See discussions, stats, and author profiles for this publication at: [Link]

net/publication/224296401

Experimental kinematics for wheeled skid-steer mobile robots

Conference Paper · December 2007

DOI: 10.1109/IROS.2007.4399139 · Source: IEEE Xplore

CITATIONS READS
207 7,840

6 authors, including:

Anthony Mandow Jorge L. Martínez

University of Malaga University of Malaga
99 PUBLICATIONS 2,341 CITATIONS 85 PUBLICATIONS 2,273 CITATIONS

SEE PROFILE SEE PROFILE

J. Morales Jose Luis Blanco

University of Malaga University of Almería
60 PUBLICATIONS 1,747 CITATIONS 171 PUBLICATIONS 3,744 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Anthony Mandow on 15 May 2014.

The user has requested enhancement of the downloaded file.

 Experimental kinematics for wheeled skid-steer mobile robots
Anthony Mandow, Jorge L. Martı́nez, Jesús Morales, José L. Blanco,
Alfonso Garcı́a-Cerezo and Javier González
Dpto. Ingenierı́a de Sistemas y Automática, Universidad de Málaga, 29013 Málaga, Spain
Email: amandow@[Link], Tel: (+34) 952 131407, Fax: (+34) 952 131413

Abstract— This work aims at improving real-time motion are not sensed by the motor encoders. Still, an effective kine-
control and dead-reckoning of wheeled skid-steer vehicles by matic model is necessary to perform on-board computations
considering the effects of slippage, but without introducing the for dead-reckoning and motion control in the short term.
complexity of dynamics computations in the loop. This traction
scheme is found both in many off-the-shelf mobile robots due Some authors have studied stability of wheeled skid-
to its mechanical simplicity and in outdoor applications due to steering with non-linear control techniques by explicitly
its maneuverability. In previous works, we reported a method considering dynamics and drive models [10] [11]. Moreover,
to experimentally obtain an optimized kinematic model for kinematics has been addressed in some works as the relation
skid-steer tracked vehicles based on the boundedness of the of linear and angular velocities with the position of the
Instantaneous Centers of Rotation (ICRs) of treads on the
motion plane. This paper provides further insight on this vehicle [12] [13]. However, these do not consider major skid
method, which is now proposed for wheeled skid-steer vehicles. effects, which arise at a lower level, in the relation between
It has been successfully applied to a popular research robotic drive velocities and vehicle velocities.
platform, Pioneer P3-AT, with different kinds of tires and A kinematic model that relates to identification of slip
terrain types. parameters from actual inertial readings was proposed in [14]
for velocity control of tracked vehicles. For wheeled skid-
I. I NTRODUCTION
steering, on-line adaptive control has been considered for
Wheeled skid-steering drive mechanisms are found in estimation of tire/ground friction of a simplified dynamic
many all-terrain vehicles, such as loaders, farm machinery, model [15].
mining and military. This traction scheme is also useful for The kinematic relation is explicitly expressed by [8] as
off-road mobile robots [1], with field applications such as a terrain-dependent 3 × 2 matrix of constant coefficients.
planetary exploration [2], land-mine detection [3] and rescue This matrix does not only capture wheel skid but also tire
[4]. Moreover, commercial robotic research platforms also pressure differences and transmission inaccuracies, although
employ this locomotion system [5]. no physical meaning is given to its elements. A custom-made
Steering is based on controlling the relative velocities of caster wheel trailer is used to measure the actual velocities
the left and right side drives, similarly to differential drive in order to derive the coefficients with linear regression
wheeled vehicles. However, since all wheels are aligned with techniques.
the longitudinal axis of the vehicle, turning requires wheel Kinematic equivalences between skid-steering and differ-
slippage. This locomotion system functions like that of a ential drive vehicles have been proposed for tracked vehi-
tracked vehicle. Tracked locomotion usually provides better cles. A simple experiment to directly obtain a symmetric
traction, but it is mechanically more complex [6]. model was presented in [16]. Moreover, constant kinematic
Wheeled skid-steering presents two major advantages over parameters are derived in [17] as optimized values for
alternative wheel configurations, such as Ackerman or axle- the Instantaneous Centers of Rotation (ICRs) of treads on
articulated. First, it is simple and robust in mechanical the plane by using laser scan-matching motion estimations.
terms. Second, it provides better manoeuverability, including These correspond to the position of ideal differential drive
zero-radius turning, using only the components needed for wheels for a particular terrain. This is based on the fact that
straight-line motion [2]. positions of tread ICRs are dynamics-dependent, but they
However, this locomotion scheme poses special difficulties lie within a bounded area at moderate speeds. Kinematic
when addressing motion control and odometry. Skid-steering identification also incorporates fine-tuning for misalignments
kinematics is not straightforward, since it is not possible to and other mechanical inaccuracies.
predict the exact motion of the vehicle only from its control This paper presents further work on the experimental
inputs. Thus, pure rolling and no-slip assumptions considered ICR kinematic method, which is now proposed for wheeled
in kinematic models for non-holonomic wheeled vehicles do skid-steer vehicles. It has been applied to Pioneer P3-AT, a
not apply in this case. commercial platform widely used as a test-bed for robotics
Nevertheless, we find that scarce work has been reported research. For instance, a fuzzy path tracker [12] and inde-
on this problem. Additional internal sensor, such as gyro- pendent control for each wheel to reduce odometry errors
scopes [7], inertial units [8] or a small passive trailer with [18] have been reported.
encoders [9] can be applied to detect heading changes that The major contributions presented by this paper are the
following: treads with respect to the robot frame (Vl , Vr ). Then, direct
• A geometric interpretation of the kinematic matrix in kinematics on the plane can be stated as follows:
[8] based on the geometric analogy with an ideal 
vx

 
differential drive model.  vy  = fd Vl
(1)
• An easy to reproduce experimental identification Vr
ωz
method based on external sensor motion estimation and
Genetic Algorithms to derive an approximate kinematic where v = (vx , vy ) is the vehicle’s translational velocity with
model for wheeled skid-steer vehicles. respect to its local frame, and ωz is its angular velocity.
• A kinematic analysis of the Pioneer P3-AT robot. When turning, the Instantaneous Center of Rotation (ICR)
Configuration parameters of the manufacturer’s implicit of the vehicle on the motion plane is expressed in local
kinematic model are related to the proposed method. coordinates as ICRv = (xICRv , yICRv ), as shown in Fig.
• Dead reckoning results obtained from the application of 2. Besides, the ICRs for the left and right treads can be
the experimental model with different wheel and terrain defined in the local frame as ICRl = (xICRl , yICRl ) and
configurations. ICRr = (xICRr , yICRr ), respectively. These result from the
Following this introduction, section 2 establishes kine- composition of the motion of the vehicle and that of linear
matic analogies of skid-steering with ideal differential drive. tread velocity (i.e., Vl or Vr ). It is known [17] that ICRl
Section 3 presents an experimental procedure for kinematics and ICRr lie on a line parallel to the local X axis that
identification with external sensors. Section 4 describes the also contains ICRv . Note that treads have the same angular
application to the P3-AT mobile robot as well as validation velocity ωz as the whole vehicle.
experiments. Section 5 is devoted to conclusions and ideas Y
for future work. Finally, acknowledgements and references
v
complete the paper. Vl Vr
ICRv ICR l ICR r
II. A NALOGY OF SKID - STEERING WITH DIFFERENTIAL
DRIVE ωz ωz

Dynamic models for skid-steering may result too costly for ωz

real-time motion control and dead-reckoning. Alternatively,
ωz X
this section discusses geometric relationships that can be
used instead.
When differential drive is applied to control a pair of pure
rolling wheels, no slippage occurs and each tread’s contact is
Vl Vr
limited to a single point. This assumption is usually adequate
for non-ideal wheels with a relatively small contact patch Fig. 2. Vehicle and tread ICRs on the plane.
(see Fig. 1.a). On the other hand, tracked vehicles with
differential drive epitomize skid-steering because of the large Thus, the geometrical relation between ICR positions
contact area of both treads (see Fig 1.b). A similar case and the vehicle’s translational and rotational velocities is
is represented by wheeled skid-steering, where each side’s expressed by:
tread consists of several contact patches that correspond to
mechanically linked wheels, as those within the dashed lines −vy
xICRv = (2)
in Fig 1.c-d. ωz
αl · Vl − vy
xICRl = (3)
ωz
Y
αr · Vr − vy
Y Y Y
xICRr = (4)
ωz
X
X X X
vx
yICRv = yICRl = yICRr = (5)
ωz
(a) (b) (c) (d)
where the nominal tread speeds have been affected by
correction factors (αl , αr ) to account for a number of fuzzy
Fig. 1. Left and right treads on the plane for differential drive (a),
tracked vehicle (b), four-wheel skid-steer vehicle (c), and six-wheel skid-
mechanical issues such as tire inflation conditions or the
steer vehicle (d). transmission belt tension.
The vehicle’s xICRv coordinate reaches infinite values
The local frame of the vehicle is assumed to have its origin when ωz = 0 in (2). Conversely, boundedness is an important
on the center of the area defined by the left and right contact property of tread ICR coordinates: In the proximity of
surfaces on the plane, and its Y axis is aligned with the straight line motion, numerators and denominators in (3)-
forward motion direction. Skid-steer vehicles are governed (5) are infinitesimals of the same order, which result in
by two control inputs: the linear velocity of its left and right finite values for xICRl , xICRr and yICRv , respectively. It
must be noted that this boundedness is limited to the case
of kinematic motion, in which centrifugal dynamics are Vl Vr
neglected and slippage is due only to steering. Interestingly,
it has been demonstrated that in order to achieve stability
Y
with fast motion (i.e., with some slippage contributed by
ICRl ICRr
centrifugal forces), it must be guaranteed that yICRv is
X
bounded [13]. Vl Vr
From Eqs.(2)-(5), the kinematic relation (1) can be ob-
tained as :
  (a)
vx  
 vy  = A · Vl
(6)
Vr
ωz Vl Vr
Y
where the elements of matrix A only depend on tread ICR ICRr
ICRl
coordinates and correction factors: X

 
−yICRv · αl yICRv · αr
1
A= ·  xICRr · αl −xICRl · αr  .
xICRr − xICRl
−αl αr (b)
(7)
Note that in the case of an ideal symmetrical kinematic
model (i.e., ICRs lie symmetrically on the local X axis and Fig. 3. Kinematic equivalence between the tread ICRs of a skid-steer
vehicle (a) and wheel contact points of an ideal differential drive vehicle
yICRv = 0), matrix A takes the following form: (b).
 
0 0
α
A= ·  xICR xICR  (8) ICRs:
2xICR
−1 1
L
χ= , (0 < χ ≤ 1) (9)
where α = αl = αr and xICR = −xICRl = xICRr . xICRr − xICRl
Besides, this locomotion system introduces a non- where L is the distance between tread centerlines. Index χ
holonomic restriction in the motion plane because the non- is equal to 1 when no slippage occurs (i.e., ideal differential
square matrix A has no inverse. drive). Besides, a normalized eccentricity index can be
It must be noted that the above expressions also represent defined as follows:
kinematics for ideal wheeled differential drive vehicles, as
illustrated by Fig. 3. Therefore, for instantaneous motion, xICRr + xICRl
e= (10)
kinematic equivalences can be considered between skid-steer xICRr − xICRl
and ideal wheel vehicles. which is equal to 0 when tread ICRs are symmetrical with
The difference between both traction schemes is that respect to the local Y axis.
whereas ICRs for single ideal wheels are constant and The major consequence of the study above is that the effect
coincident with the ground contact points, tread ICRs are of vehicle dynamics is introduced in the kinematic model
dynamics-dependent and always lie outside of the tread just by two points in local coordinates: ICRl and ICRr .
centerlines because of slippage. Thus, less slippage results These points are bounded variables for skid-steer vehicles
in tread ICRs that are closer to the vehicle. and coincide with wheel contact points for ideal differential
Significantly, the center of mass of a vehicle affects tread drive.
ICRs, as illustrated by Fig. 3. If it is closer to one side,
then that side’s wheels will slip less on account of pressure III. E XPERIMENTAL KINEMATICS IDENTIFICATION
and its ICR will be closer to the vehicle. Conversely, the A simple experiment, first proposed in [16], can be used to
opposite tread ICR will be farther. Furthermore, the area of derive a symmetric kinematic model. When equal opposite
the contact patch and its distribution also affect tread ICRs. tread speed control inputs Vr and Vl are issued, the vehicle
Thus, if the center of mass is closer to the front or rear exhibits an approximately pure rotation about its Z-axis.
ends of the vehicle, pressure distribution concentrates onto a Then, the following equation can be applied:
portion of the tread contact surface, which results in closer R R
Vr dt − Vl dt
tread ICRs and a displacement of the yICRv coordinate. In xICR ≈ , (11)
this sense, tire pressure also affects ICR positions. These 2·φ
effects can be quantified by stating appropriate indexes. where φ represents the actual rotated angle. Similarly, the
A steering efficiency index χ of the vehicle can be defined value α can be adjusted by measuring the actual traveled
as the inverse of the normalized distance between the tread distance d in straight motion:
 the crossover and mutation operators. As a result, constant
2d kinematic parameters can be optimized off-line for a specific
α≈ R R . (12)
Vr dt + Vl dt robotic task, according to typical path motions, particular soil
This simple procedure offers a kinematic solution that ac- types, and speed ranges.
counts for wheel slippage. However, it does not contemplate IV. C ASE STUDY: A PPLICATION TO THE P3-AT ROBOT
neither the asymmetric effects of the center of mass nor
A. The Pioneer P3-AT robot
mechanical misalignments.
Therefore, a more exhaustive experimental setup is pro- The P3-AT has a 4-wheel drive skid-steer platform. It
posed in order to obtain the complete kinematic relation measures approximately 0.5 m in length and is 0.25 m tall.
expressed by (7). In this case, the parameter identification Tread length (i.e., the distance between the front and rear
method comprises two steps [17]. The first is to obtain wheels) is 0.27 m. Two different wheels sets are provided
experimental data from both external and internal sensors. by the manufacturer: Solid rubber tires of 19 cm diameter
Then, an optimization tool is employed to find the parameter (intended mainly for indoor use), and pneumatic tires of
values that best fit the data. Optimization is required, since 22 cm diameter (see Fig. 4). The distance between longi-
actual tread ICR positions vary during navigation on account tudinal tread centerlines is L = 0.4 m for both tire types.
of dynamics. The drive systems use reversible-DC motors equipped with
Genetic algorithms (GAs) are proposed as an appropriate an optical quadrature shaft encoder. It has a maximum speed
search method for this problem. This is a derivative-free of 1 wheel revolution per second. The robot weights 23.6 kg
stochastic optimization tool, where each point in a solution and can carry a 35 kg payload. It can run over two hours on
space is encoded into a bit string (chromosome) and is three 12 V DC batteries.
associated with a fitness value. Starting from an initial
random population, points with better fitness values are used
to construct a new population of solutions by means of
genetic operators. Thus, after several iterations (generations),
the overall fitness value is improved.
GAs are useful for finding optimized values for the set
of kinematic parameters K = {xICRl , xICRr , yICRv , αr , αl }
from experimental data because no significant a priori knowl-
edge exists about them, it is straightforward to code them
as bit strings, and a fitness value can be assigned to new
solutions. Note that the number of parameters is one less
than the number of elements in matrix A. Fig. 4. Two tire types for the robot.
Data collection experiments consist of several representa-
tive paths on a particular soil type, with diverse rotational
and translational speeds and accelerations. Data logged from
internal readings (e.g., shaft encoders or inertial sensors)
are needed to replay odometric estimations according to
alternative kinematic parameters. External sensor readings
(e.g., from a laser scanner or Differential GPS) occur at a
lower rate, and are necessary to accurately determine the
actual motion of the vehicle. Thus, training paths can be
broken into a number N of short segments according to the
external sensor rate. This way, each segment contains several
consecutive internal readings in order to estimate a kinematic
pose increment (∆x̂, ∆ŷ, ∆φ̂) by integrating (6) with (7),
as well as an accurate motion increment (∆x, ∆y, ∆φ)
calculated from external readings. Fitness of the alternative
solutions evaluated by the GA can be assessed as the sum of
the squared odometric errors for the N recorded segments:

Fig. 5. DGPS base (foreground right) and P3-AT robot (left).

J(K) = (13)
XN   The robot receives orders from a PC and sends operational
(∆xi − ∆x̂i )2 + (∆yi − ∆ŷi )2 + (∆φi − ∆φ̂i )2 . information under a built-in client-server architecture. Odo-
i=1 metric estimations are calculated from an implicit symmetric
At each GA iteration, a set of solutions with better kinematic model, which can be adjusted by the user through
overall errors is chosen for computing a new population with a pair of integer parameters, namely Ticksmm and Revcount.
 pneumatic (50psi) pneumatic (20psi) solid rubber
B. Kinematic parametrization yICRv -0.0080 -0.0120 -0.0120
There are two alternative ways to implement approximate xICRr 0.2998 0.3071 0.2588
xICRl -0.2758 -0.2553 -0.2667
kinematics of the P3-AT robot for motion control and dead- αr 0.9253 0.9049 0.9128
reckoning. First, the server incorporates a symmetric model, αl 0.9464 0.9271 0.9047
as in (8), that can be specified through the Ticksmm and χ 0.6949 0.7112 0.7612
e 0.0417 0.0921 -0.0150
Revcount parameters. Secondly, an asymmetric model (7)
can be implemented on the client by directly processing tick TABLE I
counts from the encoders. Both approaches are discussed E STIMATED KINEMATIC PARAMETERS AND INDEXES FOR ASPHALT.

below.
1) Server model parametrization: The Ticksmm parameter
pneumatic (50psi) pneumatic (20psi) solid rubber
describes the number of encoder ticks per millimeter. It yICRv -0.0159 -0.0128 -0.0139
depends on wheel radius r, expressed in millimeters, and xICRr 0.2758 0.2899 0.2703
is affected by the correction factor α: xICRl -0.2878 -0.2772 -0.2659
αr 0.9418 0.9183 0.9139
  αl 0.9457 0.9222 0.9047
138 χ
T icksmm = round α · r · , (14) 0.7098 0.7053 0.7460
110 e -0.0213 0.0224 0.0082

where 138/110 is a constant value that depends on encoders TABLE II

and transmission gears. Besides, Revcount represents the E STIMATED KINEMATIC PARAMETERS AND INDEXES FOR SMOOTH

number of ticks for a complete revolution of the robot turning CONCRETE .

in place, which depends on slippage:

 
2π · xICR · T icksmm
Revcount = round , (15) have been recorded every 10 ms. Each parameter has been
8
coded as an 8 bit string, yielding 40 bit chromosomes for
where the denominator scales the binary range of the integer. the GA search. Tables I and II present kinematic parameters
The factory values for these parameters yield a default obtained for asphalt and smooth concrete flat terrains, respec-
symmetric kinematic model with α = 1 and xICR = 0.3 m tively. In each case, three different wheel configurations have
(assumed for the pneumatic tires). Note that a non-slip model been considered: pneumatic tires at 20 psi (nominal value)
would have xICR = 0.2 m. and 50 psi (maximum pressure), and solid rubber. Steering
2) Client model implementation: The server does not pro- efficiency and eccentricity indexes have also been computed.
vide the means to specify the complete asymmetric model, so Independently of terrain and wheel configurations, yICRv
it has to be implemented on the client PC. Dead reckoning is always about 1 cm behind of the frame origin, which
can be computed as (6) with (7) by estimating Vr and Vl confirms that it only depends on the center of mass. The
from tick counts read from the server. loss of thrust power while turning is greater in asphalt than
As for motion control, the client has to compute the driver in concrete due to greater friction. This can be observed as
velocities (in mm/s) that are to be sent to the robot with a lower steering efficiency value χ. In this sense, compact
the setvel2(Vl , Vr ) command. Control inputs can be obtained wheels provide better efficiency than pneumatic tires (but less
from (3) and (4) as: traction). Note also that the α values are closer to 1 for 50 psi
than for 20 psi because the effective wheel radius is closer
vy + xICRl · ωz to the nominal value. The results also show no significant
Vl = (16)
αl asymmetry with respect to the longitudinal axis Y , since the
vy + xICRr · ωz eccentricity values are always very close to zero.
Vr = , (17)
αr
where vx references cannot be directly addressed due to the D. Dead-reckoning validation
non-holonomic restriction of the locomotion system.
Validation experiments have been carried out for the
C. Experimental Identification six navigation conditions discussed above. Fig. 6 shows a
The P3-AT has been furnished with a 15.8 kg structure representative example path as estimated by three alternative
that houses the onboard GPS systems as well as a laptop kinematic models: the default P3-AT symmetric model, an
computer (see Fig. 5). This structure has been considered experimental symmetrical model for asphalt and 20 psi
not to affect the center of gravity of the robot significantly. pneumatic tires (α = 0.91 and xICR = 0.275 m) obtained
The onboard Javad GPS receiver provides a precision under from (11) and (12), and the corresponding optimized asym-
1 cm by accepting RTCM/RTK differential corrections from metric model. The ICR-based model clearly improves dead-
a nearby GPS base station at a rate of 5 Hz. The portable reckoning as compared with accurate path positions provided
computer acts as the client in the robotic architecture. by the DGPS. Mean squared performance values obtained
The experiments for kinematic identification have con- from all segments in the validation path are shown in Table
sisted on several joystick controlled paths, where tick counts III.
 GPS data
11
Asymmetric model
ACKNOWLEDGMENTS
Symmetric model
10 P3−AT model This work was partially supported by the Spanish CICYT
9
projects DPI 2005-01391 and DPI 2005-00207.
8 R EFERENCES
7 [1] L. Champeny-Bares, S. Coppersmith, and K. Dowling, “The Terregator
mobile robot,” Carnegie Mellon University, Pittsburgh, Pennsylvania,
Tech. Rep. CMU-RI-TR-93-03, 1991.
y (m)

6
[2] B. Shamah, M. D. Wagner, S. Moorehead, J. Teza, D. Wettergreen, and
5 W. Whittaker, “Steering and control of a passively articulated robot,”
in SPIE Sensor Fusion and Decentralized Control in Robotics Systems
4
IV, 2001.
3 [3] K. Wedewad, S. Bruder, V. Yodaiken, and J. Guilberto, “Low-cost
outdoor mobile robot: a platform for landmine detection,” in Proc.
2 IEEE 42nd Midwest Symposium on Circuits and Systems, 1999, pp.
131–134.
1 [4] A. Chetchetka, S. Hughes, R. Glinton, M. Koes, M. Lewis, I. Nour-
bakhsh, D. Rosenberg, K. Sycara, and J. Wang, “Robocup rescue robot
−12 −10 −8 −6 −4 −2 0 league team raptor (USA),” in Proc. RoboCup US Open 2005 Rescue
x (m) Robot League Competition, Atlanta, Georgia, USA, 2005.
[5] Pioneer 3 and Pioneer 2 H8-Series Operations Manual, version 3,
Fig. 6. Validation path for asphalt and 20 psi pneumatic tires. ActivMedia Robotics, LLC, Amherst, NH, August 2003.
[6] J. Y. Wong and W. Huang, “Wheels vs. tracks: A fundamental
Asymmetric P3-AT Symmetric evaluation from the traction perspective,” Journal of Terramechanics,
mean squared ∆x error 0.00010 0.00183 0.00162 vol. 43, pp. 27–42, 2006.
mean squared ∆y error 0.00011 0.00468 0.00455 [7] H. Chung, L. Ojeda, and J. Borenstein, “Accurate mobile robot dead-
mean squared ∆φ error 0.0359 0.0359 0.0359 reckoning with a precision-calibrated fiber-optic gyroscope,” IEEE
mean squared J error 0.0361 0.0424 0.0421 Transactions on Robotics and Automation, vol. 17, no. 1, pp. 329–
336, 2001.
TABLE III [8] G. Anousaki and K. J. Kyriakopoulos, “A dead-reckoning scheme for
M EAN PERFORMANCE VALUES FOR ALL SEGMENTS IN THE VALIDATION skid steered vehicles in outdoor environments,” in Proc. IEEE Int.
Conf. on Robotics and Automation, New Orleans, LA, 2004, pp. 580–
PATH .
585.
[9] Z. Fan, J. Borenstein, D. Wehe, and Y. Koren, “Experimental evalua-
tion of an encoder trailer for dead-reckoning in tracked mobile robots,”
in Proc. of the IEEE International Symposium on Intelligent Control,
V. C ONCLUSIONS Monterey, CA, 1995, pp. 571–576.
[10] L. Caracciolo, A. D. Luca, and S. Iannitti, “Trajectory tracking control
The paper proposes a method for improving real-time of a four-wheel differentially driven mobile robot,” in Proc. of the
motion control and dead-reckoning of wheeled skid-steer IEEE Int. Conf. on Robotics and Automation, Detroit, USA, 1999, pp.
2632–2638.
vehicles without introducing the complexity of dynamics [11] K. Kozlowski and D. Pazderski, “Modelling and control of a 4-wheel
computations in the loop. The method experimentally derives skid-steering mobile robot,” Int. J. Appl. Math. Comput. Sci., vol. 14,
an approximate kinematic model based on the boundedness no. 4, pp. 477–496, 2004.
[12] E. Maalouf, M. Saad, and H. Saliah, “A higher level path tracking con-
of the Instantaneous Centers of Rotation of both treads troller for a four-wheel differentially steered mobile robot,” Robotics
on the motion plane. The model relates drive and vehicle and Autonomous Systems, vol. 54, pp. 23–33, 2006.
velocities by considering the effects of slippage. Moreover, [13] K. Kozlowski and D. Pazderski, “Practical stabilization of a skid-
steering mobile robot - a kinematic-based approach,” in Proc. of
it incorporates fine-tuning parameters for misalignments and the IEEE 3rd International Conference on Mechatronics, Budapest,
other hidden mechanical inaccuracies. Hungary, 2006, pp. 519–524.
The major contributions of the paper are the following: [14] A. Le, D. Rye, and H. F. Durrant-Whyte, “Estimation of track-soil
interactions for autonomous tracked vehicles,” in Proc. of the IEEE
• The geometric analogy with an ideal differential drive Int. Conf. on Robotics and Automation, vol. 2, Albuquerque, USA,
model provides a physical interpretation to the coeffi- 1997, pp. 1388–1393.
cients of the kinematic matrix presented by [8]. [15] J. Yi, D. Song, J. Zhang, and Z. Goodwin, “Adaptive trajectory track-
ing control of skid-steered mobile robots,” in Proc. IEEE International
• An easy to reproduce experimental method based on ex- Conference on Robotics and Automation, Rome, Italy, 2007, pp. 2605–
ternal sensor motion estimation and Genetic Algorithms 2610.
to identify an approximate kinematic model for wheeled [16] S. Pedraza, R. Fernández, V. Muñoz, and A. Garcı́a-Cerezo, “A motion
control approach for a tracked mobile robot,” in Proc. of the 4th IFAC
skid-steer vehicles. International Symposium on Intelligent Components and Instruments
• A kinematic analysis of the Pioneer P3-AT robot. Con- for Control Applications, Buenos Aires, Argentina, 2000, pp. 147–152.
figuration parameters of manufacturer’s implicit kine- [17] J. L. Martı́nez, A. Mandow, J. Morales, S. Pedraza, and A. Garcı́a-
Cerezo, “Aproximating kinematics for tracked mobile robots,” The
matic model have been related to the proposed method. International Journal of Robotics Research, vol. 24, no. 10, pp. 867–
• Analysis of experimental dead reckoning results from 878, 2005.
the application of the experimental model with different [18] L. Ojeda and J. Borenstein, “Reduction of odometry errors in over-
constrained mobile robots,” in Proc. of the UGV Technology Confer-
wheel and terrain configurations. ence at the 2003 SPIE AeroSense Symposium, Orlando, FL, 2003.
For future work, it would be interesting to integrate this
enhanced kinematic model with inertial measurements for
irregular terrains.

Vie