gray path) is computed by the devel-
opment environment described in [27]. In order to test the
capability of reducing the position error on the proposed
localization algorithm, at the configuration the robot starts
with an initial position error of 0.26 m. Frame (a) shows the
trajectory obtained with the a priori fixed ,
m , for , for any
integer . Frame (b) shows the trajectory obtained with AEKF
which has been implemented with initial values and
equal to corresponding values used for EKF. Frame
(c) shows the behavior of the estimated assuming
6. This figure evidences a sharp increase of in
correspondence of the configuration of the mobile base
when the sonar measures start to be actually used. On the
configuration the effects of the sonar readings is a rapid
increase of residuals and, by (23), increases too.
As a consequence also the filter gain increases and this allows
the filter to properly weigh the incoming observations. The
estimated robot position is rapidly corrected and this provokes
a decrease of residual samples and hence of the estimated
JETTO et al.: DEVELOPMENT AND EXPERIMENTAL VALIDATION OF AN ADAPTIVE EXTENDED KALMAN FILTER 225

(a) (b)

(c)
Fig. 7. Dots path is the planned trajectory, the dark path is the realized trajectory and the gray dots are the actually used sonar measures: (a) localization
with the EKF, (b) localization with the AEKF, and (c) behavior of the estimated  ^ 2 (k ) assuming l = 6.

. The adaptive estimation of did not produce
very significant changes with respect the initial values because
of the large 60 chosen on the basis of experimental tests.
This choice was also motivated by the proximity of the initial
value to the values provided by the off-line calibration
of the sonars. This test evidences the improvement introduced
by the adaptation algorithm, which is able to reduce to 0.01 m
the initial position error by better exploiting the information
carried by the sonar readings in correspondence of the obstacle,
while the EKF yielded a final error of 0.11 m.
Fig. 8 illustrates the results of the second experiment. Part
(a) of this figure represents the realized trajectory with lo-
calization deduced by odometric measures only. In order to
test the limitation of the odometric measures, the planned
trajectory is composed by a large set of orientation changes.
The black path is the realized trajectory with only odometric
measures. In this case, at the end of the test, the robot is
out of the planned trajectory. Part (b) shows the same test
with localization based on the EKF implemented with “a
priori” chosen and . Part (c) shows the same test
226 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO. 2, APRIL 1999

(a) (b)

(c)
Fig. 8. Dots path is the planned trajectory, the dark path is the realized trajectory, the gray dots are the actually used sonar measures: (a) localization
with only odometric measures, (b) localization with EKF, and (c) localization with the AEKF.

with localization based on the AEKF. In all the experiments A long trajectory developing through three different rooms
the values 2 have been assumed; this means has been considered in the third experiment to test the pro-
that the innovation process vector samples have posed localization algorithm in realistic operating conditions.
been alternately used to estimate and For convenience, the graphic representation of the results
according to the algorithm described in Section III- has been reported on a sequence of figures [Fig. 9(a)–(e)]
A. The plot clearly evidences the improvement introduced by where the goal configuration of each frame is the start
the adaptation mechanism. configuration of the subsequent. Fig. 9 reveals that the
JETTO et al.: DEVELOPMENT AND EXPERIMENTAL VALIDATION OF AN ADAPTIVE EXTENDED KALMAN FILTER 227

(a) (b)

(c) (d)

(e)
Fig. 9. Planned trajectory is represented in the frames sequence from (a)–(e). The gray path is the planned trajectory, the black path is the realized
trajectory, and the black dots are the actually used sonar measures.
228 IEEE TRANSACTIONS ON ROBOTICS AND AUTOMATION, VOL. 15, NO. 2, APRIL 1999
Fig. 10. Localization with the presence of an unknown obstacle in side of
a wall. The gray path is the planned trajectory, the black path is the realized
trajectory and the black dots are the actually used sonar measures.
realized and planned trajectories are almost every where nearly
coinciding.
Fig. 10 shows the effect of the sonar measure selection. An
unknown obstacle of small dimensions, which is not intro-
duced in the environment map, is placed at the side of a wall
and the corresponding readings are seen as undesired interfer-
ences (the considered problem is not obstacle avoidance). No
perturbation affects the robot motion because these measures
are eliminated by the procedure outlined in Section III-B. Of
course this procedure, based on sonar measure selection is
successful as far as the unknown obstacle dimensions are small
enough compared with the reference wall dimensions.
V. CONCLUSION
This paper proposed a new method for the accurate lo-
calization of a mobile robot. The approach is based on the
use of a linearized Kalman filter endowed with an adaptive
algorithm for the adjustment of the input and measurement
noise covariance matrices. The adaptation mechanism has been
introduced to allow the filter to cope with realistic operating
conditions. If the planned trajectory is relatively simple and
not too long, some “a priori” engineered noise statistics may
produce satisfactory results, but filter divergence may occur
over complex trajectories and the localization algorithm pro-
vides unreliable estimates. In this latter case the introduction
of an adaptive algorithm seems to be the most efficient and
simple remedy to prevent filter divergence. The experiments
reported in the paper confirmed that high performances of the
localization algorithm are really obtainable in a wide range of
experimental situations.
REFERENCES
[1] C. M. Wang, “Localization estimation and uncertainty analysis for mo-
bile robots,” in Proc. Int. Conf. Robot. Automat., 1988, pp. 1230–1235.
[2] J. Borenstein and L. Feng, “Measurement and correction of systematic
odometry errors in mobile robots,” IEEE Trans. Robot. Automat., vol.
12, pp. 869–880, Dec. 1996.
[3] M. Drumheller, “Mobile robot localization using sonar,” IEEE Trans.
Pattern Anal. Machine Intell., vol. 9, no. 2, pp. 325–332, Mar. 1987.
[4] J. L. Crowley, “World modeling and position estimation for a mo-
bile robot using ultrasonic ranging,” in Proc. IEEE Int. Conf. Robot.
Automat., 1989, pp. 674–680.
[5] J. L. Leonard and H. F. Durrant-Whyte, “Mobile robot localization by
tracking geometric beacons,” IEEE Trans. Robot. Automat., vol. 7, pp.
376–382, June 1991.
[6] , Direct Sonar Sensing for Mobile Robot Navigation. Boston:
Kluwer, 1992.
[7] A. A. Holenstein, M. A. Müller, and E. Badreddin, “Mobile robot
localization in a structured environment cluttered with obstacles,” in
Proc. IEEE Int. Conf. Robot. Automat., 1992, pp. 2576–2581.
[8] A. Curran and K. J. Kyriakopoulos, “Sensor-based self-localization
for wheeled mobile robots,” in Proc. IEEE Mediterranean Symp. New
Direction Contr. Automat., 1993.
[9] P. van Turennout, G. Honderd, and L. J. van Schelven, “Wall-following
control of a mobile robot,” in Proc. IEEE Int. Conf. Robot. Automat.,
1992, pp. 280–285.
[10] L. Charbonnier and A. Fournier, “Heading guidance and obstacles
localization for indoor mobile robot,” in Proc. Int. Conf. Adv. Robot.,
ICAR’95, 1995, pp. 507–513.
[11] A. Elfes, “Sonar-based real world mapping and navigation,” Au-
tonomous Robot Vehicles, J. I. Cox and G. T. Wilfgong, Eds. Berlin,
Germany: Springer-Verlag, 1990.
[12] R. Kuc and V. B. Viard, “A physically based navigation strategy for
sonar-guided vehicles,” Int. J. Robot. Res., vol. 10, no. 2, pp. 75–87,
1991.
[13] F. Figueroa and A. Mahajan, “A robust navigation system for au-
tonomous vehicles using ultrasonics,” Contr. Eng. Practice, vol. 2, no.
1, pp. 49–59, 1994.
[14] A. Angeloni, T. Leo, S. Longhi, and R. Zulli, “Real time collision
avoidance for mobile robots,” Proc. 6th Int. Symp. Meas. Contr. Robot.,
1996, pp. 239–244.
[15] Ö. Bozma and R. Kuc, “Characterizing pulses reflected from rough
surfaces using ultrasound,” J. Acoust. Soc. Am., vol. 89, no. 6, pp.
2519–2531, 1991.
[16] M. Beckerman and E. M. Oblow, “Treatment of systematic errors in
the processing of wide-angle sonar sensor data for robotic navigation,”
IEEE Trans. Robot. Automat., vol. 6, pp. 137–145, Apr. 1990.
[17] J. Borenstein and Y. Koren, “Error eliminating rapid ultrasonic firing
for mobile robot obstacle avoidance,” IEEE J. Robot. Automat., vol. 11,
pp. 132–138, Feb. 1995.
[18] J. Vaganay, M. J. Aldon, and A. Fournier, “Mobile robot attitude
estimation by fusion of inertial data,” in Proc. IEEE Int. Conf. Robot.
Automat., 1993, pp. 277–282.
[19] K. Kobayashi, K. C. Cheok, and K. Watanabe, “Fuzzy logic rule-based
Kalman filter for estimating true speed of a ground vehicle,” Intell.
Automat. Soft Comput., vol. 1, no. 2, pp. 179–190, 1995.
[20] G. Garcia, P. Bonnifait, and J.-F. Le Corre, “A multisensor fusion
localization algorithm with self-calibration of error-corrupted mobile
robot parameters,” in Proc. Int. Conf. Adv. Robot., ICAR’95, 1995, pp.
391–397.
[21] A. H. Jazwinsky Stochastic Processes and Filtering Theory. New York:
Academic, 1970.
[22] B. D. O. Anderson and J. B. Moore, Optimal Filtering. Englewood
Cliffs, NJ: Prentice-Hall, 1979.
[23] R. G. Brown and P. Y. C. Hwang, Introduction to Random Signals and
Applied Kalman Filtering. New York: Wiley, 1997.
[24] W. D. Blair and Y. Bar-Shalom, “Tracking maneuvering targets with
multiple sensors: Does more data always mean better estimates?,” IEEE
Trans. Aerosp. Electron. Syst., vol. 32, pp. 450–456, 1992.
[25] J. Borestein and Y. Koren, “Noise rejection for ultrasonic sensors in
mobile robot applications,” in Proc. IEEE Int. Conf. Robot. Automat.,
1992, pp. 1727–1732.
[26] A. De Luca and G. Oriolo, “Local incremental planning for nonholo-
nomic mobile robots,” in Proc. IEEE Int. Conf. Robot. Automat., 1994.
[27] M. Bonifazi, F. Favi, T. Leo, S. Longhi, and R. Zulli, “A developing
environment for the solution of the navigation problem of mobile robots
with nonholonomic constraints,” in Proc. 4th IEEE Mediterranean Symp.
New Direction Contr. Automat., Krete, Greece, 1996, pp. 107–112.
JETTO et al.: DEVELOPMENT AND EXPERIMENTAL VALIDATION OF AN ADAPTIVE EXTENDED KALMAN FILTER 229

[28] G. Conte, S. Longhi, and R. Zulli, “Motion planning for unicycle and Giuseppe Venturini was born in Chieti, Italy, on
car-like robots,” Int. J. Syst. Sci., vol. 27, no. 8, pp. 791–798, 1996. July 30, 1971. He received the Doctor degree in
electronic engineering from the University of An-
cona, Italy, in 1996.
Since 1996, he has been with the Alcatel Telecom
Italia R&D Laboratory (Elasis Consortium), Chieti,
Leopoldo Jetto received the Laurea degree in elec- where he is a Software Designer. His research inter-
tronic engineering from the University of Ancona, ests include digital signal processing and intelligent
Italy, in 1976 and the postgraduate diploma degree control for wireless communications systems.
in automatic control from the University of Roma,
Rome, Italy, in 1981.
After working in the industry, he joined the Di-
partimento di Elettronica e Automatica, University
of Ancona, in 1979, first as a Lecturer and then
as a Researcher. In 1993, became an Associate
Professor of Automatic Control at the Dipartimento
di Automatica e Informatica, Politecnico di Torino,
and now holds the same position with the Dipartimento di Elettronica e
Automatica, University of Ancona. He is also a Consultant for the Project
of Scientific Cooperation, Foreign Office, with growing countries. His main
research interests lie in the theory of algebraic control in the theory and
applications of optimal linear filtering and control, both in finite and infinite
dimensions, and in signal processing. He is author and co-author of about 70
papers in the above fields.

Sauro Longhi (M’92) was born in Loreto, Italy, on

September 11, 1955. He received the Doctor degree
in electronic engineering from the University of An-
cona, Italy, in 1979 and the Post-Graduate Diploma
degree in automatic control from the University of
Rome, Italy, in 1985.
From 1980 to 1981, he held a Fellowship at the
University of Ancona. From 1981 to 1983, he was
with the Telettra S.p.A., Chieti, Italy. Since 1983,
he has been with the Dipartimento di Elettronica e
Automatica, University of Ancona, where he is an
Associate Professor in Control Systems Technologies. His research interests
include analysis and control of linear systems, control of mobile base robots,
and control of underwater vehicles.
Dr. Longhi is a member of the Technical Committee on Marine Systems,
IFAC.