See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/4296941

Validation of soil parameter identification for track-terrain interaction

Dynamics

Conference Paper · October 2007

DOI: 10.1109/IROS.2007.4399018 · Source: IEEE Xplore

CITATIONS READS
6 80

4 authors, including:

Yahya Hashem Zweiri Kaspar Althoefer

Kingston University London King's College London
97 PUBLICATIONS   877 CITATIONS    435 PUBLICATIONS   5,551 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

MAP armoured vehicle View project

Automated manufacturing View project

All content following this page was uploaded by Yahya Hashem Zweiri on 19 November 2014.

The user has requested enhancement of the downloaded file.

Proceedings of the 2007 IEEE/RSJ International ThA11.3
Conference on Intelligent Robots and Systems
San Diego, CA, USA, Oct 29 - Nov 2, 2007

Validation of Soil Parameter Identification for Track-Terrain

Interaction Dynamics
S Hutangkabodee, Y H Zweiri, L D Seneviratne, and K Althoefer

Abstract— This paper considers a tracked vehicle traversing even though these empirical algorithms (Vehicle Cone Index
unknown terrain, and proposes an approach based on the (VCI) method, Mean Maximum Pressure (MMP) method, etc
Generalized Newton Raphson (GNR) method for identifying [8], [9]) are widely used, they have some drawbacks. Firstly,
all the unknown soil parameters required for tractive force
prediction. For the first time, the methodology, based on if either a tracked vehicle specification or a terrain type does
measurements of track slip, i, and tractive force, F , to find not match those in the database, its performance cannot be
unknown soil parameters is developed. The tractive force is the predicted. Secondly, tracked vehicle performance cannot be
force generated by a tracked vehicle to drive itself forwards. predicted in real-time as the terrain type needs to be known
This tractive force depends to a large extent on certain soil prior to the operation of the tracked vehicle.
parameters, namely soil cohesion (c), soil internal friction
angle (φ), and soil shear deformation modulus (K). Accurately One approach to improve the performance and autonomy
identifying parameters of the soil on which a tracked vehicle is of tracked vehicles is to acquire information of the terrain in
moving will potentially lead to accurate traversability predic- real-time. The real-time acquisition of soil parameters based
tion, effective traction control, and precise trajectory tracking. on a physical model and numerical techniques is required
The soil parameter identification algorithm is validated with for autonomous trajectory tracking, traversability prediction
the experimental data from Wong [3] and from in-house track-
terrain interaction test rig showing good identification accuracy and traction control [10].
and fast execution speed. It is also shown to be relatively robust There has been increasing interest in parameter identifica-
to initial condition. The identified soil parameters are, in turn, tion in various engineering applications [11] - [16]. In [13], a
used to predict the tractive forces showing good agreement with Linear Least Square estimator is employed on-line to identify
all the experimental data. The technique presented in this paper two key soil parameters using on-board vehicle sensors. This
is general and can be applied to any tracked vehicle.
estimator is based on a simplified linearized model of the
I. I NTRODUCTION vehicle wheel-terrain interaction dynamics. Hutangkabodee
Tracked vehicles have many potential applications, in- [15] extended it further to identify three soil parameters
cluding space exploration, defense, agriculture, mining and based on the application of the Newton Raphson method
construction. Research into the traversability of off-road to non-linearized wheel-terrain interaction dynamics.
vehicles has been carried out since Bekker first pioneered In this paper, for the first time, an approach, based on
this field [1], [2]. Empirical relationships for vehicle-terrain measurements of tractive force, F , and track slip, i, to
interactions for both tracked and wheeled vehicles have been find unknown soil parameters is developed. A track-terrain
developed based on Mohr-Coulomb theory [1] - [4]. In [5], interaction dynamic model is used for soil parameter iden-
a slip-based traction force model is developed theoretically tification. The paper presents an algorithm, based on the
and is used to establish an effective control law for a vehicle Generalized Newton Raphson (GNR) method, to identify
travelling on rough terrain. The slip estimation for a tracked all the soil parameters required for tractive force prediction.
vehicle from trajectory measurements is achieved using an The algorithm is validated employing experimental data from
Extended Kalman Filter [6] and a sliding-mode observer [7]. Wong [3] and a specially designed test rig for tractive force
The slip estimates, together with knowledge of the terrain and slip measurement.
parameters, allow the prediction of vehicle tractive forces.
II. G OVERNING M ODEL
These forces are essential for vehicle traversability prediction
and traction control. The semi-empirical model proposed in [3] to predict
Most of the research on tracked vehicle performance on track-terrain interaction dynamics is employed in this paper.
deformable terrain have been developed empirically based on This model characterizes the interaction forces between the
experiments on vehicles traversing various types of terrain. vehicle tracks and a general terrain including loose sand,
A database of these empirical data has been created and saturated clay, dry fresh snow, and most dispersed soils. The
used as a tool for tracked vehicle performance prediction; tractive force for a tracked vehicle traversing a general terrain
is given by:
This work receives financial support from EPSRC (GR/S31402/01)
S. Hutangkabodee, L.D. Seneviratne, and K. Althoefer are ·
with the Department of Mechanical Engineering, King’s College K³ ´¸
London, UK [email protected], F = (Ac + W tan φ) 1 − 1 − e−il/K , (1)
[email protected], il
k.althoefer.kcl.ac.uk
Y.H. Zweiri is with the Department of Mechanical Engineering, Univer- where A is the contact area of the tracks, W is the normal
sity of Mu‘tah, Karak, Jordan [email protected] load due to the weight of the tracked vehicle, i is the track

1-4244-0912-8/07/$25.00 ©2007 IEEE. 3174

slip, c is the soil cohesion, φ is the soil internal friction angle, Thus given the measurement vector xi ∈ Rm , (3) provides
and K is the soil shear deformation modulus. a single equation with n unknown elements of p. For the
It is noted that the parameters c and φ only depend on the GNR method, to find p, q independent equations are required
soil and are thus constant for a given type of soil whereas (q > n). These equations are generated by evaluating the
parameter K depends not only on the type of soil but also on function f at q different time samples (t1 , t2 , ..., tq ). This
the normal soil pressure which is dependent on the vehicle results in q non-linear equations:
(particularly on the vehicle’s weight) [17]. However, for a
given vehicle traversing a given terrain, all three parameters, f1 (p1 , p2 , ..., pn , x1 (t1 ) , x2 (t1 ) , ..., xm (t1 )) = 0
c, φ, and K, can be considered constant. These parameters f2 (p1 , p2 , ..., pn , x1 (t2 ) , x2 (t2 ) , ..., xm (t2 )) = 0
are sufficient to characterize an unknown terrain for the .. ..
purpose of tractive force prediction for a tracked vehicle. . .
We further assume that the tracked vehicles under consid- fq (p1 , p2 , ..., pn , x1 (tq ) , x2 (tq ) , ..., xm (tq )) = 0
eration have very long and wide tracks so that vehicle sinkage (4)
is relatively small, in particular when compared to wheeled Expanding (4) about pi using a Taylor series expansion
vehicles where sinkage can have a profound influence on the and neglecting second and higher order terms, the GNR
model and can vary drastically during operation. Based on identification algorithm can be rewritten in matrix form as:
this assumption, which is valid for a large number of real
tracked vehicles operating on different outdoor grounds, the      ∂f1 ∂f1
−1  
p1 p1 ∂p1 ··· ∂pn f1
sinkage term has not been taken into account in the proposed ∂f2 ∂f2
p2 p2   ··· f2

model (1).
    
  ∂p1 ∂pn 
 .
    
It is noted that for a given tracked vehicle traveling on a  ..  ≈ ..  − .. .. ..   ..
. .   . . . .
     
 
given terrain type, the term (Ac + W tan φ) is a constant, pn pn ∂fq ∂fq fq
i+1 i ∂p1 ··· ∂pn i
and (1) can be rewritten as i
(5)
Let p0 be an initial estimate of the unknown parameter
· ´¸
K³
F =L 1− 1 − e−il/K , (2) vector p. Then, (5) can be iteratively applied until p con-
il
verges to its true value.
where L = (Ac + W tan φ).
The GNR method has the same advantages as those of the
This model is used in this paper for soil parameter identifi- NR method, namely its ability to identify unknown individual
cation based on vehicle slip and tractive force measurements. parameters, its robustness to noise and its fast speed of
Once L and K are accurately identified, the tractive forces convergence. In addition, the robustness of the GNR method
produced by the vehicle on a given terrain over the entire is even better than that of the NR method [11] because it
slip range can accurately be predicted. considers more equations than the number of unknowns.
III. S OIL PARAMETER I DENTIFICATION BASED ON THE However, it is noted that if the initial estimate is selected
G ENERALIZED N EWTON R APHSON (GNR) M ETHOD at a point where the derivative of the function is relatively
small, the convergence speed may be slow [18].
The proposed soil parameter identification approach for a
tracked vehicle traversing unknown terrain is based on the
Generalized Newton Raphson (GNR) method. This method
is well suited for real-time parameter identification since soil
parameters vary only slowly during terrain traversing.
The concept of the GNR method is based on the Newton
Raphson (NR) method. The NR method works by itera-
tively modifying an initial estimate and, after a number
of iterations, arrives at a converged solution. Commonly,
the iterative process is halted when the error between the
identified value and the current value falls below a pre-
defined threshold. The main difference between the GNR and
the NR methods is that the NR method employs the same
number of equations as the number of unknowns whereas
the GNR method uses more equations than the number of Fig. 1. Diagram showing implementation of the GNR method for soil
parameter identification
unknowns. The GNR algorithm is described below.
Assuming that there are n unknown system parameters in
vector p ∈ Rn , the system equation can then be expressed The GNR method is implemented to identify the soil
as a function of p and measurement vector xi ∈ Rm at time parameters for vehicle track-terrain interaction dynamics,
ti , Fig. 1. Here, vector p comprises two parameters, the lumped
parameter (L) and the shear deformation modulus (K). The
f (p, xi ) = 0. (3) vector x consists of two measured signals, tractive force,

3175
F and slip, i. The track-terrain interaction dynamic model
described by (2), and the measurement vector x are used to
identify the unknown soil parameters. Subsequently, (3) can
be expressed for time t1 , t2 , ..., tq as:
 
f1 (L, K, F1 , i1 )
 f2 (L, K, F2 , i2 ) 
 = 0, (6)
 
 ..
 . 
fq (L, K, Fq , iq )
T T
where p = [L,
· K] ¸, and x = [F, i] .
L
Let p0 = be an initial estimate of the unknown
K 0
parameters. Applying the GNR method to (6, (5) becomes:

  Fig. 2. Track-terrain interaction test rig, where A is the track assembly, B is

f1 (L, K, F1 , i1 ) the carriage, C is the chain connected to the carriage, D is the chain motor-
gearbox-encoder to drive the chain, E is the pillar parts of track assembly,
·
L
¸ ·
L
¸  f2 (L, K, F2 , i2 ) 
F is the test rig frame, G is the soil box, and H is the experimental soil
= − J −1  , (7)
 
K K ..
0  . 
fq (L, K, Fq , iq )
 ∂f1 ∂f1  Slip estimation is made possible through recent advances
∂L
∂f2
∂K
∂f2
employing a sliding-mode observer whose inputs are track
sprocket angular speeds and vehicle heading speed [7]. The
 
∂L ∂K
where J −1 is the pseudo inverse of J = 
 
.. ..
track sprocket angular speeds can be measured by using shaft

 . . 
∂fq ∂fq encoders on the vehicle’s sprockets. The vehicle heading
∂L ∂K 0
since J is not a square matrix. speed can be obtained using several techniques such as
Differential Global Positioning System (DGPS) [20], Inertial
IV. T EST R IG AND E XPERIMENTATION P ROCEDURE Navigation System (INS) [21], and Optical Flow technique
[22].
The track-terrain interaction test rig used in this study is
The test rig consists of a driven horizontal carriage
shown in Figs. 2 and 3. The idea to develop this test rig
mounted to a driven track, Fig. 2. The track is not vertically
arose from the fact that we were interested in exploring the
constrained by the carriage and can move freely in the
proposed approach in a controlled environment allowing us
vertical direction (to allow the driven track to follow the
to acquire all needed parameters in a simple and straightfor-
soil surface freely). During experimentation, two motors -
ward manner and to prove the feasibility and effectiveness
one driving the track and another driving the carriage - are
of the GNR-based soil parameter identification approach.
operated at different speeds to generate track slip. By varying
In particular, it is difficult to measure slip parameters on
the differential speeds between the track and the carriage
a real vehicle manoeuvring through terrain. Based on the
motors, various track slip values can be generated.
current developments in slip estimation ([6], [7] and [20] -
[22]), we assume slip to be measurable on-line enabling the
proposed soil parameter estimation process. With this test rig,
the track slip, i, and the tractive force, F , can be measured
very accurately. This allows the soil parameter identification
algorithm to be validated in an effective way.
In practice, the signals needed for soil parameter identifi-
cation in a real tracked vehicle can be acquired as follows. To
measure the tractive force (F ) produced from diesel-engine-
driven tracked vehicle, a technique based on a non-linear
observer is used [19]. Here, the indicated torque produced
from the diesel engine is estimated from input variables
measured from a number of low-cost sensors which can
be easily retrofitted to the vehicle. These sensors measure
instantaneous angular velocities and angular displacements
of the engine crankshaft as well as the dynamometer cou- Fig. 3. Track assembly part of Track-terrain interaction test rig, where I is
pling. The vehicle tractive force is then obtained by dividing the gearbox, J is the DC motor, K is the optical encoder, L is the sprocket
wheel, M is the track belt, N is the sprocket-wheel-distance adjuster, and O
this estimated indicated torque by the radius of the sprocket is the slot for distance adjuster
wheel of the track.

3176
 The test rig is equipped with various sensors to acquire Iron sand, and Garise 14/25 soil respectively, acquired from
the measured signals needed for soil parameter identification. track-terrain test rig (data set B).
Two optical encoders are used to measure the angular speeds
TABLE III
of the track and carriage motors. The track slip can then be
I DENTIFICATION RESULTS FOR G ARSIDE I RON SAND (S ET B)
calculated using
rt ωt Soil Parameters Actual Values Identified Relative Error
i=1− , (8) Values (%)
rc ωc L (N) 50 - 80 56.82 within range
where i is the track slip, rt is the radius of track sprocket K (m) 0.01 - 0.025 0.01043 within range
Elapsed time (s) 0.015
wheel, ωt is the angular speed of the track sprocket wheel,
rc is the effective radius of the chain wheel, and ωc is the
angular speed of the chain wheel. TABLE IV
Tractive force F is calculated from F = T /r where r is I DENTIFICATION RESULTS FOR G ARSIDE 14/25 SOIL (S ET B)
the radius of the sprocket wheel. Sprocket wheel torque (T )
is acquired from the sprocket wheel motor input current (I) Soil Parameters Actual Values Identified Relative Error
Values (%)
using torque-current, T − I, relationship. L (N) 59.55 59.37 0.30
The track used in the experiment is 50-mm in width, 115- K (m) 0.01 - 0.025 0.008456 15.44
mm in length (between the two sprocket wheel centers), and Elapsed time (s) 0.015
100-mm sprocket wheel diameter. The track sprocket wheel
maximum angular speed is 1.5 rad/s which gives a track The identification results for measured data set A are
circumferential speed of 0.15 m/s (without slip). The test rig shown in Table I and those for measured data set B are
is built in a 1m × 1m × 0.5m frame. The experimental soil is depicted in Tables II–IV for Garside 60 soil, Garside Iron
contained in a 0.1m × 0.85m × 0.1m soil box. Experiments sand, and Garside 14/25 soil respectively. The identification
were performed on three types of soils - Garside 60 soil, errors for all the estimates with respect to the actual soil
Garside Iron sand, and Garside 14/25 soil. These soils were parameters are also shown in these tables with the execution
purchased from the Garside Sands company, UK. time. The identification algorithms were run on an Intel
Pentium(R) 4 processor with a 2.80 GHz CPU and 1.00
V. E XPERIMENTAL R ESULTS GByte of RAM.
In this section, the soil parameter identification results are According to [3], K varies from 0.01 m for firm sand to
presented based on two sets of measured data. The first set 0.025 m for loose sand. Parameter K of all the soils used
(set A) is the measured data from Wong’s experiments using in the test rig, Tables II–IV, falls within this range as all of
a real-size tracked vehicle [3]. The second set (set B) is from them are of sand nature with varying degree of compaction
the track-terrain interaction test rig described in Section IV. (varying in firmness). K is not given a specific value since its
The lumped term, (Ac + W tan φ) and shear deformation value is dependent on soil normal pressure (due to vehicle’s
modulus, K are the two unknown parameters to be identified. weight and track size) [17]. Values of c and φ for all three
soils were obtained employing a standard shear box test.
TABLE I
However, for the only wet soil used (Garside Iron sand, Table
I DENTIFICATION RESULTS USING DATA FROM W ONG [3] (S ET A)
III), its moisture content changes during the experimentation
Soil Parameters Actual Values Identified Relative Error (sand getting dried up) leading to changes in parameters c
Values (%) and φ (subsequently L). Therefore, for Garside Iron sand, its
L (kN) 278.75 280.69 0.70 actual L value is also given in range (50 - 80 N).
K (m) 0.0180 0.01924 6.89
Elapsed time (s) 0.015
It is seen from Table I that the accuracy of identification
is very good, with relative errors 0.70% for L and 6.89%
for K. For the results from the test rig, Tables II–IV, the
TABLE II identification accuracy is relatively good for Garside 60 soil
I DENTIFICATION RESULTS FOR G ARSIDE 60 SOIL (S ET B) and Garside 14/25 soil (9.37% and 0.30% for L and 0% and
15.44% for K, respectively) whereas, for Garside Iron sand,
Soil Parameters Actual Values Identified Relative Error both L and K are in range. The speed of convergence of
Values (%)
L (N) 64.15 70.16 9.37 the GNR method is fast with a 0.015 s execution time for all
K (m) 0.01 - 0.025 0.01652 within range the tests. The execution time can be further reduced if the
Elapsed time (s) 0.015 code is optimized. In this paper, the code is created using
Matlab 6.5 running in command mode; the execution time
From all the measured data obtained from experiments, the can be significantly reduced, if an optimized code is exe-
entire sets of measured F and i are used for soil parameter cuted. Therefore, the GNR method has promising potential
identification based on the GNR method. There are 18 sets for on-line soil parameter identification during track-terrain
of measured data taken from Wong [3] (data set A) and interactions.
there are 20, 17, and 18 sets for Garside 60 soil, Garside The tests to check the robustness to initial conditions

3177
 TABLE V 60
T HE RANGE OF INITIAL ESTIMATES OF SOIL PARAMETERS THAT
PRODUCES THE CONVERGED SOLUTION FOR W ONG DATA (S ET A) 50

Tractive force, F (N)

40
Lower bound Soil paramters Upper bound
0 (35) L (kN) 1000 (490)
0.001 (0.006) K (m) 0.051 (0.05) 30

TABLE VI 20

T HE RANGE OF INITIAL ESTIMATES OF SOIL PARAMETERS THAT Predicted data from

10 identified soil parameters
PRODUCES THE CONVERGED SOLUTION FOR TEST RIG DATA (S ET B) Measured data

0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
Soil types Lower bound Soil paramters Upper bound Slip, i
All soil types 0 (9) L (N) 1000 (468)
Garside 60 0.001 (0.006) K (m) 0.029 (0.05) Fig. 5. Comparison between measured and predicted tractive force using
Garside iron 0.001 (0.006) K (m) 0.021 (0.05) the GNR method for measured data set B on Garside 60 soil (test rig
Garside 14/25 0.001 (0.006) K (m) 0.019 (0.05) experiment)

50
(initial guesses) of the soil parameter identification using
45
GNR method are conducted on both Wong’s measured data
40
and track-terrain test rig’s data. Tables V and VI give the
35
ranges of initial conditions for L and K that allow successful

Tractive force, F (N)

convergence of the identified soil parameters. It is noted 30

that values in brackets in Tables V and VI are the practical 25

range of the soil parameters in real world according to [3]. 20

From [3], the practical range of c and φ are [0 - 68.5 kPa] 15

and [6 - 40.1 degrees], respectively. The practical range for 10 Predicted data from
identified soil parameters
L is calculated by substituting the practical c and φ into 5
Measured data

Ac + W tan φ. 0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7
From Table V, the range of effective initial estimates Slip, i
covers the practical range for both L and K, thus, showing
good robustness of the identification algorithm to the initial Fig. 6. Comparison between measured and predicted tractive force using
the GNR method for measured data set B on Garside Iron sand (test rig
conditions for data Set A. However, for Data Set B, even experiment)
though the algorithm shows good robustness for L, it is
not quite robust for K as the upper bound of effective
initial guess for K falls below the upper bound of the
practical K by quite a big margin. From this observation, values with the measured data set A [3], Fig. 4. The rms
it is then recommendable to choose initial guess for K near error in this case is 6.54 kN over the 270 kN span of F over
its practical lower bound, e.g. 0.01 m. entire slip range. This shows a very good prediction of F
over the entire slip range using the soil parameters identified.
300 Similarly, the soil parameter identification results calculated
from experimental data using the test rig are used to predict
250 the tractive force, Figs. 5–7. Over an approximately 50 N
span of F , the rms errors are 4.64 N, 2.54 N, and 8.87 N for
Tractive Force, F (kN)

200
Garside 60 soil, Garside Iron sand, and Garside 14/25 soil
150 respectively. This again reflects good prediction accuracy of
F for the test rig experiments. The relatively high prediction
100 error of Garside 14/25 soil (rms error = 8.87 N) results from
Predicted data from
identified soil parameters
its experimental data being more spread than those of the
50 Measured data
other two soils.
0 Thus by measuring tractive force samples at a few slip
0 0.05 0.1 0.15 0.2 0.25
Slip, i values, soil parameters can be identified and used to predict
the tracked vehicle tractive forces over the entire range of
Fig. 4. Comparison between measured and predicted tractive force using the slips. Hence, accurate tractive force prediction of a vehicle
GNR method for measured data set A (real-size tracked vehicle experiment)
over a particular terrain can be achieved. Consequently, using
this approach it is known how much tractive force the tracked
Next, the estimates of L and K (Table I) are used to predict vehicle will produce over a particular terrain when a certain
the tractive force and compare the resultant tractive force slip is induced. This directly benefits traction control of a

3178
 70
[3] J.Y. Wong, “Theory of Ground Vehicles”, Third Edition, John Wiley
& Sons, Inc. 2001.
60
[4] J.Y. Wong, “Terramechanics and Off-Road Vehicles”, Springer, Else-
vier Science Publishers B.V. , Netherlands, 1989.
50
[5] K. Yoshida, H. Hamano, “Motion Dynamics of a Rover With Slip-
Tractive force, F (N)

Based Traction Model”, IEEE International Conference on Robotics

40
and Automation, Vol. 3, pp. 3155–3160, 2002.
[6] Anh Tuan Le, “Modeling and Control of Tracked Vehicles”, Ph.D.
30 thesis, University of Sydney, 1999.
[7] Z. Song, Y.H. Zweiri, L.D. Seneviratne, K. Althoefer, “Driver Support
20 System Based on a Non-Linear Slip Observer for Off Road Vehicles”,
Predicted data from 16th IFAC World Congress, Prague, Czech Republic, 2005.
identified soil parameters
10 Measured data [8] B. Maclaurin, “The Soft-Soil Performance of Wheeled and Tracked
Vehicles”, Journal of Defence Science, Vol. 5, No. 4, pp. 380–385,
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9
2000.
Slip, i [9] M. Geeves, “The Mobility of Tracked and Wheeled Vehicles”, Journal
of Defence Science, Vol. 5, No. 4, pp. 364–374, 2000.
Fig. 7. Comparison between measured and predicted tractive force using [10] Z. Fan, Y. Koren, D. Wehe, “A Simple Traction Control for Tracked
the GNR method for measured data set B on Garside 14/25 sand (test rig Vehicles”, Proceedings : American Control Conference, Vol. 2, pp.
experiment) 1176–1177, 1995.
[11] Y.H. Zweiri, L.D. Seneviratne, and K. Althoefer, “Parameter Estima-
tion for Excavator Arm Using Generalized Newton Method”, IEEE
Transactions on Robotics, Vol. 20, pp. 762-767, 2004.
[12] Choopar Tan, Y.H. Zweiri, L.D. Seneviratne, and K. Althoefer, “On-
tracked vehicle as it can control the tractive forces by varying line Soil Property Estimation for Autonomous Excavator Vehicles”,
the track-terrain slip. IEEE International Conference on Robotics and Automation, pp. 121–
126, Taipei, Taiwan, 14–19 September 2003.
Traversability prediction can also be achieved from the [13] K. Iagnemma, S. Kang, H. Shibly, and S. Dubowsky, “Online Terrain
predicted tractive forces. If the maximum traction of a Parameter Estimation for Wheeled Mobile Robots With Application
tracked vehicle is reached, but it is still not enough to to Planetary Rovers”, IEEE Transactions on Robotics, Vol. 20, No. 5,
pp. 921-927, October 2004.
overcome the motion resistance exerted from the terrain, the [14] K. Iagnemma, S. Dubowsky, “High Speed Hazard Avoidance for
tracked vehicle will get stuck in this terrain, with its tracks Mobile Robots in Rough Terrain” SPIE Conference on Unmanned
spinning. Another traversability limit occurs when a tracked Ground Vehicle Technology IV, Vol. 4715, pp. 256–266, 2002.
[15] S. Hutangkabodee, Y.H. Zweiri, L.D. Seneviratne, and K. Althoefer,
vehicle gets stuck in a highly cohesive muddy terrain. In this “Soil Parameter Identification for Wheel-Terrain Interaction Dynam-
case, the track is not able to produce enough shear force to ics and Traversability Prediction,” International Journal of Automation
fail the soil and create any slip, thus no traversing is possible. and Computing (IJAC), Vol. 3, No. 3, pp. 244–251, July 2006.
[16] S. Hutangkabodee, Y.H. Zweiri, L.D. Seneviratne, and K. Althoefer,
VI. C ONCLUSION AND F UTURE W ORK “Multi-Solution Problem for Track-Terrain Interaction Dynamics and
Lumped Soil Parameter identification,” Field and Service Robotics
This paper investigates soil parameter identification for a (Results of the 5th International Conference), Springer Berlin Hei-
tracked vehicle traversing unknown terrain based on a track- delberg, Vol. 25, pp. 517–528, 2006.
[17] B.M.D. Wills, “The Measurement of Soil Shear Strength and De-
terrain interaction dynamic model and sensor feedback (slip, formation Moduli and a Comparison of the Actual and Theoretical
tractive force). The paper presents a new real-time approach Performance of a Family of Rigid Tracks,” Journal of Agricultural
for identifying all the unknown soil parameters required for Engineering Research, Vol. 8, No. 2, 1963.
[18] J.D. Faires, R. Burden, “Numerical methods”, Third Edition, Pacific
tractive force prediction based on the Generalized Newton Grove, CA. 2003.
Raphson (GNR) method. The track-terrain interaction test rig [19] Y.H. Zweiri, and L.D Seneviratne, “Diesel engine indicated and load
was constructed to validate the feasibility and effectiveness torque estimation using a non-linear observer”, Proceedings of the
IMECHE Part D: Journal of Automobile Engineering, Vol. 220, No.
of the approach. The experimental data from a real tracked 6, pp. 775-785, 2006.
vehicle [3] was also used. The approach was applied to [20] K. Ohno, T. Tsubouchi, B. Shigematsu, S. Maeyama, and S. Yuta,
these experimental data and showed good accuracy and speed “Outdoor Navigation of a Mobile Robot between Buildings based on
DGPS and Odometry Data Fusion”, IEEE International Conference
in identifying soil parameters. It was also shown to be on Robotics and Automation, pp. 1978-1984, Taipei, Taiwan, 14–19
relatively robust to initial condition, especially for parameter September, 2003.
L. The identified soil parameters are in turn used to predict [21] B. Barshan, and H.F. Durrant-Whyte, “Inertial Navigation Systems
for Mobile Robots”, IEEE Transactions on Robotics and Automation,
the tractive forces showing good agreement with all the Vol. 11, No. 3, pp. 328–342, 1995.
experimental data. Using the proposed algorithm, the tractive [22] S. Chhaniyara, P. Bunnun, Y.H. Zweiri, L.D. Seneviratne and K.
forces of tracked vehicles traversing an unknown terrain can Althoefer, “Feasibility of Velocity Estimation for All Terrain Ground
Vehicles using an Optical Flow Algorithm”, 3rd International Confer-
be predicted and can subsequently be used for traversability ence on Autonomous Robots and Agents (ICARA), Palmerston North,
prediction, traction control, and trajectory tracking. New Zealand, 12–14 December, 2006.
Future work will focus on the outdoor testing with a real
tracked vehicle or tracked mobile robot.
R EFERENCES
[1] G. Bekker, “Theory of Land Locomotion”, University of Michigan
Press, 1956.
[2] G. Bekker, “Introduction of Terrain-Vehicle Systems”, University of
Michigan Press, 1969.

3179

View publication stats