See

discussions, stats, and author profiles for this publication at: http://www.researchgate.net/publication/228561343

Modelling of Mobile Robot Dynamics

ARTICLE · JANUARY 2010

CITATIONS

3 AUTHORS, INCLUDING:

Edouard Ivanjko
University of Zagreb
25 PUBLICATIONS 77 CITATIONS

SEE PROFILE

Available from: Edouard Ivanjko

Retrieved on: 23 August 2015
 MODELLING OF MOBILE ROBOT DYNAMICS
Edouard Ivanjko1 , Toni Petrinić2 , Ivan Petrović1
1
University of Zagreb, Faculty of Electrical Engineering and Computing
10000 Zagreb, Unska 3, Croatia
2
HEP d.d., 10000 Zagreb, Ulica grada Vukovara 37, Croatia
edouard.ivanjko@fer.hr(Edouard Ivanjko)

Abstract

This paper presents two approaches to modelling of mobile robot dynamics. First
approach is based on physical modelling and second approach is based on exper-
imental identification of mobile robot dynamics features. Model of mobile robot
dynamics can then be used to improve the navigational system, especially path
planing and localization modules. Localization module estimates mobile robot
pose using its kinematic odometry model for pose prediction and additional sen-
sor measurements for pose correction. Kinematic odometry models are simple,
valid if mobile robot is travelling with low velocity, low acceleration and light
load. Disadvantage is that they don’t take any dynamic constraints into account.
This leads to errors in pose prediction, especially when significant control signal
(translational and rotational velocity reference) changes occur. Problem lies in the
fact that mobile robot can’t immediately change its current velocity to the desired
value and mostly there exists a communication delay between the navigation com-
puter and mobile robot micro-controller. Errors in predicted pose cause additional
computations in path planning and localization modules. In order to reduce such
pose prediction errors and considering that mobile robots are designed to travel
at higher velocities and perform heavy duty work, mobile robot drive dynamics
can be modelled and included as part of the navigational system. Proposed two
modelling approaches are described and first results using a Pioneer 3DX mobile
robot are presented. They are also compared regarding to complexity, accuracy
and suitability of implementation as part of the mobile robot navigational system.
Keywords: Modelling, mobile robot, estimation, dynamic model.

Presenting Author’s Biography

Edouard Ivanjko received his PhD in 2009. from Faculty of Electrical
Engineering and Computing, University of Zagreb, Croatia where he is
currently working. His main research interests are: mobile robotics, lo-
calization of mobile robots and environment modelling for mobile robots.
He published 3 papers in international journals and 19 papers in proceed-
ings of international conferences. He is a member of IEEE, KoREMA and
Croatian Society for Robotics. He speaks fluent German and English.
1 Introduction occur when significant velocity changes (big accelera-
tion or deceleration values) are requested from the mo-
In recent years huge interest in autonomous industrial bile robot. Appropriate control strategies and trajec-
vehicles can be noticed. Control systems for such ve- tories need to include mobile robot dynamics proper-
hicles should take into account all disturbances which ties [4].
can appear during their missions. Their control systems
should react quickly and adapt to changing environment Recent published research also uses mobile robot dy-
conditions. Many ”classical” methods for designing namics to cope with workspace floor characteristics.
control systems (optimum-control methods, algebraic In [5] authors developed a two-level mobile robot
methods) for such vehicles require a physical system motion control strategy that can cope with different
description together with all of its parameters. For ac- workspace floor characteristics. Low level part is a clas-
curate mobile robot system description it’s necessary to sical wheel velocity controller whereas high level part
generate an appropriate dynamic model. The dynamic uses measured wheel velocities to adapt generated tra-
model allows considering such properties as: mass, in- jectory if there are significant differences from desired
ertia, friction forces, centrifugal force, torque, etc. Such values. For this purpose authors modelled the mobile
models are built in order to better understand the struc- robot as a rigid body that rolls on two drive wheels and
ture and operation of the controlled mechatronic sys- one castor including velocity changes constraints by the
tem. Creation of such a model becomes even more im- used actuators.
portant if high complex systems have to be developed.
Preparation and application of needed dynamical model Mobile robot dynamic model can be also used to im-
allows early detection of flaws and mistakes in the de- prove estimated pose in the localization module. This
scription (model) of the real system. Their modification module mostly uses a kinematic model to predict mo-
is simpler and less expensive in the virtual construction bile robot pose using appropriate control input values.
stage in comparison with physical prototypes. In this case control input consists of translational and
rotational velocities. Predicted pose is then corrected
Knowledge about mobile robot dynamics is very im- using additional sensor measurements. A typical ex-
portant for planning feasible mobile robot trajectories. ample is usage of non-linear Kalman filters for this
First a path consisting of a series of poses is generated task [6]. Quality of so estimated pose depends on used
in the path planning module. Such a path is generated kinematic model accuracy, used additional sensors and
regarding to mobile robot dimensions in the sense that workspace model (map) accuracy. Kinematic model ac-
the mobile robot can traverse free workspace underly- curacy has a significant influence in such a framework.
ing planned path. In this case dynamic properties are
not important. Desired path following is only done in Control input for kinematic model is usually computed
geometric space and path planning criterion is that the in a navigational computer connected via a commu-
mobile robot doesn’t collide with any stationary obsta- nication link with a low level micro-controller. The
cles. Typical path planning algorithms for such appli- micro-controller handles drive wheels velocity control
cations are the A* and D* algorithms [1]. and their current velocity measurement. Velocity mea-
surement is then returned to the navigational computer
Things change when moving obstacles are part of mo- and used for pose estimation or path planning compu-
bile robot workspace. They can be people or other mo- tations. Control input values of the navigational com-
bile robots or other moving machines. In these situa- puter can differ from values that the micro-controller
tions a certain pose must be reached by the mobile robot currently uses for drive wheels velocity control. Rea-
in a certain time frame to avoid collision. In such cases sons are communication link delay, mobile robot dy-
mobile robot motion velocity has to be constantly al- namic properties and used drive mechanical character-
tered in an appropriate way that mobile robot can avoid istics. Mechanical characteristics include influence of
collision with a moving obstacle. Only in that way a friction, backlash, etc. Result is that used kinematic
collision free motion in a workspace with moving ob- model predicts mobile robot movement, while micro-
stacles can be guaranteed [2]. During trajectory plan- controller hasn’t received movement command yet or
ning phase mobile robot dynamic properties are used to predicts a mobile robot movement that can’t be per-
create a set of velocity profiles that can be performed by formed by the used mobile robot. So, there exists a con-
the mobile robot in a safe manner. In this process mo- stant prediction error that can’t be taken into account by
bile robot velocity limitations must be respected with means of calibration or path planning restrictions.
preserved trajectory curvature [3].
It would be beneficial to model such features and in-
Mobile robot dynamical model is very important in clude them in the pose prediction step as part of the
cases when mobile robot velocities generate forces localization module. Also path planning could be im-
which influence can’t be neglected during movement. proved, especially the moving obstacle avoidance part
A good example is mobile robot soccer where used mo- when generated path is altered in time space i.e. a
bile robots are small and velocities are significant com- trajectory is created. To do this mobile robot physi-
pared to their size and mass. Typical shape is a cube cal properties have to be examined and their influence
of about 7.5 (cm) size and velocities can be faster than on interesting variables has to be determined. In case
2 (m/s). Small size and mass combined with such ve- of mobile robot navigation, crucial variable is its pose
locity values can cause slippage of drive wheels or turn which is predicted i.e. estimated using drive wheel ve-
overs in a curvature trajectory. Slippage can especially locities. Input variables are velocity references and ap-
propriate model should use them too as an input to pro- left actuated
vide current mobile robot velocity value on its output. Y wheel
r V
These velocity values can then be used by the localiza- L

tion and other modules. This would be the first step or center of mass
V
C
model creation. Second step would include validation
of obtained model. It isn’t good to validate the model C A Θ
y
d
on real mobile robot in its working environment due to
center of axle
danger of damage. More preferable is simulation test- b
ing where velocity data from mobile robot experiments
castor wheel
are used. Such a way is used in this article also. V
R
right actuated
This paper presents two approaches to modelling dy- wheel
namic mobile robot features including influence of me- x
chanical drive characteristics. First approach is based X
on making a physical model of mobile robot body and
components used for velocity control like velocity con- Fig. 1 Geometrical dependencies of a differential drive
troller, motor, gearbox, etc. Second approach is based mobile robot.
on experimental fitting of recorded mobile robot veloc-
ity data regarding reference velocity data. Both models
are validated using velocity data recorded using a Pio- Dynamic motion equation can be derived using Euler-
neer 3DX mobile robot. Lagrange formulation [8, 9]:
The rest of the paper is organised as follows. Sec- ( )
ond section contains description of both modelling ap- d ∂L ∂L
− = Qi , (1)
proaches. Third section gives an overview on models dt ∂ q̇i ∂qi
implementations in M ATLAB/S IMULINK. After that
obtained results are given followed with models com- where L stands for difference of kinetic, T , and poten-
parison. Paper ends with conclusion and description of tial, U , energy, qi stands for generalized coordinate, and
future work on this topic. Qi stands for generalized force that acts on the mechan-
ical system.
2 Modelling approaches Under assumption that mobile robot moves only on a
plane surface, potential energy of robot is zero (U = 0)
As mentioned above, this article presents two ap-
and we have to find only kinetic energy of the mobile
proaches to mobile robot dynamics modelling. First
robot. Kinetic energy of the whole structure is given by
steps needed for physical modelling are described and
the following equation:
secondly steps taken for experimental identifications
are described. Both models need to take into account
maximal values of rotational and translational veloci- T = Tt + Tr + Trwr , (2)
ties including maximal rotational and translational ac-
celeration and deceleration values. These model fea- where Tt is kinetic energy of mobile robot translation,
tures can be easily taken into account by using satura- Tr is kinetic energy of mobile robot rotation, and Trwr
tion and slope limitation functions. is kinetic energy of rotation of wheels and rotors of DC
motors, all in (kgmm2 /s2 ) respectively. Values of in-
2.1 Physical modelling
troduced energy terms can be expressed by following
Used Pioneer 3DX robot is a two wheeled differen- equations:
tial drive robot, where each wheel is driven indepen-
dently. Forward motion is produced by both wheels 1 1
driven at the same rate, turning right is achieved by Tt = mv 2 = m(x2c + yc2 ), (3)
2 c 2
driving the left wheel at a higher rate than the right
wheel and vice versa for turning left. This type of mo- 1
bile robot can turn on the spot by driving one wheel Tr = IA Θ̇2 , (4)
forward and second wheel in the opposite direction 2
at same rate. Third wheel is a castor wheel needed
for mobile robot stability. Drive wheels are equipped 1 1
Trwr = I0 Θ̇2R + I0 Θ̇2L , (5)
with encoders and their angular velocity readings be- 2 2
come available through simple routine calls. Kinematic
model of a differential drive mobile robot can be found where m is the mass of entire mobile robot (kg), vc
in [7] and geometrical dependencies are given in Fig. 1, is linear velocity of the mobile robot’s center of mass
where r is drive wheel radius (mm), vL and vR are left C (mm/s), IA is the moment of inertia of entire mo-
and right drive wheel velocities, respectively (mm/s), bile robot with respect to the point A (kgmm2 ), Θ is
x and y present mobile robot position in cartesian co- mobile robot orientation (rad), Θ̇ is mobile robot ro-
ordinates in (mm), and b is axle length between drive tational speed (rad/s), I0 is the moment of inertia of
wheels (mm). combined drive motor (rotor) and wheel (kgmm2 ), and
Θ̇R , and Θ̇L are angular velocities of the right and left
drive wheel, respectively (rad/s). B Θ̈R + AΘ̈L = ML − K Θ̇L , (15)
Components of the velocity of point A, can be ex-
( ( ) )
pressed in terms of Θ̇R and Θ̇L : mr2 IA + md2 r2
A= + + I0 , (16)
r 4 b2
ẋA = (Θ̇R + Θ̇L ) cos(Θ), (6)
2 ( ( ) )
mr2 IA + md2 r2
r B= − . (17)
ẏA = (Θ̇R + Θ̇L ) sin(Θ), (7) 4 b2
2
2.2 Experimental identification
r(Θ̇R − Θ̇L ) Number of needed values is much smaller in this case
Θ̇ = , (8)
b and it consist of maximal rotational and translation ve-
locity including their maximal acceleration and decel-
where ẋA presents velocity of point A in direction of eration values.
the X-axis (mm/s), and ẏA presents velocity of point
A in direction of the Y -axis (mm/s). Other features can be obtained by creating appropri-
ate experiments. To obtain this features, first critical
Components of the velocity of point C considering ve- velocity change cases have to be defined. Such cases
locity of point A are now: are partly covered with the above mentioned maximal
values. Other cases are, when mobile robot changes
ẋC = ẋA + dΘ̇ sin Θ, (9) its translational or rotational velocity direction, when
it starts its movement (step wise velocity reference
change), and when velocity is constant (steady move-
ment) or constantly changing (mobile robot is acceler-
ẏC = ẏA − dΘ̇ cos Θ, (10) ating or decelerating). Figs. 2 to 5 show mobile robot
reference and measured velocity relationship for men-
where d is the distance between points A and C in tioned cases. Only translational velocities are presented
(mm), and ẋC presents velocity of point C in direc- whence rotational velocity responses show similar be-
tion of the X-axis (mm/s), and ẏC presents velocity haviors. Velocity responses where obtained using a Pi-
of point C in direction of the Y -axis (mm/s). oneer 3DX mobile robot controlled with an application
Total kinetic energy of the mobile robot can be calcu- using 100 (ms) sampling time. Only for stand still area
observation smaller sampling time was used (20 (ms)).
lated in terms of Θ̇R and Θ̇L :
150
( ) reference value
mr 2 (IA +md2 )r 2 I0
T (Θ̇R ,Θ̇L ) = 8 + 2b2
+ 2 Θ̇2R measured value

( )
translational velocity [mm/s]

mr 2 (IA +md2 )r 2 I0
+ 8 + 2b2
+ 2 Θ̇2L 100
( )
mr 2 (I +md2 )r 2
+ 4 − A b2 Θ̇R Θ̇L .

(11)
50

Now, Lagrange equations:

( ) 0
d ∂L ∂L
− = MR − K Θ̇R , (12)
dt ∂ Θ̇R ∂ΘR 0 0.5 1
time [s]
1.5 2

( ) Fig. 2 Mobile robot velocity response in case of a step-

d ∂L ∂L
− = ML − K Θ̇L , (13) like velocity reference change.
dt ∂ Θ̇L ∂ΘL

are applied. MR and ML are right and left actua- It can be seen that in case of a step-like velocity refer-
tion torques, respectively in (kgmm/s2 ) and K Θ̇R and ence change (Fig. 2), mobile robot velocity response is
K Θ̇L are viscous friction values of right and left wheel- similar to a ramp determined by maximal acceleration
motor systems, respectively in (kgmm/s2 ). Finally, value. Such behavior is expected but there is also an
dynamic motion equations can be expressed as: additional time delay in the velocity response. It can
be explained as a communication delay. Fig. 3 displays
case of a constant reference. Mobile robot can hold de-
AΘ̈R + B Θ̈L = MR − K Θ̇R , (14) sired velocity with influence of noise. On the right side
 cluding a greater error when the ramp begins. One part
of the greater error happens due to used drive controller
−140
reference value
influence and friction. Fig. 5 displays the case of a ve-
locity direction change. An area where mobile robot
translational velocity [mm/s]

measured value
−145 stands still can be observed. Deceleration to stand still
happens constantly like the reference but acceleration in
the opposite direction continues after velocity reference
−150 reaches a certain value. This feature can be explained
with influence of movement in previous direction and
−155
friction.
In order to reduce error between control input values
used for velocity values prediction and true mobile
−160
3 4 5 6 7 8 9 10 11 robot movement values, appropriate dynamic model
time [s] can be used. It has to take into account all mentioned
Fig. 3 Mobile robot velocity response in case of a con- cases. Additionally, model has to be simple so that it
stant velocity reference. doesn’t make any additional burden on the navigation
computer. Also a small number of model parameters
is preferable to enable a possibility of their on-line es-
timation. Communication delay can be modelled us-
35 ing transport delay, and behavior around standstill us-
reference value ing a variable threshold. After reference value reaches
30 measured value
threshold value, estimated velocity values start to rise.
25 Crucial model values that have to be obtained experi-
translational velocity [mm/s]

mentally are communication delay and threshold at zero

20
velocity value. Communication delay value is deter-
mined by comparing sent velocity reference value and
15
measured velocity value in time space. Accordingly,
10
each reference and velocity value had a corresponding
time stamp. Obtained time delay value is 250 (ms) for
5 translational part and 270 (ms) for rotational part. It
has to be mentioned here that mobile robot controller
0 works internally with a sampling time of 5 (ms), and
13 13.1 13.2 13.3 13.4 13.5 13.6 sends out averaged velocity values every 100 (ms).
time [s]
This explains why obtained communication delay value
Fig. 4 Mobile robot velocity response in case of a ramp- isn’t an integer multiple of the sampling time. For this
like velocity reference change. reasons communication delay is modelled as a combi-
nation of integer delay and ZOH discretized first order
transfer function. That means that delay of 270 (ms)
30
in case of the translational part is modelled as delay of
two discretization steps and the rest of 70 (ms) as a
20
reference value first order discrete transfer function with time constant
measured value
of 100 (ms) and discretization time of 100 (ms).
translational velocity [mm/s]

10 Needed threshold value depends on the ramp slope, i.e.

velocity acceleration or deceleration and its value is
0 then computed from a dependency recorded from a se-
ries of experiments. Experiments where done for char-
−10 acteristic values of velocity change and a good enough
approximation can be made using a straight line. Fig. 6
−20 displays obtained threshold values for translational part
of the model. Rotational part is obtained in an analog
−30 way.
25.8 26 26.2 26.4 26.6 26.8 27
time [s]
3 Model implementation
Fig. 5 Mobile robot velocity response in case of a ramp-
like velocity reference change crossing stand still point. This section describes how obtained models were im-
plemented in M ATLAB/S IMULINK. Both models are
implemented with equal requirements. First require-
ment is of course best possible accuracy. Second re-
of the figure, influence of the mentioned time delay can quirement was that model can accept measurements ob-
be observed. Fig. 4 displays case of a ramp-like velocity tained from a real mobile robot and compare velocity
reference change. Also a time delay can be observed in- values for model accuracy validation. For the sake of
 16
ometric center of that element. This is a good approxi-
mation because all elements have a symmetrical shape
threshold value arround stand still point [mm/s2]
14
and constitution like batteries, wheels, motor with gear-
12 box, case, etc. PI controller for drive wheel angular ve-
locity is used. Integral time constant compensates dom-
10
inant time constant of the velocity control loop. Pro-
8 portional gain is chosen in order to damping factor of
measured threshold value regulation loop be satisfied (ξ = 0.9). PI parameters
6 estimated threshold value were so KR = 11996 and TI = 4.58 (s).
4 Obtained model can be seen in Fig. 7. Model input
variables are mobile robot rotational and translational
2
velocity references. Their values and known mobile
0
robot kinematic model with velocity and acceleration
0 50 100 150 200 250
2
300 constrains are then used to compute left and right drive
mobile robot acceleration/deceleration value [mm/s ]
wheel speed references including time delay. These ref-
Fig. 6 Threshold value for translational part. erences are then used as input for left and right drive
wheel speed controller. Coupling between left and right
side is also modelled. End part of the model on the right
side computes final mobile robot rotational and transla-
comparison, deviation of the estimated velocity from tional velocities. Number of encoder ticks per revolu-
the measured one is computed including average and tion equals 500.
maximal error values.
translation
3.1 Physical model velocity
esimated
reference
translation
[mm/s]
Pioneer 3DX drive system uses high-speed, high- velocity
[mm/s]
torque, reversible-DC motors, each equipped with a
high-resolution optical quadrature shaft encoder for velocity and PI DC
acceleration left wheel angular controller motor
constrains Mobile
precise position and speed sensing and advanced dead- velocity reference Mobile
+ robot
robot
reckoning. Motor parameters can be found in [10] and kinematic dynamics
kinematic
model right wheel angular +
model
most important ones are given in Tab. 2. Mobile robot + PI DC encoders
velocity reference
time delay controller motor
parameters, motor gear head ratios and encoder ticks
per revolution can be found in [11] and most important estimated
rotation rotation
ones are given in Tab. 1. It has to be noticed that val- velocity velocity
reference [o/s]
ues I0 and IA were estimated and the rest is taken from [o/s]
mobile robot manufacturer data.
Fig. 7 Block scheme of proposed physical model.
Tab. 1 Mobile robot parameters
Parameter Value Unit Time delay is in this case difficult to model because
m 28.05 (kg) amount of time needed for velocity data transmission
r 95 (mm) and their evaluation in navigational computer or mo-
b 320 (mm) bile robot micro-controller isn’t documented. It can be
d 57.8 (mm) heuristically determined and verified in simulation. So,
I0 9.24 · 10−2 (kgmm2 ) time delay of 3 time steps (300 (ms)) was used.
IA 175 · 103 (kgmm2 ) 3.2 Experimental model
Fig. 8 presents proposed experimental model block
Tab. 2 Mobile robot drive parameters scheme consisting of parts described in subsection 2.2.
It can be used for translational and rotational velocity
Parameter Value Unit part. Model input is generated velocity reference and
KA 0.013 (A/V
√) output is estimated mobile robot velocity. To obtain es-
KM 0.029 (N m/ W ) timated values of both, rotational and translational ve-
TA 1.1 (ms) locities, two models have to work in parallel.
TM 8.5 (ms) According to model blocks in Fig. 8, proposed model
K 35 · 10−7 (N ms/rad) can be implemented mostly using standard S IMULINK
blocks. Only problematic part is the block labelled
Coordinates of mobile robot center of mass and mo- ”Threshold at zero velocity”. It has to influence the es-
ment of inertia with respect to the mobile robot cen- timated velocity value only when mobile robot velocity
ter of mass were computed by separating used mobile is changing its direction. And then, only in the case
robot on distinctive elements whose mass and pose in- when absolute velocity value is beginning to rise. This
side mobile robot could be easily measured. It is as- part is solved by detecting appropriate velocity change
sumed that mass of each element is concentrated in ge- situation and then applying one of the following cases:
 Velocity reference Tab. 5 Error values in case of rotational velocity estima-
tion in (◦ /s) for second test values set

Communication delay Error Physical model Experimental model

Maximal 6.25 8.94
Average 1.19 0.92
Velocity and
acceleration constrains
Tab. 6 Error values in case of translational velocity es-
timation in (mm/s) for second test values set
Threshold at
zero velocity Error Physical model Experimental model
Maximal 53.96 32
Average 6.77 4.66
Estimated velocity

Tab. 7 Error values in case of no model for second test

Fig. 8 Block scheme of proposed experimental model. values set in (◦ /s) and (mm/s)
Error Rotation Translation
(i) estimated velocity value isn’t changed, (ii) estimated Maximal 12.87 48.83
velocity value is changed to zero. In that way separately Average 2.13 14.76
deceleration to mobile robot stand still for the case of
negative and positive velocity are detected. Estimated
velocity changes are detected by comparing previous Obtained values of maximal velocity error and its av-
and current values. A wrap to zero block is used to erage value are given in Tabs. 3 to 6. For comparison
make needed changes to estimated velocity value when Tab. 7 contains error values for the case with no model.
estimated value is below the set threshold value. Significant improvement can be observed. Apart er-
ror values, obtained velocity responses were also ex-
amined. They are given in Figs. 9 and 11. Only a part
4 Obtained results of recorded velocity values are presented for the sake of
In order to test proposed dynamic models two data sets a better representation.
obtained from a real Pioneer 3DX mobile robot were Figs. 10 and 12 present error between measured veloc-
prepared. First data set consists of clearly separated ity and estimated velocity. Fig. 10 presents translational
critical velocity change cases. Second data set is taken part for first data set and Fig. 12 rotational part for sec-
from a navigation algorithm testing experiment. Navi- ond data set. Other cases can be presented in a similar
gation algorithm generated velocity references consid- fashion.
ering current obstacle situation in mobile robot environ-
ment and respecting mobile robot motion constraints. It
presents a more realistic situation and is therefore better
150
situated for described models accuracy comparison.
100
translational speed [mm/s]

Tab. 3 Error values in case of rotational velocity estima-

tion in (◦ /s) for first test values set
50

0
Error Physical model Experimental model
Maximal 4.31 4.59 −50

Average 0.81 0.48 reference

−100 measurement
physical model
experimental model
−150

Tab. 4 Error values in case of translational velocity es- 26.8 27 27.2 27.4
time [s]
27.6 27.8 28

timation in (mm/s) for first test values set

Error Physical model Experimental model Fig. 9 Translational velocity response detail in case of
Maximal 18.8 12.4 the first data set.
Average 3.36 1.62

5 Model comparison
For the sake of a better model comparison using devia-
tion values between this two data sets, average velocity As it was expected, second data set results with a less
values have to be known. First data set had maximal accurate estimation. One reason for the less accurate
translational velocity of 300 (mm/s) and maximal ro- estimation is more frequent change of the velocity ref-
tational velocity of 50 (◦ /s). Second data set had max- erence. Second reason are situations that can’t be de-
imal values of 600 (mm/s) and 63 (◦ /s), respectively. tected by an off-line model like drive wheel slippage or
 8
value can be used as a good base for real mobile robot
6 travelling velocity estimation in combination with other
sensors [12].
translational speed difference [ mm/s]

2 From data given in tables above, it can be observed that

0 experimental model has smaller average error values
−2
which means it mostly better models examined mobile
robot drive dynamics. Maximal error value is larger in
−4
the case of the experimental model. Such spikes are re-
−6 lated to cases of significant velocity reference changes.
−8
In such cases internal states define partly velocity re-
sponse and they are not included in the experimental
−10
model. That presents a drawback of this model. Physi-
−12
0 5 10 15 20 25 30 35
cal model contains internal states of mobile robot drive
time [s] system because it models the controller, DC motor,
gearbox with encoders, and communication delay with
Fig. 10 Estimated translational velocity error for first appropriate transfer functions. So it performs better in
test values set and physical model. cases where this internal states are significantly influ-
encing velocity response. Currently this model doesn’t
−6 include friction and backlash influence. Adding these
features could improve this model accuracy but it would
−8 also complicate its implementation as part of mobile
robot navigational system. So its more beneficial to use
−10
a simpler model with equal accuracy and, if possible,
rotational speed [ °/s]

combine it with an on-line parameter estimation frame-

−12
work.
Another crucial aspect of the proposed dynamic mod-
−14 els comparison is possibility of their implementation as
part of mobile robot navigational system. Also possi-
−16 reference bility of making an on-line parameter estimation frame-
measurement
physical model work would be a good feature. In both cases only com-
−18 experimental model munication delay can be estimated on-line. One has
81 81.5 82 82.5 83 83.5 84 84.5 to consider that only available measurement is mobile
time [s]
robot velocity. Regarding possibility of implementation
Fig. 11 Rotational velocity response detail in case of the experimental model is simpler and doesn’t suffer from
second data set. the accuracy loss when discretized.

5
6 Conclusion & future work
4 This paper presents two modelling approaches regard-
3
ing mobile robot drive dynamics. First approach con-
siders modelling every element of mobile robot drive
rotational speed difference [ °/s]

2
system and corresponding control framework. Second
1 approach models characteristic velocity change cases.
0 It results in a simpler model that gives smaller average
velocity error.
−1

−2 Model development consists of two phases. First phase

consists of mobile robot examination to create an ap-
−3
propriate model concept i.e. modelling phase. This
−4 phase includes also creation of appropriate experiments
−5 so that all dynamic properties can be included in the
0 20 40 60 80 100
time [s] model concept. Second phase includes firstly testing
of the model concept i.e. simulation phase. Testing is
Fig. 12 Estimated rotational velocity error for second done in M ATLAB/S IMULINK simulation using experi-
test values set and experimental model. mental data from the first phase. Additionally data from
a real mobile robot are used for simulation testing. Such
an approach ensures an enough accurate model to be
finally implemented as part of a mobile robot naviga-
travelling over uneven floor. Mobile robot drive wheels tional system.
are in such situations under random environment in-
fluence and their drive wheel rotation velocity doesn’t As mentioned both models are tested in M AT-
reflect real mobile robot travelling velocity. But, their LAB /S IMULINK environment using velocity data ob-
tained from real mobile robot in real navigational con- [4] Gregor Klančar, Borut Zupančič, and Rihard
ditions. First test results confirm an improvement in Karba. Modelling and simulation of a group of
comparison to usage of velocity reference values for mobile robots. Simulation Modelling Practice and
mobile robot motion prediction. When used by navi- Theory, 15:647–658, 2007.
gational system, mobile robot pose prediction could be [5] A. Albagul and Wahyudi. Dynamic modelling and
more accurate and a more precise generated trajectory adaptive traction control for mobile robots. In-
following can be assured. ternational Journal of Advanced Robotic Systems,
1(3):149–154, 2004.
So future work will go into direction of including this
dynamic model into mobile robot navigational system [6] Edouard Ivanjko, Ivan Petrović, and Mario Vašak.
(localization and path planing module) and expanding Sonar-based pose tracking of indoor mobile
it with an on-line estimation framework. robots. AUTOMATIKA - Journal for control, mea-
surement, electronics, computing and communi-
cations, 45(3-4):145–154, 2004.
7 Acknowledgement [7] Edouard Ivanjko, Ivan Komšić, and Ivan Petrović.
This research has been supported by the Ministry of Simple off-line odometry calibration of differ-
Science and Technology of the Republic of Croatia un- ential drive mobile robots. In Proceedings of
der grant No. 036-0363078-3018. Authors wish also 16th International Workshop on Robotics in Alpe-
to thank Marija Dakulović for help in providing needed Adria-Danube Region, pages 164–169, Ljubljana,
velocity data from her navigation experiments for dy- Slovenia, june 7-9, 2007.
namic model testing. [8] S. M. LaValle. Planning Algorithms. Cambrige
University Press, New York, 2006.
[9] Jasmin Velagić, Bakir Lacević, and Nedim Osmić.
8 References Motion Planning, chapter Nonlinear Motion Con-
[1] Anthony Stentz. Optimal and efficient path plan- trol of Mobile Robot Dynamic Model, pages 529–
ning for unknown and dynamic environments. In- 550. InTech Education and Publishing, june 2008.
ternational Journal of Robotics and Automation, [10] Mobile robot drive motors parameters.
10:89–100, 1993. www.allmotion.se/files/pdf/1707.pdf.
[2] M. Djakulović and I. Petrović. Dynamic window [11] Mobile robot construction parameters.
based approach to mobile robot motion control in http://www.activrobots.com/robots/p2dx.html.
the presence of moving obstacles. In Proceedings [12] P. Goel, S. I. Roumeliotis, and G.S. Sukhatme.
of IEEE International Conference on Robotics Robust localization using relative and absolute po-
and Automation - ICRA 2007, Roma, Italy, 10-14 sition estimates. In Proceedings of the IEEE/RSJ
April 2007, pages 1986–1991, 2007. International Conference on Intelligent Robots
[3] G. Oriolo, A. D. Luca, and M. Vendittelli. Wmr and Systems (IROS), 1999.
control via dynamic feedback linearization: De-
sign, implementacion, and experimental valida-
tion. IEEE Transactions on Control Systems Tech-
nology, 10(6):835–852, 2002.