Professional Documents
Culture Documents
RP1 PDF
RP1 PDF
15 2003, Singapore 1
1 Introduction
of public transportation for energy efficiency [4]. To take matical model of the B2 which was developed. Second, we
advantage of this observation while reducing the pollution conduct an open-loop experiments of the stable unloaded
and congestion associated with automobiles is the princi- vehicle whose center-of-mass below the wheel axle. This
ple target of our work. The majority of our vehicles for the has the advantage of giving us detailed model information
pilot program [16] are of the conventional type. However, before introducing unknown moments due to passengers.
the initial cost, size, and maintenance problems are prob- Later works will build on these results by implementing a
lematic. The impetuous to design a two wheeled road vehi- preliminary LQR and conducting closed-loop experiments
cle follows from this experience. The sum of the opposing to accurately characterize machine dynamics.
wheel speed dictates longitudinal speed while differential
wheel speed gives direction. Thus, the weight, complexity,
and cost of steering mechanisms disappear. In addition, the 2 Dynamic Model of the B2
B2 can rotate on itself, thus making it highly maneuverable
in European city centers (of which several are limiting or The dynamic model of the B2 is now presented. The ba-
entirely restricting access to cars). Finally, the total size of sis of the model is a 3D vehicle carrying a rigid mass load
the B2 is such that 10 can fit into the space of one ordi- as illustrated in Figure (2). Several simplifying assump-
nary car. Thus, the two wheeled vehicle concept is highly tion are made which are important. First, a no-slip con-
compatible with our main intentions. Figure (1) shows an dition is placed on the wheels (A1). Second, the vehicle
image of our concept of a two wheeled road vehicle. body composed of chassis and persons are assumed to be
The nature of this two wheeled vehicle poses several a single rigid body (A2). Third, the vehicle is assumed to
interesting controls questions. For instance, while a per- operate on a relatively flat surface so that rolling motion is
son occupies the vehicle, their movements change the ve- not possible (A3). The first assumption provides the nec-
hicle center-of-gravity. Consequently, if the standard in- essary conditions for a kinematic model which connects
verted pendulum control were to be applied the control al- wheel motion to inertial forces which act on the inverted
gorithm would be obliged to maintain a steady acceleration body (the principle mechanism of pendulum stabilization).
to achieve the control objective [14, 2, 3] This appears to be The second assumption will be justified or modified based
the principle means of controlling the Segway. However, as on observations. The third assumption is justified restrict-
already stated, this is not compatible with our final objec- ing the vehicle to 20% grades (about 11o ). Still, it may be
tives. As such, a control algorithm must be found which worth further consideration in the future since turning the
seeks the natural equilibrium point of the system, thus re- vehicle on an incline causes the pendulum to experience a
jecting passenger motion so that speed can be automated. rolling motion.
Judging by the model, the balance problem of the B2 2.1 Lagrangian Equations and nonlinear model
appears to be a linear one since the pitch of the vehicle
is restricted to be ≤ 15o and inertial damping limits the Model development is straightforward and calculated by
pendulum body (up to 12kg − m2 ) rotations. Both facts the Lagrangian Method [5]. Working from left diagram of
which mitigate nonlinearities common to inverted pendu- Figure (2) we write the coordinates for the point interme-
lums [17, 8]. However, the mass and placement of the oc- diary between the two wheels of B2
cupants will normally be unknown so any control applied
R
should address uncertainty explicitly. ẋ = (θ̇lw + θ̇rw )sin(φ)
2
On the other hand, the occupants aren’t simply para- R
metric uncertainty, instead they represent both unmodeled ẏ = − (θ̇lw + θ̇rw )cos(φ)
2
dynamics and disturbance. In part they react to the behav- R
ior of B2; when it stops suddenly they move forward. Si- φ̇ = (θ̇lw − θ̇rw )
W
multaneously, they are true disturbance which may at any
time perform an action which impacts B2 dynamics. Con- Where the subscript notation is described in Figure (2).
sequently, it will be seen that the balance command of the Note that if we define φ(t = 0) = 0 then we have as a di-
B2 falls into the category of robust linear controls such as rect constraint on rotation φ = (R/W )(θlw − θrw ). Given
µ-synthesis [6, 19]. Ironically, to make good use of robust the center position of the vehicle in the xy-plane the wheels
control and achieve good results, one is obliged to char- positions are given as
acterize the uncertainty of the plant. Often this requires a W W
detailed experimental study to identify the plant. xlw = x − cos(φ) xrw = x + cos(φ)
2 2
Hence, this article deals with modeling and identifica- W W
ylw =y− sin(φ) yrw =y+ sin(φ)
tion of the vehicle. First, we briefly outline a 3D mathe- 2 2
x Computational Intelligence, Robotics, and Autonomous Systems, Dec. 15 2003, Singapore
The Conference on 3
y
yb yrw
ylw y
ψ g
Mb , Jbψ , Jbφ
v xlw CM
Rθ̇lw
x
xb θrw zb
xrw
W
R θlw
φ
Rθ̇rw
L
Mw , J w , R
Figure 2: Illustrations of the kinematic (left) and dynamic (right) models for the B2 operating on a plane with a effective
pendulum height of zb . The pendulum body swings about the ψ-axis and the entire vehicle can rotate in xy-plane about
the φ-axis. There are three dimensions of significance: L, R, and W which are the effective the pendulum length, wheel
radius, and wheel base. Masses are designated by M and moments as J. Subscripts are bψ, bφ, correspond to body moment
about ψ and φ axes. Subscripts lw and rw indicate either left or right wheel moment or masses. Several assumptions have
been made. First, that the vehicle operates only on a flat surface. Second, that the rotation of the wheels is proportional to
longitudinal speed (no-slip). Third, that the passengers and body of the vehicle behave as a rigid body with center-of-mass
(CM ). Occupant movement can be modeled as a torque disturbance along the ψ axis.
xb = x + Lsin(φ)sin(ψ) M = diag(Mw , Mw , Mw , Mw , Mb , Mb , Mb )
yb = y − Lcos(φ)sin(ψ) Examination of the variable L shows that there are three
zb = R + Lcos(ψ) independent variables, which we use to define the general-
ized coordinate vector q = (θlw , θrw , ψ). The equations of
Where L is the distance of the center-of-mass from the motion can then be derived from the Lagrangian equations
wheel axle as illustrated in Figure (2). Given the kinemat-
d ∂L ∂L
ics of the vehicle components, it is is a straightforward pro- − = Ξi i=1,2,3 (1)
dt ∂ q̇i ∂qi
cess to write the energy expression for the complete plant
Here Ξi are generalized forces acting parallel to each inde-
1 pendent coordinate. These are used to describe external or
L = T ∗ − V = (Ψ̇T JΨ̇ + Ẋ T MẊ) − M (g · X)
2 internal, conservative or nonconservative forces necessary
to express the model [5]. Since the variations δqi are inde-
Where Ψ and X are translational and rotational coordinates pendent, the generalized forces fi can be calculated from
defined as the equations
Ψ = (ψ, φ, θlw , θrw , η(θlw − ψ), η(θrw − ψ)T N
X ∂Rj
Ξi = fj · i=1,2,3
j=1
∂qi
and
X = (xlw , ylw , xrw , yrw , xb , yb , zb )T This calculation is necessary to generate correct forcing
functions for the motor torque as well as damping matrices
The matrices J and M are the moments and masses of the due to gearbox and tire friction. Evaluating Eqs. (1) we
vehicle. In our case, they are simply diagonal matrices derive the following dynamic equations which we place in
given as the form
J = diag(Jbψ , Jbφ , Jw , Jw , Jm , Jm ) H(q)q̈ + C(q, q̇)q̇ + G(q) = Qu (2)
The Conference on Computational Intelligence, Robotics, and Autonomous Systems, Dec. 15 2003, Singapore 4
The mass matrix is found to be body would not be useful since this value is independent of
the CM. (the terrain may not be flat). Another alternative
h0 h1 h2 −hc hc hb would be to use an accelerometer as a tilt sensor. However
H(q) = ha h1 h0 h2 + hc −hc hb the measurement is a complicated function of the body tilt,
h2 h2 h3 hb hb 0 vehicle motion, and sensor position. Instead the B2 uses
a gyroscope with gain Kgyro to measure the tilting rate of
with coefficients ha = 1/(4W 2 ), hb = LRMb cos(ψ)/2,
and hc = L2 R2 Mb cos(2ψ)/(2W 2 ) and elements defined the CM ψ̇. As a measurement sensor, gyroscope are largely
as independent of vehicle accelerations and entirely indepen-
dent of sensor position. On the other hand, A notable dis-
h0 = (4Mw W 2 + 4Jbφ + (2L2 + W 2 )Mb )R2 + advantage of the sensor is susceptibility to very low fre-
quency noise in the form of temperature drift. Even so,
4W 2 (Jm η 2 + Jw )
the frequency range of the drift is on the order of minutes,
h2 = −4W 2 η 2 Jm whereas the response of the body dynamics in the range of
h1 = (Mb (W 2 − 2L2 ) − 4Jbφ )R2 fractions of seconds. Therefore, an effective technique is to
h3 = 4W 2 (Mb L2 + Jbψ + 2η 2 Jm ) apply a very low frequency high pass filter to the gyroscope
output.
The nonlinear damping matrix is a comprised of gyro- Vehicle speed is estimated by encoders Kenc attached
scopic and damping to each motor rotor (thus further amplified by the gearbox
ratio η) and anchored to the vehicle body. Since both the
C(q, q̇) = stator and rotor of the encoder are subject to rotation, the
encoder registers the value θ − ψ and not simply the wheel
2
1 −1 − W 4LR
sec(ψ)
2 speed.
− W 4LR
sec(ψ)
ca
−1 1 Rotation about the φ can be accessed by measuring the
(θ̇rw −θ̇lw ) (θ̇lw −θ̇rw )
2ψ̇ 2ψ̇
0 difference of the wheel encoders (valid by the no-slip as-
sumption A1). However, this would be inherently noisy.
βgb,lw + βt 0 −βgb,lw
In addition, although great precision is not needed for con-
+ 0 βgb,rw + βt −βgb,rw
trol of the vehicle, it is desirable to have good odemetric
−βgb,lw −βgb,rw βgb,lw + βgb,rw
estimates. This goal can be achieved by use of gyroscopes
2 2
in conjunction with encoders as described by Borenstein
where ca = L R MW b sin(2ψ)ψ̇
2 . The terms labeled β repre- [11, 10]. Consequently, a second gyroscope is mounted
sent plant friction. Due significant causes are the gearboxes along the φ-axis. So that the measurement is consistent
and tires which are indicated by subscripts gb and t. The R
with the model, it is written as Kgyro W (θlw − θrw ).
principle cause of instability of the pendulum is because of Thus, the B2 has the following partial state measure-
the negative nonlinear stiffness nonlinear stiffness due to ment capabilities
gravity
G(q) = (0, 0, −gLMb sin(ψ))T y = (ηKenc (θ̇lw − ψ̇), ηKenc (θ̇rw − ψ̇),
Finally, motors at each wheel can apply torque which pro- R
Kgyro ψ̇, Kgyro (θlw − θrw ))T
vides a means of controlling vehicle speed and pendulum W
attitude Finally, some elaboration in the definitions of Mb , L,
Kamp,lw 0
Jbψ , and Jbφ is necessary. Each of these values depends on
ulw the number of passengers Np and their mass Mp . For our
Qu = ηKmot Kda 0 Kamp,rw
urw work, we compute the mass and pendulum length parame-
−Kamp,lw −Kamp,rl
ters by the formula
Here Kamp represents the amplifier which is modeled as
Np Mp Lp + Mb0 Lb0
a simple gain because they have bandwidth greater then Mb = Mb0 + Np Mp and L=
1kHz. While Kmot and Kad are the motor torque constant Np Mp + Mb0
and AD converter gain respectively. Here Lp and Lb0 are the distances from the passenger and
The crucial state for this control problem is CM atti- unloaded vehicle center-of-masses to the wheel axle. The
tude ψ. This is the angle subtended by the gravity vector unloaded mass of the vehicle is assigned the constant Mb0 .
and the vector connecting the wheel axis and CM. There is The moments Jbψ and Jbφ also depend on the passengers
no simple way to measure it directly. For example, mea-
suring the distance from the ground to the bottom of the Jbψ = Jb0ψ + Jp and Jbφ = Jb0φ + Jp
The Conference on Computational Intelligence, Robotics, and Autonomous Systems, Dec. 15 2003, Singapore 5
Np M p 2 Np M p 2
Jpψ ≈ Hp + Np Mp L2p and Jpφ ≈ Wp 4
12 12
2
where Hp and Wp are seated height and width of passen- Coloumb Friction
N−m
0
−6
As commented in the introduction, we wish to apply ro- Figure 3: Measured friction model for the B2 Left mo-
bust control techniques to the control of the B2. A nec- tor (similar to right motor). As can be seen, the stic-
essary step in this process is to be able to characterize the tion friction is approximately 7N − m, the Coulomb fric-
plant model. In this paper we only consider identification tion is about 4N − m, and the viscous friction is about
where Np = 0 and open-loop is stable (Mb0 rests below the 1.4N − m/rad/sec.
wheel axle). To this end, the following plan was executed
20
20
10
N−m
N−m
0
0
Opposing motor torques
Identically applied motor torques
−10 −20
−20 8 10 12 14 16 18 20 22
4 5 6 7 8 9 10 11 12 13 14 15
100
100
deg/sec
50
deg/sec
0
0
−50 −100 Possible overdamping in gearbox friction model
Gearbox locking due to stiction
−100
8 10 12 14 16 18 20 22
4 5 6 7 8 9 10 11 12 13 14 15
100
100
deg/sec
50
deg/sec
0
0
−50 −100
−100
8 10 12 14 16 18 20 22
4 5 6 7 8 9 10 11 12 13 14 15
50 10
Body−wheel system rocking
5
deg/sec
deg/sec
0 0
−5
−50 Motion caused by differences between left/right
Gearbox locking due to stiction −10 gearbox friction.
8 10 12 14 16 18 20 22
4 5 6 7 8 9 10 11 12 13 14 15
4
50
2
deg/sec
deg/sec
0 0
−2
−50
Good representation of cross−coupling
−4
8 10 12 14 16 18 20 22
4 5 6 7 8 9 10 11 12 13 14 15
time (sec)
time (sec)
Figure 4: Validation experiment planar of dynamics. From Figure 5: Validation experiment of twist (right) dynamics.
top to bottom: Applied torque at each wheel ητlw (-) and From top to bottom: Applied torque at each wheel ητlw (-)
ητrw ) (-.), left wheel encoder θ̇lw , right wheel encoder θ̇rw , and ητrw ) (-.), left wheel encoder θ̇lw , right wheel encoder
ψ̇ gyroscope output, and φ̇ gyroscope output . Experimen- θ̇rw , ψ̇ gyroscope output, and φ̇ gyroscope output . Exper-
tal values are shows as a solid line and simulation as dashed imental values are shows as a solid line and simulation as
(-.) dashed (-.)
phase of the project, we implement a simple LQR con- [10] Borenstein J. and Feng L. Gyrodometry: A new
trol. Although not ideal, this implementation allows us method for combining data from gyros and odometry
to quickly access the closed loop performance of the sys- in mobile robots. Proceedings of the 1996 IEEE In-
tem. In particular, we are able to experiment with people ternational Conference on Robotics and Automation,
mounted on the vehicle. Thus, we can learn how significant pages 423–428, 1996.
human mechanical properties (flexible modes) are to the
control problem. Furthermore, we are able to repeat these [11] Borenstein J. and Feng L. Measurement and correc-
experiments numerous times under different loads and con- tion of systematic odometry errors in mobile robots.
ditions providing a tool for building an uncertainty model IEEE Transactions on Robotics and Automation,
of the loaded B2. Therefore, from usage experience of the 12(6):869–880, 1996.
LQR command, control design specifications are revised to [12] E. Koyonagi, S. Iida, K. Kimoto, and S. Yuta. A
improve human-machine interaction. wheeled inverse pendulum type self-contained mo-
bile robot and its two-dimensional trajectory control.
Proc. of ISMCR’92, pages 891–898, 1992.
References
[13] Baldwin N. The Wolseley, ISBN 0747802971. Shire
[1] B. Armstrong. Friction: Experimental determination, Publications Ltd, Buckinghamshire, UK, 1995.
modeling, and compensation. Proceedings of IEEE
International Conference on Robotics and Automa- [14] Control of convey-crane based on passivity. J. col-
tion, 3:1422–1427, 1988. lado and r. lozano and i. fontoni. Prec. of the ACC,
Chicago, Ill. June, 2000.
[2] K. Astrom and K. Furuta. Swinging up a pendulum
by energy control. IFAC 13th World Congress, San [15] M. Parent. The cybercars. ITS World Congress
Francisco, California, 1996. Chicago Oct 2002, 2002.
[3] W. L. Brogan. Modern Control Theory. Prentice Hall, [16] B. Saugy, K. Malone, and J.W. van der Wiel. Cyber-
Englewood Cliffs, New Jersey 07632, 1991. netic transport systems: Lessons to be learned from
user needs analysis and field experience. IEEE Intel-
[4] R. Coffey and Lowson M.V. A comparative analysis ligent Vehicles Symposium, 2002.
of energy consumption and emission of urban trans-
port systems. In Recio Eds. Baldasano and Sucharov, [17] M. Spong. The swing up control problem for the ac-
editors, Urban Transport and the Environment II. robot. IEEE Control Systems, February 1995.
Computational Mechanics Publications, 1997.
[18] J. Uwland and M. Janse. Cybercars on the move.
[5] S. Crandall, D. Karnopp, E. Kurtz, and D. Pridmore- Transport Matters, TNO Traffic and Transport Jour-
Brown. Dynamics of Mechanical and Electromechan- nal, 2002.
ical Systems. Krieger Publishing Company, Malabar,
[19] K. Zhou and J. Doyle. Essentials of Robust Control.
Florida, 1968.
Prentice Hall, Englewood Cliffs, New Jersey 07632,
[6] J. Doyle, B. Francis, and A. Tannenbaum. Feed- 1997.
back Control Theory. Macmillan Publishing Com-
pany, New York, 1992.
[7] E. Garcia, P. Gonzalez, and C Canudas de Wit. Ve-
locity dependence in the cyclic friction arising with
gears. International Journal of Robotics Research,
21(9):761–771, 2002.
[8] Kajiwara H., Apkarian P., and Gahinet P. lpv tech-
niques for control of an inverted pendulum. IEEE
Control Systems, 19(1):44–54, 1999.
[9] Yun-Su Ha and Shin’ichi Yuta. Trajectory track-
ing control for navigation of the inverse pendulum
type self contained mobile robot. Robotics and Au-
tonomous Systems, 17:65–80, 1996.