You are on page 1of 16

L. G.

Bushnell
D. M. Tilbury
Steering Three-Input
S. S. Sastry Nonholonomic Systems:
Electronics Research Laboratory
Department of Electrical Engineering and Computer Sciences
The Fire Truck
University of California at Berkeley
Berkeley, California 94720 Example

Abstract constant controls (Monaco and Normand-Cyrot 1992).


In this article we will look at using sinusoidal and poly-
In this article we steer wheeled nonholonomic systems that can
be represented in a so-called chained form. Sufficient condi- nomial control inputs for steering the fire truck system, a
tions for converting a multiple-input system with nonholonomic nonholonomic system with three inputs.
velocity constraints into a multiple-chain, single-generator Our investigation of the fire truck system is fundamen-
chained form via state feedback and a coordinate transfor- tal to understanding multiple-steering, multiple-trailer
mation are presented along with sinusoidal and polynomial nonholonomic systems, which may be used in practice in
control algorithms to steer such systems. Our example is the manufacturing plants, nuclear power plants, or any area
three-input nonholonomic system of a fire truck, or tiller truck. unsafe for human operators. The contributions of this ar-
In this three-axle system, the control inputs are the steering ticle are the introduction of a three-input system as an
velocities of both the first and third (or tiller) axles and the
example of a nonholonomic system that can be steered
forward driving velocity of the truck. Simulation results are in chained form, steering algorithms for multiple-input
given for parallel parking, left hand turning, right hand turn- chained form systems, sufficient conditions for converting
ing, and changing lanes. Comparison is made to the same
a nonholonomic system with three or more inputs into a
vehicle without tiller steering.
chained form system, and simulations that show how ex-
tra steering wheels can result in greater maneuverability.
We are interested in steering mechanical systems with
1. Introduction
nonholonomic, or nonintegrable, constraints on their ve-
Chained form systems were first introduced in Murray locities. These constraints can be written as
and Sastry (1993) as a class of systems inspired from
w2(x)X = 0 i =
1, 2, ... , k,
Brockett (1981) to which one could convert a num-
ber of interesting examples, including a car and a car where x E R n is the state of the system and wi(x) E
with one trailer, and for which it was easy to derive Rn, i =
1, ... , 7~ are row vectors. Such constraints arise,
steering control laws. These examples had two inputs, for example, when a wheel rolls without slipping. If these
and their chained forms had one chain. In this article constraints are integrable, they are called holonomic con-
we convert systems with more than two inputs into a straints and yield level surfaces h2(x} ci for some
=

chained form with more than one chain. The method we constant ci to which the trajectory of the system is re-
present for transforming a drift-free, nonholonomic sys- stricted. If these constraints are not integrable, they are
tem into chained form is analogous to the method for called nonholonomic constraints.
exact linearization of a nonlinear system presented in We assume that the wi, i = l, ... , k are linearly inde-
Isidori (1989). pendent and smooth. The corresponding codistribution
A method for steering chained form systems with si- SZ(x) =
span{(.c) 1 (x), ...,LU~(x)} has dimension k.
nusoids was first stated in Murray and Sastry (1990). Therefore, we can find an (n - k)-dimensional dis-
Other practical and efficient methods for steering include tribution A(x) = span{g} (x),..., gn-~(x)}, with
polynomial controls (Tilbury et al. 1993) and piecewise gi(x) E R&dquo;~, i = 1, ... , n - k such that A = 5~1;
i.e., w2(x) ~ gj (x) =
0, for all w2 E SZ,g~ E A. Then
a mechanical system with the above nonholonomic lin-
This work was supported in part by NSF-IRI 9014490. DMT would like to
acknowledge an AT&T Ph.D. Fellowship for partial support. ear velocity constraints can be represented as a control
The International Journal of Robotics Research, system with inputs uz as follows:
- - . - . - - .. , - - - - - - - - .

366
Our motion planning problem therefore consists of con- and a polynomial algorithm to control these systems. In
trolling the drift-free system Section 5, simulation results are given for the fire truck
example. In Section 6, conclusions are drawn.

where x E open set U C Rn, ~ci(t) E R for i 1, ... , m,


=
2. The Fire Truck
m < n, and the gi are smooth, linearly independent
vector fields. All assumed to Fire trucks areused to carry aerial ladders, tools, and
subsequent conditions are

hold on the open set U. Given xo and xf , we wish to find equipment, and their main purpose is rescue and ventila-
a control law u (ui (t), ... , um(t)) to steer x(O) xO to
= = tion. They are mainly used by fire departments in large
x(T) xf on the time interval [0, T].
= cities in the United States and have great maneuverability
Recent work (Murray and Sastry 1993) in the area of through narrow city streets as a result of the extra steer-
controlling nonholonomic systems by using sinusoidal ing on the third axle, or tiller.
inputs has concentrated on systems with two inputs. If the The fire truck is an example of a three-input nonholo-
system meets certain sufficient conditions allowing it to nomic system. It is mathematically modeled as two planar
be transformed into what we call a single-chain, single- rigid bodies supported by three axles. The support of the
rear body, or trailer, is over the center of the rear axle of
generator chained form,
the front body, or cab (axle-to-axle hitching). The first
and third axles are allowed to pivot, while the middle
axle is rigidly fixed to the cab body. The wheels are as-
sumed to roll but not slip, thus giving linear velocity
constraints.
The derivation of the kinematic equations for the fire
truck is shown in Figure 1, where we emphasize the
by a change of coordinates and state feedback, then the truck’s two rigid bodies. The states of the mathemati-
system may be steered by setting the inputs vl and v2 cal model, all functions of time, are chosen as follows:
to be sinusoids at integrally related frequencies. We call (~1, yl) is the Cartesian location of the center of the
this a chained form system because the derivative of each rear axle of the cab, 01 is the steering angle of the front
state depends on the state directly above it in a chained wheels with respect to the cab body, and 81 is the orienta-
fashion. This particular chained form is reminiscent of tion of the cab body with respect to the horizontal axis of
Brunovsky normal form. Indeed, with the input v, set the inertial frame. The states (X2, Y2, 42, 02) are described
to 1, the coordinates z2, ... , zn are in Brunovsky canon-
similarly for the trailer, except that ~2 is the angle of the
ical form. We note, however, that chained form systems rear wheels with respect to the trailer body.
are nonlinear, drift-free, and bilinear in the input and state Let the distance between the front and rear axles of
variables. The input vi that appears in the chain is called the cab be Lo and the distance between the centers of
the generating input. the rear axles of the cab and trailer be L 1. This gives the
In this article we extend the above idea to systems holonomic constraints
with three or more inputs. The plan is to transform a
three-input system with nonholonomic velocity constraints
into a two-chain, single-generator chained form by state
feedback and a coordinate transformation. In this special
form, it can be controlled using sinusoidal or polynomial The six coordinates x : _ (~ i , y ~ , ~ ~ , 91, ~~, B2 ) are suf-
input functions. This idea is also extended to the general ficient to represent the positions of the cab, trailer, and
case of m + 1 inputs in the Appendix. wheels.
The outline of the article is as follows. In Section 2, For mechanical systems with wheels rolling and turning
our three-input example of a nonholonomic system is on a surface, the nonslipping constraint states that the
introduced. The kinematic equations are derived and
velocity of a body in the direction perpendicular to each
represented as a control system. This example, the fire wheel must be zero. In terms of coordinates, for a wheel
truck, will be used throughout the article to illustrate centered at location (~, y) and at an angle cp with respect
the analytical results. In Section 3, we give sufficient to the horizontal axis of the fixed frame, the constraint is
conditions for transforming a three-input nonholonomic
system into a two-chain, single-generator chained form
and show how the example can be transformed into this
special form. In Section 4, we present a step-by-step To simplify our kinematic model of the fire truck, each
sinusoidal algorithm, an all-at-once sinusoidal algorithm, pair of wheels is modeled as a single wheel centered at

367
Fig. 1. The configuration of the fire truck.

the midpoint of the axle.’ Requiring that the wheels do for all wi E 0, gj E A. A simple calculation shows that
not slip gives the three linear velocity constraints the following vector fields form a basis for A:

These constraints can be expressed more compactly as


w’(x) - ii 0, where the w’(x) are row vector fields in
=
R6:

The nonholonomic constraints w’(x) - X =


0, for all
c,~i C 11 are equivalent to i E A (i.e., x is a linear com-
The corresponding codistribution is H(x) = bination of vector fields in A). Therefore, the kinematic
span { &dquo;&dquo;,1 (x), &dquo;&dquo;,2(X), &dquo;&dquo;,3 (x)}. Because fi has dimen- model of the fire truck as a control system with three
sion three and the state space is of dimension six, we
inputs is written as
can find a three-dimensional distribution A(x) :=

Span f 91 (x), g2(X), 93 (X) such that c~i(x} ~ gj (x) 0, =

1. It may be shown (Alexander and Maddocks 1988) that in fact the two We have chosen the basis 19 1, g2, 93 1 for A so that the
wheels have different angles and their normals all intersect at a single
input u j corresponds to the forward driving velocity of
&phis;
1 &phis;
1
2
point. If, for example, we let and be the angles of the front wheels the truck, u2 corresponds to the steering velocity of the
of the cab, we can derive a holonomic constraint that can be used to
eliminate one of these two variables. Thus, we need keep track of only one front wheels of the cab, and U3 corresponds to the steer-
of the wheel angles. ing velocity of the rear wheels of the trailer.

368
3. Converting to Two-Chain,
Single-Generator Chained Form
The foregoing kinematic equation for the fire truck be-
comes complicated when we investigate how to find
control inputs ~~c 1, u2, u3 ~ that will steer the state x E R6
from an initial point to a desired point. Chained form sys- such that the distributions
tems are constructed in such a way that they are easily
steered with sinusoidal, polynomial, or piecewise constant
inputs.
Recall that the Lie bracket of vector fields f and g is
defined as

have constant dimension on U and are all involutive, and


Gn_ has dimension n - 1 on U.
where a4g(x) := g(x). The Lie bracket is invariant under
coordinate transformations. A distribution A is said to be The proof is in the Appendix.
involutive if f, g E A implies [f, g] 6 A.
Remark 1. The condition that Go, ... , Gn - I all be invo-
Deriving conditions to transform a nonholonomic sys- lutive is somewhat redundant, as in the exact linearization
tem with two or more inputs into chained form is not
conditions in Isidori (1989), Section 5.2.6, as the involu-
difficult if we consider the method for exact linearization
of a nonlinear system with drift via state feedback and a tivity of some distributions in the sequence may imply the
coordinate transformation, as presented in Chapter 5 of involutivity of others.
Isidori (1989). Analogous to this method, we state and Remark 2. Recall that a system of the form X =

prove the following sufficient conditions for transforming 2:;:1 1 fi(x)u2 is completely controllable if the involutive
a three-input, drift-free, nonholonomic system to chained
closure of the distribution Ao =
span{f1, ... , ,f~} at each
form. The Appendix extends this to an (m + 1 )-input
configuration is equal to the entire state space Rn (see
system. Chow’s Theorem in Hermann and Krener (1977)). For
PROPOSITION 1. CONVERTING TO TWO-CHAIN, SINGLE- the case of three inputs, the system is completely control-
GENERATOR CHAINED FORM. Consider a three-input, lable, as the involutive closure of Ao =
span { f I, f2, f3}
drift-free, nonholonomic system is spanned by the vector fields fl, f2, adf,f2, ... , a~2f2,
f3.’ adf, F3, ... , a~3f3, which span Rn. Thus, the system
X = gl(X)Ul + g2(x)~2 + 93(X)U3 (3) in two-chain,. single-generator chained form (4) is com-
with smooth, linearly independent input vector fields pletely controllable, because controllability is unaffected
gl, g2, g3. There exists a feedback transformation on some by state feedback and coordinate transformation. This
open set U c R&dquo; argument easily extends to the case of m + 1 inputs.

Example. Proposition I is now used to find the chained


form equations for the fire truck system. The vector fields
fl, f2, f3 that satisfy the conditions of the proposition are
to transform thesystem (3) into two-chain, single-
generator chained form

if there exists a basis fl,f2,f3 for Ao := span{gi,g2,g3}


that has the form

369
where fl =sec Olgl with gi as in eq. (1). These can be The coordinate transformation (~, C, 77) =
4b(x) from
considered as the original input vector fields after the equation (15)is computed as follows:
input transformation Ul u~ cos 01. In this representation
=

three of the states (~ i , 1>], 02) are controlled directly, so


their velocities are the inputs.
The distributions G, in eq. (6) are now constructed,
and their involutivity is checked. By way of notation,
define f4 := adt, f2, f5 := adf,f3 and f6 := acti,f2’
Recall
that x =
(Xl, YI, ~# 1 , 6 j , ~z, a2).

This is a valid coordinate transformation, as the matrix


a~ is nonsingular.
The derivatives with respect to time give the follow-
ing equations, showing the required state feedback to
transform the system equations into the two-chain, single-
generator chained form:

Remark 4. Using the fire truck system as a reference,


The distribution G2 has dxmension2 n - 1 5 on =
we can infer characteristics of chained form systems. The
U = {x : ~i -6’2,<~]~2,6’) 7~ ±~} C R6. It may also be
generating input, vi, is related to the driving velocity
verified that Go, Gi, and G2 are involutive on U and the of the system by UI =
sec 81 v ~ . For the fire truck, the
functions hi = ~1, h2 yl, and h3
=
6*2 satisfy eq. (14).
=
first chain corresponds to the first two axles (one steer-
Note that there is a lack of uniqueness in the h functions. able, one passive), and the second chain corresponds
to the third steerable axle. In general multiple-steering,
Remark 3. In the open set U c R6, the fire truck is not
in a jack-knife configuration-that is, the wheels are less multiple-trailer systems, the single-generator chain form
can be applied with the number of chains corresponding
than 7r/2 relative to their respective bodies, and the cab’s
to the number of steerable axles in the system. Tilbury
body angle is not ~~r/2. et al. (1994) show that such a composite system can be
converted into multiple-input chained form, but in gen-
2. This can be easily checked by showing that the five vector fields that
define G
eral, dynamic state feedback is needed to achieve the
2 along with f 1 are linearly independent; i.e.,
transformation.
When the attachment point between the cab and the
trailer is not located at the center of the rear axle of

370
the cab but at some distance off of the axle (kingpin of eq. (4). The first two algorithms use sinusoidal con-
hitch), as is commonly the case for cars pulling trail- trol inputs; one steers one level in the chains at a time
ers, the above procedure applies with no modification. in a step-by-step fashion, and the other steers all levels
The kinematic equations and the h functions needed for in the chains at once. The third algorithm uses poly-
transforming this system to chained form are derived as nomial inputs. Other methods for steering two-chain,
follows. single-generator chained form systems exist, such as
Refer to Figure 2 for the system states and parameters the method using piecewise constant inputs described
and let x = (xr, ~~, 0 1, 81, ~z, 02). The
three row vectors in Tilbury and Chelouah (1993) for the fire truck exam-
representing the nonslipping constraints have the form: ple. The method of steering nonholonomic systems using
piecewise constant inputs was first introduced in Monaco
and Normand-Cyrot (1992) as multirate digital control.
We have found that although all of these steering algo-
rithms will find a path between any two configurations,
the resulting trajectories look &dquo;nicer&dquo; for some methods
than for others. Therefore, the choice of method will de-
pend on the type of trajectory desired. The step-by-step
sinusoidal method is not recommended in practice, but is
included here only to emphasize the chained form struc-
In the same way as before, we can construct a control ture. The all-at-once sinusoidal method works well for
system with vector fields 91, g2, g3, which are orthogonal trajectories that have a reversal (e.g., parallel parking).
to flex) =
span { &dquo;&dquo; 1 (x), &dquo;,,2 (x), &dquo;,,Bx)}. We choose these The polynomial and piecewise constant methods work
vector fields such that the inputs have the interpretation well for trajectories without backups.
that ui is the driving velocity of the rear wheels of the Recall that we want to solve the following con-
cab, U2 is the steering velocity of the front axle, and u3 trol problem: Given (~0, ~°, r~°) and (~f , (f,,qf),
is the steering velocity of the tiller. The control system is find a control law v = (V I (t), V2(t), v3(t)) to steer
then written as (~(0), (0), &dquo;,(0)) = (~0, CO, 770) to (~(z’)~ ~(Z’)~ ~l(Z’)) =

(~f , ~f , r~f ) in the given time interval [0, T].


The step-by-step sinusoidal algorithm is an extension
of the algorithm for two-input systems in Murray and
Sastry (1993). The algorithm exploits the decoupling of
the two chains, allowing for simultaneous steering. The
main idea, considering for a moment only the C chain,
is that if v, - a sin ot and V2 = ~ cos ~c.~t, then (0
will have a frequency component at ~cv, (1 will have a
frequency component at (~ &horbar; 1)w, ..., and (< will have
a frequency component at zero. By simple integration
over one period, this yields net movement in (< while
It isstraightforward to check that eq. (7) reduces to the (0,&dquo; . &dquo; , (g- j return to their previous values. Thus, at the
original system in eq. (2) when L2 0. = ~th step in the algorithm, the states at the ~th level in the
The chained form coordinates for this system are found chain are driven to their final positions.
in an analogous manner to those for the original fire truck For the time interval [O, T], the tops of the chains are

system using steered with constant inputs

driving &xi;
, &zeta;
0 0 to their final positions. For level
0 and &eta;
Note that these are the same h functions we used for the l = 1, ... , n
,n
3 2 in the chains, the inputs
3 + 1, ... , n are

fire truck system with axle-to-axle hitching. set as

4. Steering Two-Chain, Single-Generator over the time interval [ ], where


lT, (l + 1)
T w
= 2&pi;/ and
T
Chained Form Systems are chosen such that
&alpha;, &beta; and &gamma;
In this section, we present three algorithms for steering
a system in the two-chain, single-generator chained form

371
Fig. 2. The fire truck with off-axle hitching.
.

This causes the states (f. and q< to reach their final val- system (4) is integrated over [0, T] starting at the given
ues and the first f coordinates in each chain-namely, initial state (~, (0, i7o)
with the inputs (8). The input
~o, (o, (£- 10 T/o, 71(- 1 -to return to their final val-
... , parameters are then found by selecting a value for a ~I
ues. When one of the chains is shorter than the other, that and taking the inverse transformation ~-1 (~~f , ~f , r~f ). In
chain’s input can be set to zero when there are no more practice, the final state is not within an E ball around the
states to steer. initial state. In this case, we connect the two given states
Although this method works well, it can be tedious in by a finite number of E balls and apply the above method
practice because of the many steps that are needed. The within each ball. It is not clear how to apply this method
trajectories that are generated consist of many segments when the transformation to chained form has singularities,
and do not always follow a direct path between the start because in the original coordinates the parameter E may
and goal configurations. be a function of x.
In Tilbury et al. (1993), an &dquo;all-at-once&dquo; sinusoidal The third steering method uses polynomial inputs and
method was proposed for the two-input case. This method was first presented in Tilbury et al. (1993) for two-input
was extended in Tilbury et al. (1994) for multiple-input systems and in Tilbury et al. (1994) for the multiple-input
systems and will be summarized here for the three-input case. For three-input systems, we set the first input, vl, to
case. In this method, only one step is used with all of the be constant over the entire trajectory and the other inputs
necessary frequencies set in the inputs as follows: to be Taylor polynomials:

Given initial and final states such that I(~O, (°,17°) - with the number of parameters on each input chosen to
(~f , ~f , r~f ) ~ < E for a small E, the local existence of be equal to the number of states in its chain. The time
these parameters was proven in Tilbury et al. (1993) for needed to steer the system is determined from the change
single-chain systems and is easily extendible to multiple- in the ~o coordinate, T =
~o - ~~ .
Integrating the chained
chain systems. The basic idea of the proof is to define form equations with inputs as in (9) and initial condi-
a local diffeomorphism, 0 : Rn ~ R’~, which maps tion (~(0),((0),77(0)) = (~o, (0,,qO), and evaluating
the input parameters (a.o, bo, ... , brt, , co, ... , cn3 except at time T, the parameters bfl, ,bn2’ co, ... , c,,, can be
... ,

a, to (if , (f, 17f), where 0(ao, bo, ... , bn2’ co, ... ,I Cn3) is found in terms of the initial and final states from setting
the value of (~(T), ~(T), 17(T)) when the chained form (~(T), ~(T), r~(T)) _ (~f , ~f , ~f ). This equation is easily

372
solved for the parameters bi, ci by inverting a matrix that fire truck from y, = 5 to zero, with all other coordinates
is linear in the parameters. starting and ending at zero. The inputs for this trajectory
For trajectories that requiregj 0= as in the parallel are found from the last step of the step-by-step sinusoidal

parking maneuver, the polynomial method will not work. algorithm, where the only state that must be changed is
An easy way to remedy this situation is to pick an inter- yi =
(2, the last coordinate of the C chain, or, from the
mediate point, with ~o not equal to the given ~o, ~o ,
and all-at-once sinusoidal method,
then plan the path in two pieces.

5. Simulation Results
The simulation of the fire truck system is performed on
the transformed system in the chained form with ao = &o = ~i = co = ci = 0, ai = 2, &2 ~ &horbar;1.5915,
and = 1 over the interval [0,27r].
To limit the size of the configuration states or the in-
puts, one must take into account the coordinate trans-
formation to chained form, because the simulation is

The transformed states are steered from an initial point performed on the system in chained form. From the equa-
tions of the fire truck system, the steering wheel angles
to a final point by using the transformed inputs as con- 3
can take values between -90 and 90 degrees. In reality,3
structed in the step-by-step sinusoidal steering algorithm,
the front steering wheel angle is limited to -45 degrees
the all-at-once sinusoidal steering algorithm, or the poly-
< ol < 45 degrees, and the tiller steering wheel angle
nomial steering algorithm. Then the inverse coordinate
transformation
is limited to -15 degrees < ~Z < 15 degrees. Limiting
these angles can be approximated in the simulation by
limiting the two chained form inputs v2 and v3. For the
limited inputs, we used the same input parameters as for
the regular inputs in equation (12), with w 1/3 over =

the interval [0,67r]. Figure 4 shows the x - y phase plot


with and without the control limitation and the corre-
sponding inputs. The effect of limiting the control inputs
is to stretch out the trajectory four times as long in the x
direction.
In Figure 5, the chained form and original input func-
tions needed to parallel park the fire truck as in Figure 3
with Lo = 1, and Li = 4 is calculated to extract the are shown. The chained form inputs are the open-loop
trajectory of the fire truck in the original coordinates. The control laws for the system in two-chain, single-generator
results are presented for the parallel parking maneuver chained form. The physical inputs, however, depend on
using sinusoidal inputs. To see the advantage of having the states of the system, as can be seen in the following
the extra steering wheel, we show both left and right
equations:
comer trajectories and a change-lane trajectory using

polynomial inputs for the fire truck and the same system
without tiller steering.
Because the coordinate transformation to chained form
is a diffeomorphism on the open set U =
(x : 6j -
~2~).~2,~! ~ ±-~-} } C Rn, we need to avoid these
singular points in the configuration space when picking
the initial and final configurations for the simulation. One
practical solution is to plan a path that does not start or
end the fire truck in a singular configuration, and then
rotate the resulting trajectory about the origin to yield the Polynomial control inputs are now used to show the
desired trajectory. This will be explained later. control design and performance differences when the fire
Figure 3 shows the parallel parking maneuver that truck does not have tiller steering. For this system, we
results from using the step-by-step sinusoidal steering used the coordinate transformation to chained form (from
method or the all-at-once sinusoidal steering method. In
the original coordinates, this corresponds to steering the 3. Data from Berkeley Fire Department Station No. 2.

373
Fig. 3. The parallel parking trajectory using sinusoidal inputs. The fire truck starts at yi = 5 and ends at zero, with all
other coordinates starting and ending at zero.

using h =
X2 and h2
Y2 in eq. (15) without the 71
= In all of the following simulations, the initial and final
chain and following the procedure as for the fire truck, or states of the two systems are the same.
following the procedure in Tilbury et al. (1993)), Figures 6 and 7 show the advantage of having tiller
steering when making a 90-degree turn. For both of these
trajectories, the fire truck goes through the singular point
01 =
’IT /2. To avoid this singularity for the right hand
turn, we simulated the trajectory with the initial and fi-
nal configurations at 7r/4 and -’IT /4, respectively, then
rotated the entire resulting path by ~r/4. The left hand
turn was simulated similarly, noting that the path is wider
than the right hand turn in the same intersection due to
the convention of driving on the right hand side of the
road.
The inputs with and without tiller steering for the right
hand turn trajectory are shown in Figure 8. We see that
which kinematically is a two-axle car pulling one trailer.
the steering velocity for the system without tiller steering,
For the fire truck in the chained form of equation (10),
which is just the steering velocity of the front wheels,
there are two chains, one of length 3 and one of length 2,
switches back and forth more. Of greater interest, how-
so we use the polynomial control inputs ever, is that the magnitude of the input u2 without tiller
steering is larger than the inputs of the fire truck with
tiller steering. In some sense, the control for the system
without tiller steering has to work harder.
Figure 9 shows how the tiller steering assists the ve-
to steer the hicle when it changes lanes on a freeway. Figures 10
system. For the fire truck without tiller steer-
five-state system), there is only one chain of length and 11 confirm our findings that the system without
ing (a
4, with the two inputs being tiller steering works harder than with tiller steering.
In addition, we notice that in the right hand comer
trajectories, the system without a tiller crosses over
into the lane of oncoming traffic. All of the above

374
Fig. 4. The parallel parking trajectory with and without control limitations using the same initial and final positions as
in Figure 3. The x - y phase plot shows the trajectories corresponding to the regular inputs (solid line) and limited in-
puts (dashed line). For the input plots, the driving velocity ul is the solid line, the steering velocity of the front wheels
U2 is the dotted line, and the steering velocity of the tiller u3 is the dashed line.

maneuvers are smoother for the fire truck, which justi- steering algorithms presented provide methods for open-
fies our initial hypothesis that the tiller adds maneuver- loop, point-to-point control without obstacles.
ability. The example used to illustrate the ideas of this article
was a fire truck. The kinematic equations were derived

using nonholonomic linear velocity constraints for the


6. Conclusions
nonslipping conditions of the wheels rolling on the road.
Sufficient conditions have been given for transforming a The system was transformed into a two-chain, single-
three-input, drift-free nonholonomic system into a two- generator chained form and steered using sinusoids and
chain, single-generator chained form via a coordinate polynomial control inputs. Simulation results compared
transformation and state feedback. In this special form, the fire truck to the same system without tiller steering:
the system was shown to be completely controllable and The fire truck has greater maneuverability and can exe-
easy to steer using sinusoidal input functions with in- cute the same maneuvers in a narrower space and with
tegrally related frequencies or polynomial inputs. The less steering effort.

375
376
377
Fig. 11. The inputs for the fire truck with and without tiller steering for the 75-degree right-hand turn trajectory shown
in Figure 10. The driving velocity Ul is the solid line, the steering velocity of the front wheels 2c2 is the dotted line,
and the steering velocity of the tiller U3 is the dashed line.

Future work includes controlling wheeled nonholo- Let n2 be the smallest integer less than n such that
nomic vehicles when there are obstacles present. Pre-
liminary work in this area is presented in Bushnell
et al. (1994).
and let n3 be the integer when dim Gn3-1 1 first drops by
two. Thus, a4,f2 and act;.
f3 each give new directions up
to some level, n3, when one chain saturates, and the other
chain continues to give new directions until the state
Appendix space is spanned:
Proof of Proposition 1: We first remark on the dimension
of each distribution, didim Gi. By construction,
do = 2 and by assumption, d~,_1 ~ r~ - 1. Because
GpCGIC...CGn_1,
with n3 + n2 + 2 = n - 1. Without loss of generality, this
proof will use the case n2 > n3. A basis for Gn- can

378
then be chosen, which is given by the first n2 brackets of Starting with the involutive distribution G~,3 _ 1, the
f1 with f2 and the first n3 brackets of f, with f3: distributions drop dimension by 2. For Gn3 -I, the vec-
removed from G~3 . The
tor fields ad~f2 and ad~3f3 are

c~i, d~2, ~Lf, ~2,..., dL~’~’~2


one-forms
and the
annihilate
fact that
Gn~ _ Gn3 .
1 C In addition, eq. (13)
(where f2 and f3 have been renumbered if necessary).
dL~~’~2 annihilates Gn3 gives
We note that because of the special form (5) of the
vector fields, none of the vector fields in G~L-1 has an

entry a/axl, thus which shows that annihilates G~,~ _ 1. There


dLfl’ -n3 h2
is one more function whose differential also annihilates
G~,3_l; we call it h3 and note that dh3 ~ ad~~3f3(x) :
:=

a3(X) # 0.
has dimension n on U. Then, because the distribution
Gn-i is involutive and of dimension dn_ i = n - 1 on U,
At the next step, the vector fields ad f, 13-1
f2 and
the Frobenius Theorem (Isidori 1989) proposes that there act;;i-lf3 removed from Gn3-1 to get the involutive
are

exists n - d~_ 1 = 1 smooth function hi : U ~ R such distribution G,,-2 of dimension d~,3_2 2(n3 - 1). The
=

that dhl - X =
0 for all X E Gn- and furthermore, dh1 . one-forms dh i, dh2, dLf, h2, h~, dh3 annihilate
... , dLf~z-n3
f1 (x) := aa (x) ~ 0. With f, in the special form of eq. (5), Gn3_2 C Gn3-1. To see that the one-forms dLfl2-~~+lhz
we can choose hl to be xl, which gives dhl - fl(x) = 1. It and dLf, h3 also annihilate G~3 _2, we use eq. (13) and the
can also be verified that none of the vector fields in Gn - i fact that dL~~Z-n3h2 and dh3 annihilate Gn3-1 to get
has an entry in the first coordinate, giving dh 1 ~ X 0 for=

all X E Gn-1. By the dimension argument, Gi G~,_1i


=

for i =
nz, ... , n - 2, thus dh1 1 Gi for i = n2, ... , n -1.
The distribution G.~Z _ 1 drops dimension by 1 by re-
from Gn-I and
moving the vector field adn,f2 Gnz. Be-
=

cause this distribution is involutive and dnz _ 1 = n - 2,


there exist two smooth functions whose derivatives span
G*2 -I’ One of these functions is hl, since it annihilates
Gn2 ~ Gn,- i . Let h2 be the second function independent Furthermore,
of hl, and note that dh2 - aciR2fz(x)
:= a2(x) # 0.

At the next step, the vector field aciR2 -I


f2 is removed
from G’2-1 to get the involutive distribution Gn= -2 of
dimension dn2_2 = n - 3. The functions dhl and dh2 This procedure continues to the distribution Go, which
annihilate G~-2 We will need the following property is annihilated by dhi, dh2, ...,
<iL~&dquo;~2,
dh3, ...,
of Lie derivatives to find the third smooth function that
annihilates Gn,-21
dLnj-’ h .
3 In summary, we have found three functions, h 1, h~, h3
such that

for a smooth function A. Using this property and the fact


that dh2 annihilates G~,2 _ ~, we find that dLf, h2 annihi-
lates G.~,,_Z, since

These three functions are used to define the chained


form coordinates as follows

Furtherrrrore,

This procedure continues to the distribution Gn3, which is


annihilated by dhl, dh2, dlfl h2, ...,
dLf~’--’~~-1 h2.
379
To verify that the above coordinate transformation
is valid, we show that it is a local diffeomorphism.
First, the derivative of the coordinate transforma-
tion is calculated with respect to x, a.
This is then
multiplied on the right by the nonsingular matrix M
whose columns are the r~ independent vector fields
ft,f2,ad~2,...,ad~f2,f3.adf,f3,...,ad~f3: &dquo;

Therefore, the input transformation

will result in the two-chain, single-generator chained


form (4). 1

PROPOSITION 2. CONVERTING TO m-CHAIN, SINGLE-


GENERATOR CHAINED FORM. Consider the drift-free
nonholonomic system

with smooth, linearly independent input vector fields gi.


There exists a feedback transformation on some open set
U c Rl-

to transform the system (16) into m-chain, single-


generator chained form
The functions a2(x) and a~(x) are nonzero by defini-
tion.Equation (14) is used to get the zeros in the matrix.
Using row operations, it can be shown that the above
matrix,
Am

V’~

is equivalent to a nonsingular diagonal matrix with

on the diagonal, and thus has full rank. Therefore, the Ja-
cobian matrix (8W /8x) must also be nonsingular locally,
which implies that (~, (, 1]) ~(x) is a local diffeomor-
=

phism and a valid coordinate transformation on the open


set U (see Isidori [1989], Proposition 1.2.3).
To compute the input transformation, we take deriva-
tives of the transformed coordinates, cancelling terms by if there exists a basis f,..., f,~+ ~ 1 for..::lo:=
using the zero entries of the above matrix (d~/c~a~) ~ M: span{gl, ... , g~,-~+I which has the form

380
Bushnell, L. Mirtich, B., Sahai, A., and Secor, M. 1994.
Off-tracking bounds for a car pulling trailers with
kingpin hitching. Proceedings of the IEEE Control and
such that the distributions Decision Conference, pp. 2944-2949.
Hermann, R., and Krener, A. J. 1977. Nonlinear control-
lability and observability. IEEE Trans. Auto. Control
22:728-740.
have constant dimension on U, are all involutive, and
Isidori, A. 1989. Nonlinear Control Systems. 2nd ed. New
Gn-l has dimension n - 1 on U. York: Springer-Verlag.
Monaco, S., and Normand-Cyrot, D. 1992. An introduc-
The proof follows the same method as in the proof of tion to motion planning under multirate digital control.
Proposition 1.
Proceedings of the IEEE Control and Decision Confer-
ence, Tucson, AZ, pp. 1780-1785.
Acknowledgments Murray, R. M., and Sastry, S. S. 1990. Grasping and ma-
The authors would like to thank Captain David Orth, nipulation using multifingered robot hands. Robotics:
Proceedings of Symposia in Applied Mathematics,
tillerman George Fisher, and driver Bob Humphrey at
41:91-128.
Berkeley Fire Department Station No. 2 for demonstrating R. M., and Sastry, S. S. 1993. Nonholonomic
their tiller truck. We also thank Richard Murray and Greg Murray,
motion planning: Steering using sinusoids. IEEE Trans.
Walsh for many inspiring discussions. In addition, we
Auto. Control 38(5):700-716.
would like to thank the reviewers for their suggestions
and comments. Tilbury, D., and Chelouah, A. 1993. Steering a three-
input nonholonomic system using multi-rate controls.
Proceedings of the European Control Conference,
References Groningen, The Netherlands, pp. 1428-1431.
Tilbury, D., Murray, R., and Sastry, S. 1993. Trajectory
Alexander, J. C., and Maddocks, J. H. 1988. On the ma- generation for the N-trailer problem using Goursat
neuvering of vehicles. SIAM J. Appl. Math. 48(1):38- normal form. Proceedings of the IEEE Control and
51. Decision Conference, San Antonio, TX, pp. 971-977.
Brockett, R. W. 1981. Control theory and singular Rie- Tilbury, D., S&oslash;rdalen, O., Bushnell, L., and Sastry, S.
mannian geometry. In Hilton, P. J., and Young, G. S. 1994. A multi-steering trailer system: Conversion into
(eds.): New Directions in Applied Mathematics. New chained form using dynamic feedback. Proceedings of
York: Springer-Verlag, pp. 11-27. the IFAC Symposium on Robot Control, Capri, Italy.

381

You might also like