Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Mechatronics
journal homepage: www.elsevier.com/locate/mechatronics
a r t i c l e i n f o a b s t r a c t
Article history: Odometry provides fundamental pose estimates for wheeled vehicles. For accurate and reliable pose esti-
Received 1 August 2008 mation, systematic and nonsystematic errors of odometry should be reduced. In this paper, we focus on
Accepted 3 June 2010 systematic error sources of a car-like mobile robot (CLMR) and we suggest a novel calibration method.
Kinematic parameters of the CLMR can be successfully calibrated by only a couple of test driving. After
reducing deterministic errors by calibration, odometry accuracy can be further improved by redundant
Keywords: odometry fusion with the extended Kalman lter (EKF). Odometry fusion reduces nonsystematic or sto-
Mobile robot
chastic errors. Experimental verications are carried out using a radio-controlled miniature car.
Calibration
Odometry
2010 Elsevier Ltd. All rights reserved.
Pose recognition
0957-4158/$ - see front matter 2010 Elsevier Ltd. All rights reserved.
doi:10.1016/j.mechatronics.2010.06.002
K. Lee et al. / Mechatronics 20 (2010) 582595 583
Fig. 2. Difference between a two wheel differential mobile robot and the CLMR.
neously, the test track should include both straight and curved the CW and CCW direction, odometry is computed as a dashed line
path, as explained in Section 2.2. Therefore, for examples, T- or in Fig. 5. In this case, the nominal tread is smaller than the actual
S-shaped tracks are not available. Once the CLMR is driven through tread.
the test track twice, the systematic errors can be calibrated by Since we assume that the rear wheel diameters are identical in
using the proposed scheme. One is to make a CCW round trip from Type A errors, odometry path and the real path are identical in the
SCCW . The other is to make a CW round trip from SCW . straight line section. On the other hand, in the semicircle section,
The reason why we drive the robot along the test track in both the radius of curvature from odometry data is different from the
CW and CCW directions is as follows. If the CLMR is driven in only real path because the tread is not calibrated. As shown in Fig. 5,
one direction, odometry error of the nal position could be small we assume that the error in semicircular section is represented
regardless of uncalibrated parameters. It is possible that the effects by angle a. In CW and CCW direction, nal position (x4, y4) can
of unequal wheel diameters and tread errors may be internally be obtained as following equations. Detailed equations are given
compensated each other. In such a case, odometry error of at the in [17].
nal position becomes larger when the CLMR is driven in the oppo- p
site direction. pa qa
x4;CW 2
CW : 6
The systematic error calibration is decomposed into two steps. y4;CW 2qa
First, we dene Type A error which implies the uncalibrated tread
error, while it is assumed that rear wheel diameters are assumed to x4;CCW p2pa qa
have no modeling error. Seconds, we dene Type B error that re- CCW : 7
y4;CCW 2qa
sults from an uncalibrated wheel diameter without tread error.
Fig. 5. Type A errors in CW motion from SCW (top) and CCW motion from SCCW
(bottom). Fig. 6. Type B errors in CW (top) and CCW (bottom) motions.
586 K. Lee et al. / Mechatronics 20 (2010) 582595
In an ideal case, the values for a from (11) and (14) should be
identical and the values for b from (12) and (15) should also be
identical. Since the positional error is affected by various error
sources, the experimental a and b are different from ideal values.
However, it is reasonable to assume that odometry error at the -
nal pose will converge to zero, when kinematic parameters con-
verge to real values. If the resultant odometry error is not
sufciently small, iterative change of kinematic parameters possi-
bly improve odometry accuracy. By using a and b, we can calibrate
tread and wheel diameters.
From Fig. 5, we can obtain a relationship regarding the heading
error between the real path and odometry path. The relationship
between a and the tread, b, can be represented as the following
equation.
p pa
16
bactual bnominal
Fig. 7. Geometric relation between the radius of curvature and the length of the
Finally, (16) can be rewritten as the following equation. straight line under uncalibrated tread.
K. Lee et al. / Mechatronics 20 (2010) 582595 587
3. Reduction of the nonsystematic error By substituting Ddr in Eq. (1) with Eq. (25), we can dene the
observation, zk, as follows.
Nonsystematic errors are caused by various uncertainties 0 1
Dh
including wheel slippage, uneven ground contact conditions and xk1;f p
1
Ddr cos hk1;f 2 f
B 1tan / C
so forth. Fortunately, the CLMR has redundant odometry informa- B C
zk B
B yk1;f p
1 Dh
Ddr sin hk1;f 2 f C
C 26
tion that is provided from the front and rear wheels. If some wheels @ 1tan / A
slip at certain moments, the redundant odometry can be used for hk1;f Dhf
compensating the wheel slippages except in the extreme case of
slippage in all wheels. In order to compensate nonsystematic errors The nonlinear measurement model is presented as the follow-
which are stochastic, the EKF is adopted to fuse the front and real ing equation.
odometry information. The detailed algorithms on the EKF can be zk hxk v k 27
found in existing literatures as in [1]. The sequence of odometry fu-
sion through the EKF is divided into three steps as follows. In the above, hxk denotes the measurement vector, hxk
xk yk hk T , and vk is Gaussian sensor noise vector with covariance
3.1. System model matrix Rk.
The system model is dened by the kinematic nonlinear equa- 3.3. The EKF algorithm
tions as presented in (1), with the state vector xk = [xkykhk]T and in-
put, uk = [DdDh]T, which can be written in the following form with The EKF is used for combining the front and rear odometry
noises. redundancy. The pose of the CLMR is predicted by using the rear
odometry at the prediction stage. The front odometry is fused at
xk f xk1 ; uk ; ck ; rk 20 the correction stage.
In Eq. (20), ck and rk denote the system and input noises with
the associated matrices, Uk and Qk. 3.3.1. Prediction stage
^ f X
X ^ k1 ; uk ; 0; 0
3.2. Measurement model k
28
Pk Ak Pk1 ATk Bk Ck1 BTk Q
In order to use the EKF, odometry of the front axle should be
A and B are shown in the following Jacobian.
converted to odometry of the rear axle. Fig. 8 shows the geometry
2 3
of the CLMR in a curve. The distances between I.C. and the front 1 0 Ddk sin hk D2hk
and rear axle centers are presented as follows. 6 7
Ak 4 0 1 Ddk cos hk D2hk 5 29
q
q L2 q2r 21 0 0 1
2 3
L coshk D2hk D2dk sinhk D2hk
qr 22 6 7
tan / Bk 4 sinhk Dhk Ddk coshk Dhk 5 30
2 2 2
0 1
By combining (21) and (22), we obtain the following equation.
s s
L2 1
qf L2 2 L 1 2 23 3.3.2. Correction stage
tan / tan /
Sk Hk Pk HTk Rk
The travel distance of the front and rear axle centers during Dt is
represented as follows. K k Pk HTk S1
k
31
Ddr qr Dh; Ddf qf Dh: 24 ^k X
X ^ K k zk hX
^
k k
4.1. Simulations
and the reference path show a great difference. However, the end-
position error is extremely small. This fact implies that the effect of
the tread error and wheel diameter errors were balanced and can-
celled each other at the end-position. In order to solve this prob-
lem, the CLMR should be driven in the opposite direction. The
result of CW motion is shown in Fig. 10.
In Fig. 10, the end-position error between the real position of
the CLMR and odometry is extremely large due to uncalibrated
parameters. Since the reference motion is different from the previ-
ous test, the error balancing problem can be avoided. For further
investigations, the end-position errors were computed under kine-
matic parameter errors. Fig. 11 shows the computed result.
In Fig. 11, x and y axes represent uncalibrated tread and wheel
diameters respectively. z axis correspond to the end-position error.
The end-position error should be a minimum around the real kine-
matic parameter. The valley in Fig. 11 implies that the end-position
error can be small for a wide variety of parameter sets. The valley is
resulted from the error-balancing problems. In order to overcome
error balancing, different reference trajectory should be chosen.
Fig. 9. Odometry path including both tread and wheel diameter errors: CCW
Fig. 12 shows end-position errors for both CCW and CW motions.
motion. Fig. 12 is obtained by computing end-position errors for kine-
matic parameters around the valley of Fig. 11. If the CLMR is driven
to CCW direction only, kinematic parameters may not converge to
real values due to error balancing. However, it is clear that the end-
position error increases under CW motion. Therefore, calibrations
could be successfully carried out by adopting both CW and CCW
motions.
As shown in Fig. 13, the valley of the error surface is also ob-
tained under CW motion. Therefore, it is desirable to obtain the -
nal pose error under CCW motion. In Fig. 14, the CW motions have
small end-position errors despite uncalibrated parameters. How-
ever, the end-position errors under the CCW motions show great
differences around actual parameters (b = 31.5 cm, Dr =Dl = 1.00).
Therefore, calibrations could be successfully carried out by adopt-
ing both the CW and CCW motions.
Fig. 15a represents the total end-position errors for uncali-
brated wheel diameters under calibrated treads. It is clear that
the total end-position errors monotonically increase around the ac-
tual wheel diameter ratio of 1.00. The region of convergence is
about 20%, which is acceptably large in practical applications.
Fig. 10. Odometry path including both tread and wheel diameter errors: CW Fig. 15b shows the total error for uncalibrated treads under cali-
motion. brated wheel diameters. The length of the rear tread varies from
Fig. 11. The end-position errors under tread and wheel diameter errors: CCW motion.
K. Lee et al. / Mechatronics 20 (2010) 582595 589
Fig. 12. The end-position errors for both CW and CCW motions including parametric errors with respect to Fig. 11.
Fig. 13. The end-position errors under tread and wheel diameter errors: CW motion.
21.5 cm to 41.5 cm. The total end-position errors converge to zero From Fig. 16, the global minimum is located at the point that
as the length of the tread moves close to the actual value. The re- represents the actual parameters. Therefore, calibrations could be
gion of convergence is about 30%. successfully carried out by adopting both the CW and CCW
Fig. 16 is obtained by combining Figs. 11 and 13. Fig. 16b and c motions.
shows the total error surface observed from different viewpoints.
The error surface presented in Fig. 16a reects the cases in which 4.2. Experimental systems and steering calibration
both tread and wheel diameter errors exist simultaneously. The
presented error surface represents the convergence of the total The CLMR used in experiments is shown in Fig. 17. This CLMR
end-position errors to zero under the proposed calibration scheme. has a same kinematic structure as a real car.
The error-balancing problems are successfully prevented by adopt- As shown in Fig. 17b, the commercialized position sensor (Star-
ing both CW and CCW motions. The region of convergence is about Gazer) is adopted for monitoring the real pose of a vehicle. It is an
30% for the uncalibrated tread and 20% for the uncalibrated articial landmark based device. The average positional error is
wheel diameters. We dene the error surface in the following form. about 5 cm and the orientation error is below 5. In a mobile robot,
Etotal f b; Dr =Dl 34 we can recongure the driving mechanism. For experiments, a
front wheel driven mechanism has been adopted. Passive rear
590 K. Lee et al. / Mechatronics 20 (2010) 582595
Fig. 14. The end-position errors for both CW and CCW motions including parametric errors with respect to Fig. 13.
Fig. 15. Total end-position errors under (a) wheel diameter errors and (b) tread errors.
wheels result in better odometry accuracy than active front computed from experimental curvatures. It is clear that the steer-
wheels. ing angle calibration can be accurately carried out with strong
According to the ground contact conditions or modeling errors linearity.
with regard to the steering angle, moving direction of the CLMR
may not coincide with the desired steering input. Since the accu- 4.3. Systematic error calibration results
racy of steering motion dominates the overall control performance,
the steering angle should be calibrated rst. Regardless of the other The size of the test track should not be too small because cali-
kinematic parameters, the vehicle moves along a circular path un- bration results become too sensitive with respect to odometry er-
der a constant steering input. rors. Therefore, appropriate size of the test track with respect to
Fig. 18 shows the traces of the CLMR from successive camera the robots dimension is required for calibrating the kinematic
images under constant steering angles. The steering angle changes parameters. In experiments, we set the size of the test track as
from 20.6 to 19.7 which are joint limits. From the traces of the 350 cm 250 cm by considering the physical dimension of the
CLMR, the resultant curvature of the circular path in Fig. 19a can be wheel diameter and the tread.
computed by the least square tting in [18]. From Fig. 19a, it is The CLMR is driven by open loop in the experiments. The CLMR
clear that the experimental path is circular and the curvature is al- moves along the test track by constant speed of 100 mm/s. The
most constant. Fig. 19b shows the calibration result between the steering angle of the robot is controlled by an RC-Servo motor.
steering control inputs and the resultant steering angles that are The sequence of open loop control is as follows.
K. Lee et al. / Mechatronics 20 (2010) 582595 591
Fig. 16. Integrated end-position errors under tread and wheel diameter errors.
(1) From the starting position, go straight by 2q (where q is the Experimental results are shown in Fig. 20. As represented in
curvature radius of the semicircle of the test track). dashed line in Fig. 20, the real path of the CLMR measured by the
(2) Turn left (or right) with constant radius of curvature q, until position sensor well match with the reference path. Final positions
the semicircle section is completed. after ve test drives are marked by x symbols around the starting
(3) Go straight again by 2q. location in Fig. 20. When the starting position is (x, y) = (0, 0), nal
(4) Turn left (or right) with constant radius of curvature q, until positions are shown in Table 1. As shown in Table 1, the mean of
the semicircle section is completed. ve positional errors is 0.57 cm. The error is small when it is com-
(5) The CLMR returns to the starting position. pared to the positional error after calibration. For example, the
positional error after calibration was 9.6 cm as presented in exper-
The fundamental assumption of the proposed odometry calibra- iment #1 of Table. 2.
tion scheme is that the vehicle accurately follows the reference The proposed systematic error calibration scheme is experi-
path regardless of uncalibrated parameters. The nal pose error mentally tested. Fig. 21 shows experimental results before and
should be acceptably small. Therefore, we carried out open loop after the rst and the second calibrations. In experiments, not only
driving test of the CLMR rst. systematic errors but also nonsystematic errors affect odometry
592 K. Lee et al. / Mechatronics 20 (2010) 582595
Fig. 17. The CLMR used in experiments and the position sensor.
calibration method was applied, the parameters were calibrated to Emax;syst maxr c:g:;CW ; r c:g:;CCW 37
b = 28 cm and DR/DL = 0.99. From Fig. 21, it is obvious that the loca-
Table 2
Results on the calibration of the systematic error.
Exp. Experimental conditions Before calibration Emt (cm) After calibration Eyst (cm) Figure
(Front) (Rear)
1 Details are shown in Fig. 21 52.9 9.6
4 Winding tape around left outside of rear left and right wheels 91.5 8.9
In all cases, the nal position error was remarkably decreased ow-
ing to the calibration. It is clear that our calibration scheme pro-
Fig. 22. Tape winding around the rear left wheel.
vides accurate results regardless of the changes in kinematic
parameters. The experiments veried that the proposed method
can be used to calibrate the systematic odometry error.
5. Conclusion
cation target is the mobile robots with hard wheels. It can be said [6] Doh NL, Choset H, Chung WK. Accurate relative localization using odometry.
In: IEEE international conference on robotics and automation, Taipei, Taiwan;
that the variation of kinematic parameters is not large in mobile
September 2003. p. 160612.
robots with hard wheels such as urethane wheels. Regardless of [7] Kelly A. Linearized error propagation in odometry. Int J Rob Res 2004;23(2):
the necessity of repetitive calibration, the proposed scheme signif- 179218.
icantly contributes to the improvement of the localization [8] Martinelli A. The odometry error of a mobile robot with a synchronous drive
system. IEEE Trans Rob Autom 2002;18(3):399405.
performance. [9] Philip J. Mckerrow, Danny Ratner. Calibrating a 4-wheel mobile robot, IEEE/RSJ.
In: Intl conference on intelligent robots and systems, Lausanne, Switzerland;
Acknowledgements October 2002. p. 85964 .
[10] Thrun S, Burgard W, Fox D. A probabilistic approach to concurrent mapping
and localization for mobile robots. Mach Learning 1998;31:2953.
This work was supported in part by the MKE under the Human [11] Bento LC, Nunes U, Moita F, Surrcio A. Sensor fusion for precise autonomous
Resources Development Program for Convergence Robot Special- vehicle navigation in outdoor semi-structured environments. In: IEEE int conf
on intelligent transportation systems (ITSC2005), Basel, Switzerland; May
ists. This work was also supported by Mid-career Researcher Pro- 2005. p. 24550.
gram through the NRF grant funded by the MEST. This work was [12] Surrcio A, Nunes U, Arajo R. Fusion of odometry with magnetic sensors using
also supported by the ITRC support program. This work was also kalman lters and augmented system models for mobile robot navigation. In:
In IEEE int symposium on industrial electronics, Dubrovnik, Croatia; June
supported by the project The Development of Autonomous Navi- 2005. p. 15516.
gation Systems for a Robot Vehicle in Urban Environment at the [13] Dellaert F, Fox D, Burgard W, Thrun S. Monte Carlo localization for mobile
KIST. This work was also supported by Korea University during robots. In: IEEE intl conf on robotics and automation, Detroit, Michigan; May
1999. p. 13228.
the sabbatical year of the second author.
[14] Smith A, Doucet Arnaud, de Freitas Nando, Gordon Neil. Sequential Monte
Carlo methods in practice. 1st ed. Springer; 2001.
References [15] Borges G, Aldon MJ. An optimal pose estimator for map-based mobile robot
dynamic localization: experimental comparison with the EKF. In: IEEE
[1] Thrun S, Burgard W, Fox D. Probabilistic robotics. 1st ed. The MIT Press; 2005. international conference on robotics and automation, Seoul, Korea; May
[2] Borenstein J, Everett HR, Feng L. Where am I? Sensors and methods for mobile 2001. p. 158590.
robot positioning. University of Michigan, Department of Mechanical [16] Philippe Bonnifait, Pascal Bouron. Data fusion of four ABS sensors and GPS for
Engineering and Applied Mechanics, Mobile Robotics Laboratory, 1101 Beal an enhanced localization of car-like vehicles. In: IEEE international conference
Avenue, Ann Arbor, MI 48109. on robotics & automation, Seoul, Korea; May 2001. p. 1597602.
[3] Thrun S et al. Stanley: the robot that won the DARPA grand challenge: research [17] Kooktae Lee, Woojin Chung. Calibration of kinematic parameters of a car-like
articles. J Rob Syst 2006;23(9):66192. mobile robot to improve odometry accuracy. In: IEEE international conference
[4] Urmson C et al. A robust approach to high-speed navigation for unrehearsed on robotics and automation, Pasadena, CA; May 2008. p. 254651.
desert terrain. J Field Rob 2006;23(8):467508. [18] Corral Celestino A. On implementing Kasas circle t procedure. IEEE Trans
[5] Borenstein J, Feng L. Correction of systematic odometry errors in mobile Instrum Meas 1998;47(3):78995.
robots. In: IEEE international conference on intelligent robots and systems,
Pittsburgh, PA; August 1995. p. 56974.