You are on page 1of 18

Math.

Control Signals Systems (1994) 7:58-75


9 1994 Springer-Verlag London Limited Mathematics of Control,
Signals, and Systems

Nilpotent Bases for a Class of Nonintegrable


Distributions with Applications to Trajectory
Generation for Nonholonomic Systems*
R i c h a r d M. M u r r a y t

Abstract. This paper develops a constructive method for finding a nilpotent basis
for a special class of smooth nonholonomic distributions. The main tool is the use
of the Goursat normal form theorem which arises in the study of exterior differen-
tial systems. The results are applied to the problem of finding a set of nilpotent
input vector fields for a nonholonomic control system, which can then be used to
construct explicit trajectories to drive the system between any two points. A
kinematic model of a rolling penny is used to illustrate this approach. The methods
presented here extend previous work using the "chained form" and cast that work
into a coordinate-free setting.
Key words. Nilpotent vector fields, Nonholonomic control systems, Trajectory
generation.

1. Introduction

This w o r k was m o t i v a t e d b y the recent interest in t r a j e c t o r y g e n e r a t i o n for m e c h a n i -


cal systems with n o n h o l o n o m i c constraints. C o n s i d e r the p r o b l e m of steering a
m e c h a n i c a l system whose c o n f i g u r a t i o n space is a s m o o t h (C ~ n - d i m e n s i o n a l
m a n i f o l d M from an initial c o n f i g u r a t i o n x ~ to a final c o n f i g u r a t i o n x 1, subject to
a set of i n d e p e n d e n t k i n e m a t i c c o n s t r a i n t s h a v i n g the form

(o)j(x), 2 ) = 0, j = 1 , . . . , k. (1)

T h e o)~.'s are a basis for a c o d i s t r i b u t i o n on M which restricts the velocity of the


system to be zero in certain directions. W e a s s u m e the cots are s m o o t h a n d linearly
i n d e p e n d e n t over the ring o f s m o o t h functions. Such constraints can arise when two
surfaces roll a g a i n s t each other, such as the rolling between a wheel a n d the r o a d ,
o r in s p a c e - b a s e d systems where the t o t a l a n g u l a r m o m e n t u m of the system is
c o n s e r v e d ( a l t h o u g h strictly s p e a k i n g this l a t t e r case is n o t a "constraint," it can be
t r e a t e d with the same set of tools).

* Date received: 9 February, 1993. Date revised: 13 March 1994. This research was supported in part
by a grant from the Powell Foundation.
t Division of Engineering and Applied Science, California Institute of Technology, Pasadena, Califor-
nia 91125, U.S.A. murray@indra.caltech.edu.

58
Nilpotent Basesfor a Class of Nonintegrable Distributions 59

To study such a system, we convert the path-planning problem into a control


problem. Let A be a distribution of dimension m = n - k which is annihilated by
the constraints. We represent this distribution with respect to a local basis of vector
fields:
A = span{g I, g2, ..-, gin}, gi(x) ~ TxM. (2)
In coordinates the constraint one-forms can be written as a k x n matrix and the
gi's are a basis for the fight null space of this matrix. The path-planning problem
can then be restated as that of finding an input function, u(t) ~ E", such that the
control system

Z: Yr = g l ( x ) u l + "'" + gm(x)um (3)

is driven from x ~ to x 1. As a consequence of our assumptions on the coj's, the gi's


are also smooth and linearly independent. We restrict the analysis to a local
coordinate chart on M.
The conditions for the existence of a path between two configurations is given by
Chow's theorem. We let [f, g] be the Lie bracket between two vector fields,

ef
[f, 9] = ~ f - ~x o'

and define the involutive closure of a distribution A as the closure of A under Lie
bracketing. Chow's theorem asserts that if the involutive closure of the distribution
associated with (3) spans T~M at each configuration, the system can be steered
between any two configurations. It is not apparent how the path can be explicitly
constructed: the trajectory-generation problem is to construct explicitly paths be-
tween two points which everywhere satisfy the constraints.
A set of kinematic constraints is holonomic if the constraints restrict the motion
of the system to a submanifold of dimension n - k. In this case the constraints on
the system can be rewritten as an algebraic constraint on the configuration variables
x. The constraints are nonholonomic if they do not constrain the system to lie on a
submanifold of the same dimension as the input space. In particular, we are most
interested in constraints which are maximally nonholonomic: any point in the config-
uration space can be reached. This is equivalent to saying that the corresponding
control system is controllable. If the constraints are not maximally nonholonomic,
the system can still be analyzed by restricting the initial and final configurations to
lie on the same leaf of the foliation generated by the distribution. A control system
which has the form given in (3) is called a nonholonomic control system.
The general problem of finding a feasible trajectory for a nonholonomic system
is difficult. For certain application areas--such as mobile robotics--fast, efficient
algorithms are required and one is often willing to sacrifice conditions such as
optimality if the tradeoff is finding a reasonable path in an efficient fashion (see
[ L J T M ] for a review of nonholonomic motion planning in this context). Motivated
in part by these applications, there have been recent efforts to understand and
exploit the structure of nonholonomic systems using tools from differential geomet-
ric control theory. In particular, the use of nilpotent vector fields has appeared
60 R.M. Murray

repeatedly in the work of Lafferriere and Sussmann [LS2], Jacob I-J1], [J2], and
the author [MS].
We recall the definition of a nilpotent basis for a distribution. Let adI 9 = [-f, 9]
and define ad} 9 = If, ad}-1 9]. Given a basis of vector fields {91, ..., 9m}, the basis
is nilpotent if an integer k exists such that all Lie products ad~, 9j = 0 for all i, j =
1, ..., m. The integer k is called the order of nilpotency. The advantage of using a
nilpotent basis to represent a distribution is that certain computations are greatly
simplified since high-order brackets are identically zero.
Recently a very elegant and general motion-planning method for nilpotent sys-
tems has been developed by Lafferriere and Sussmann [LS2]. One of the main tools
in their method is the Baker-Campbell-Hausdorff formula, which describes the
composition of flows of vector fields as a single flow,
(~ X1 0 (~ X2 = (~ XI +X2+(I/2)[XI'X2]+''',

The method consists of first converting a desired motion into a flow of the form
~OCtlXI'+Ot2X2+at3[XI,Xz]"[" ...

and then using the Baker-Campbell-Hausdorff formula, in combination with a


Hall basis for the Lie algebra generated by {X~}, to choose piecewise constant inputs
that achieve this flow. If the system is nilpotent, the Baker-Campbell-Hausdorff
formula has a finite number of terms and it can be shown that the process terminates.
The connection between the flow coordinates c~and the original coordinates for the
problem is made using a product expansion for the Chen series [$2], which allows
c~to be computed by simple integration. A similar approach has been used by Jacob
[Jll, [J2] using a different choice of basis for the Lie algebra, namely a Lyndon basis.
Somewhat related to these approaches, our own work has focused on the use of
specific choices of coordinates such that the control system can be explicitly inte-
grated by quadratures. The basic technique is to search for a feedback transforma-
tion which puts the system into a special nilpotent form (called chained form), and
then use sinusoids at integrally related frequencies to drive the system between states
[MS] (the use of integrally related sinusoids is motivated by the work of Brockett
[B]). In the case of two-input systems, constructive sufficient conditions for the
existence of a nilpotent basis for the associated distribution were given in [MS]. In
this paper we extend that work to give a simple rank condition which guarantees
the existence of a nilpotent basis for two-dimensional distribution and show how
the resulting basis can be applied to solve the trajectory generation problem.
The general problem of finding a nilpotent basis for a distribution has been
studied by Hermes et al. I'HLS]. They present a set of necessary conditions for the
existence of a local nilpotent basis and give two sets of sufficient conditions, one
based on solving a partial differential equation and the other based on Darboux's
theorem. The result presented here also relies on solving a partial differential
equation, but the existence of a solution is guaranteed by the rank condition.
This paper is organized as follows: Section 2 presents a rank test for determining
if a nonholonomie distribution admits a nilpotent basis. We concentrate on the
two-input case and present a constructive procedure for finding such a basis. Section
Nilpotent Basesfor a Class of Nonintegrable Distributions 61

3 gives an application of these results to the trajectory-generation problem described


above. We use as an example the case of a rolling penny [BRM]. Finally, we discuss
the implications of these results to other problems and describe areas of future work.

2. Nilpotent Bases for a Class of Nonholonomic Systems

In this section we present a set of necessary and sufficient conditions for a two-input
nonholonomic control system to have a nilpotent basis of a certain special form.
The conditions lead to a constructive method of finding a feedback transformation
which converts the nonholonomic system into a nilpotent, nonholonomic system.
This extends the work initiated in [MS] by giving a set of necessary and sufficient
conditions for the existence of a feedback transformation which converts a system
into chained form.

2.1. Pfaffian Systems


The proofs in this paper make extensive use of tools from the theory of exterior
differential systems. We present a brief overview of the relevant concepts here. A
detailed description can be found in the monograph by Bryant et al. [BCGt].
Let M be a manifold of dimension n with cotangent bundle T ' M ; let f~P(M)
denote the set of smooth exterior p-forms on M. Define f~*(M) as the set of smooth
exterior differential forms of all orders on a manifold M, f~*(M) = @~P(M). An
exterior differential system is given by an ideal J c f~*(M) that is closed under
exterior differentiation.
A Pfaffian system is an exterior differential system which is generated by a set of
linearly independent one-forms

I = {~z1. . . . . c~s}.

We call s the dimension of the system (assumed constant).


The exterior derivative induces a mapping ~: I ~ f~g(M)/I:

6:)~+d2 mod I e ~2(M).

The mapping 6 is a linear mapping over C~~

6(f~ + gfl) = df /x a + f d~ + do ^ fl + f dfl m o d I

= f da + g dfl modI

= fS(a) + Oh(r).

It follows that the kernel of 5 is a subbundle of f~(M) (i.e., at each point p s M, the
kernel of 6 is a linear subspace of T ' M ) . We call this subbundle I (1), the first derived
system of I:
I m = ker 6 = {2 ~ I : d 2 m o d I = 0}.
We represent 1 (1) using a basis of one-forms and hence I ~ is a Pfaffian system.
62 R.M. Murray

Since I (~) is itself a Pfaffian system, we can continue this construction and generate
a filtration

I = I (~ ~ I (~) ~ .-- z I ts) . (4)

If the dimension of each I (i) is constant, then this construction terminates for some
finite integer N. In this case we call (4) the derived flag of I and N the derived
length.
The derived flag describes the integrability properties of the algebraic ideal
generated by I. If I is completely integrable, then by Frobenius's theorem we have
I (1) = I (~ i.e., the length of the derived flag is zero. In fact, I (N) is always integrable
since by definition dI (s) m o d I (m - 0. I (m is the largest integrable subsystem con-
tained in I. Thus if I (m is not zero, then independent functions hi, ..., hr exist such
that {dh~} c {I}. In the context of control theory, this means that the system is not
controllable since algebraic functions exist which provide foliation of the state space
and it is impossible to move from one leaf of the foliation to another. The converse
of this controllability result is provided by Chow's theorem [C]:

Theorem 1 (Chow). Let I = {~1 . . . . . ct'} represent a set of constraints and assume
that the derived fla 9 of the system exists. Then a path x(t) between any two points
satisfyin9 cd(x)2 = 0 for all i exists if and only if an N exists such that I (m = O.

2.2. Goursat Normal Form

The main results of this section are a direct consequence of the following theorem,
whose proof can be found on pp. 54-57 of [BCG*].

Theorem 2 (Goursat N o r m a l Form). Let U be an open subset of ~" and let


I = {~1,..., ~s} be a collection of smooth, linearly independent one-forms defined on
U. I f a one-form 7z ~ 0 rood I exists such that

d~ i ------- --(Zi+1 A ~ m o d cd, ..., cd, i = 1 , . . . , s - - 1,


(5)
d~ ~ r 0 m o d I,

then a set of coordinates x exists such that

I = {dx,-x,_ 1 dxl,..., dx3-x2 dxl}.

Remark 1. A set of vector fields which is perpendicular to I is given by

01 = ~ + x2 ~ +"' + x,-1 ~-~x,' 92 - ax 2. (6)

These are precisely the vector fields which appear in the two-input, single-chained
form described in [MS]. We say that a nonholonomic system with vector fields as
in (6) is in chained form.
Nilpotent Bases for a Class of Nonintegrable Distributions 63

Let Z be a nonholonomic control system with smooth, independent, input vector


fields 9~ and 92- Define two filtrations:

E o = A, Fo~A,

E I = E o + [Eo, Eo], F1 = F o + EFo,Vo],


E 2 = E 1 + [El, E l i , F2 =F, + EL, Fo], (7)

Ei+l = Ei + [E;, Eli, F,+t = F, + EFt, Vo],

Under the assumption that each distribution is constant rank, the filtrations have
finite length (possibly different).
The filtration {F~} is the one which usually appears in the context of nonlinear
controllability and feedback linearization. In particular, Fz consists of all brackets
up to order i. The distribution Ei also contains all brackets of order i, but may
contain additional Lie products of higher order. This is due to the recursive
construction of E~, as opposed to the iterative construction of Fv The filtration El
is precisely the sequence of distributions which is perpendicular to the derived flag
of I = A•

Theorem 3. Given a two-dimensional distribution A such that

dim E i = dim F~ = i + 2, i = 0 . . . . . n -- 2,

a local basis {ex, ..., e~} and a one-form rc exist such that the Goursat congruences
(5) are satisfied for the differential system I = A •

The proof of this result is based on a suggestion by W. Sluis. The fact that a count
of the dimensions of the derived systems I (~ (or equivalently the filtration {E~}) is
not sufficient to guarantee equivalence to the Goursat normal form has been pointed
out by Giaro et al. [ G K R ] , who constructed a counterexample in five dimensions.
They also note that if dim El = i + 2, then the Goursat conditions (5) hold on
an open dense set. F o r the applications considered here, the stronger result of
Theorem 3 is needed. The proof relies on the following lemma.

Lemma 4. Under the assumptions o f Theorem 3 a basis { ~ . . . . , ~ } and a set o f


one-forms 2 i exist such that 2 i mod I =~ 0 and

d~ ~-=)? ^ a 2 m o d a 1,

d~ 2 = 2 2 A a 3 m o d ~ l , a2,

(8)
dcd -~ -= 2 ~-~ A a s mod a ~, ..., ~-~,

d~ ~ ~ 0 rood L
64 R.M. Murray

Proof. Choose a basis for I which is adapted to the derived flag:


i(o = {~, . . . . , ~s-,}.

Since {ai} is adapted to the derived flag, we have immediately that the congruences
(8) are satisfied and da i ~ 0 rood ~1 . . . . . a/. F o r later use, we also note that any
transformation of the form

~i = aio:i + ~ biaS, ai > O,


j<i

also satisfies the congruences (8) with ~i = (a//ai+l)2/. It follows from the definition
of {E/} that ,s-i(E~) = 0 for j < i and that, for any X, Y ~ Ei,

daS-i(X, Y) = X o ~ S - ' ( y ) _ yas-i(X ) _ ~ - i ( [ X , Y])


= - - g ~ - ; ( [ X , Y]).

We show 2 ~-~ m o d I # 0 by contradiction. First note that F~ c E / a n d since they


have the same dimension we must have Fz = Ev This in turn implies that

Ei+ 1 = E i + [E~, E0]


and hence every Z e E~+I such that Z r E / h a s the form Z = W + IX, Y] for some
IV,, X r Y ~ E0. Thus some such X, Y with Z = IX, Y] ~ Ei exists and
d ~ - i ( X , Y) = - ~ - i ( Z ) # O.

N o w i f U -i rood I = 0, then 2"-i(Y) = 0 and


dcd-~(X, Y) = (;t s-~ A a~-i+l)(X, Y)
= ;.~-,(x)~-,+~(y) _ ~-i(y)~-/+~(x)

= O,

which is a contradiction. Hence 2 s - i m o d I # 0. 9

P r o o f of Theorem. C h o o s e a basis for I as in the p r o o f of L e m m a 4. If we define


n = ,P-~, then (8) has the desired form d o w n to i = s - 1 and = is determined
rood el, ..., ~ . We n o w proceed by induction. Assume
doci = Tr ~i+1 mod ~i, ..., ~i

and that n is determined mod a i . . . . , a~+i. Then


da ~-~ - 2 i-~ /, ~ mod ~ i , . . . , ~ - i

and taking the exterior derivative and wedging with ~i gives


(d2o~ i - 1 ) A O~i -~ 0 ~-- . + . ) / - 1 A 7~ A ~i A a i+1 m o d a 1. . . . . a i-1

and hence
2 i-I = an + bo~ i -4- c ~ I+1 rood ~ , ..., ~/-1.

F r o m the lemma we have that 2/-1 m o d I # 0 a n d therefore a 4: 0. Hence we can


Nilpotent Bases for a Class of Nonintegrable Distributions 65

define a new one-form g as

= azc + ba ~ + ce i+I.

By scaling 5 ~. . . . . es, the congruences d o w n to i are preserved with ~ replacing n.


Furthermore, we obtain

dcd - ~ = ~ A c d mod5 ~,...7~-~

and ~ is determined m o d 5 ~. . . . . 5( This completes the induction. 9

To explore the need for the assumption on the dimension of Fi, we further examine
the relationship

= a n + bO~i + CO~i+1.

W i t h o u t loss of generality we can take b = 0 since it does not affect the congruences
at this level. If we do n o t require )J -1 rood I ~ 0, then we c a n n o t define ~ as before.
There are two possibilities: a = 0 or a = 0 at a point. It turns out that a = 0 leads
to a contradiction. To show this, we assume that c = 1 by scaling the cd's appropri-
ately, and hence we have

d 5 i - 1 ~ 5 i+1 A 5 i rood 0~1~ . . . , 5 i-1.

We apply the s t a n d a r d trick: take the exterior derivative and analyze the resulting
relationships. T a k i n g the exterior derivative

0 = d 2 5 i - 1 ~-- d 5 i+l A 5 i - - 5 i+1 A d 5 i + fl A 0~i+1 A Cti mod 51, . . . , 5 i-1,

where the additional term accounts for the possibility of adding a d5 ~-1 term due
to the congruence. N o w take the wedge p r o d u c t with 5 ~+1 to obtain

d 5 i+l A 5 i A o~i+1 = 0 m o d 51, ..., 5 i-~.

Shifting the wedge p r o d u c t s into the mod, we have

d5 ~+~ = 0 m o d 51, . .., cd, a i+1,

but this contradicts the fact that dcd +1 ~ 0 rood ~ , . . . , a~+1 and hence a ~ 0.
Thus the only possibility is that a = 0 at a point (or on a submanifold), but not
in a neighborhood. The extra assumptions on the dimension of the iterative filtra-
tion {Fg} ensure that this does n o t happen. This is illustrated in the following
example, which is similar to that constructed by G i a r o e t al. [ G K R ] .

Example 1. Consider the following distribution:

g' = + + +

a
gz - a x 2
66 R.M. Murray

The filtrations corresponding to this distribution are defined in terms of the brackets
a 0
- x4--, ad2~ ga - x3 , ad2~ gl = O,
ado~ g2 OX3 OX5 63X4 t~X5
0 a
ad~, gz = - 2 x 2 ~ - , [ado, g2, ad 2, 92] = 2 = - - .
UX 5 OX 5

The filtrations {E,} and {F~} match for i = 0, 1, 2. At the next level we have
F3 = F2 + {adgl 92},
E3 = F2 + {ado3, g2, [adgl g2, ad~l g2]}-
The filtration (Ei}, which is perpendicular to the derived flag, has constant dimen-
sion (2, 3, 4, 5) but the distribution F 3 looses rank on the submanifold x 2 = 0. Hence
a basis and a one-form ~ such that the Goursat congruences are satisfied do not
exist. This shows the nature of the obstruction that can occur when the dimension
count is correct only for the Goursat filtration {E~}.

2.3. Application to the Control Problem


The Goursat normal form can be used to give conditions under which a system is
feedback transformable to a system in chain form. The following proposition follows
immediately by application of Theorems 2 and 3.

Proposition 5. A feedback transformation which puts a nonholonomic system into


chained form (6) exists if and only if
dim E i = dim Fi = i + 2, i = 0 . . . . . n - 2.

Thus, the class of two-input systems which can be converted into chained form
(with a single chain) is completely characterized by an algebraic test. It is interesting
to note that although chained systems are not generic, they occur frequently in
applications. F o r example, all two-input, regular nonholonomic systems in three
and four dimensions are locally feedback equivalent to a system in chained form.
This follows from Darboux's theorem and Engel's theorem, respectively, and is a
special case of the results presented above (see pp. 50-51 of [ B C G t] for a proof of
Engel's theorem). These special cases have also been pointed out by Hermes [H].
Examples of systems which can be converted into chained form include a disk roiling
on a plane (rolling penny) [BRM], a kinematic model for a car [MS], and a
kinematic car towing N trailers [S 1].
The results of Proposition 5 provide a set of sufficient conditions for the existence
of a nilpotent basis for a two-dimensional distribution. The conditions are obviously
not necessary since a distribution might admit a nilpotent basis even though the
growth conditions in Proposition 5 are not satisfied. One simple example is
a ~ O
gl = ~ + Xz~x3 4- X3~X4,
a
g2 ----
~ q'- X 3 ~ ,
Nilpotent Bases for a Class of Nonintegrable Distributions 67

/
Fig. 1. Penny rolling on a plane.

which has a derived flag with dimensions (3, 2, 0) and hence dim E, = dim F i ~ i + 2.
Even if we restrict the growth of the distribution so that dim Fi = i + 2 the condi-
tions are only sufficient, as the following vector fields show:

0 0 0 0 ~6

a
g2 - - OX 2-

For this distribution the filtration {Fi) grows at the proper rate, but E4 does not
satisfy the dimension count.
One disadvantage of using Proposition 5 is the need to find an explicit change
of coordinates which puts the system into chained form. Even in relatively low-
dimensional cases, the computations can become extremely complex.

Example 2 (Rolling Penny). Consider the kinematic model of a penny rolling on


a plane, as shown in Fig. 1. Let x e ~4 denote the configuration of the penny, with
(xl, x2) being the location of the penny on the plane, x 3 the angle that the penny
makes with a fixed line on the plane, and x4 the angle of a fixed line on the penny
with respect to the vertical. We take the radius of the penny as 1. The constraints
for the penny are that it roll in the direction it is pointing, with no slipping:

a 1 = cos x 3 d x ~ + sin x 3 d x 2 - dx4, (9)

Lljrc~
~2 = sin x 3 d x 1 - cos x 3 d x 2. (lO)
Converting this to a control system gives

22 =|sinx3|
23 [ 0 |Ul + u 2. (11)
24

F o r this choice of vector fields, ul corresponds to the forward velocity of the penny
and u2 corresponds to the angular velocity of the penny about the plane's normal.
A simple calculation shows that

ad~ g2 = - a d ~ g 2
68 R,M, Murray

and hence these vector fields are not nilpotent. However, Proposition 5 can be used
to find a change of input which makes the system nilpotent.
The first step is to compute the derived flag for the exterior differential system.
The exterior derivatives of el and c~; are given by
de 1 = - s i n x 3 d x 3 A d x l + cos x 3 d x 3 A d x 2, (12)
de 2 = cos x 3 d x 3 A dx 1 + sin x 3 d x 3 A d x 2. (13)
It is easy to check that the following relationships hold:
d a 1 /x ~1 = _ d x 1 /x d x z /x d x 3 + cos x 3 d x 2 /x d x 3 /x d x 3
- sin x 3 d x 1 A d x ~ A d x a ,
d a 1 A cd A C~2 = 0 ,

do~2 /x od /x o~2 = - d x 1 ^ d x 2 i", d x 3 /x d x , # O,

and hence {el, cd} is a basis adapted to the derived flag:


I m) = {cd, ~2},
i<1) = {~1 ), (14)

V~'= {0}.
The dimension of the derived flag is (2, 1, 0), implying that the system is completely
controllable. Furthermore, it is easy to check that the conditions of Theorem 3 are
satisfied and hence the system admits a nilpotent basis.
Next, we search for a one-form 7r which satisfies the G o u r s a t congruences. Since
this system is low-dimensional, the algebra needed to find zc is straightforward.
Define a 3 and ~* so as to complete the basis:
cc3 = cos x 3 d x , + sin x 3 d x 2,
OC4 = d x 3.

rc must satisfy
dod = - c d / x 7z m o d ~1. (15)
Setting n = 23c~3 + 24 a4, (15) gives

"~a i =*" 7t = dx3"

Following the p r o o f of the G o u r s a t theorem and solving a set of first-order partial


differential equations (see pp. 54-55 of [BCG*]), the following change of coordi-
nates is used to put the system into Goursat normal form:

21 ~-- X 3 ,

g2 = - - X 1 COS X 3 - - X 2 S i n X 3 ,

z3 = - x l sin xz + x2 cos x3,


z,, = x 1 c o s x 3 + x 2 sinx 3- x 4.
Nilpotent Bases for a Class of Nonintegrable Distributions 69

In this new set of coordinates, the one-forms become


~i = dz 4 _ za dzt,
cd = - ( d z 3 - z 2 dz,)
and this is easily converted to a Goursat basis by negating ct2. Let ~ represent the
chained vector fields in Goursat coordinates; in the original coordinates these vector
fields have the form

g,(x) = \ ~ x J (h(o(x)),

where ~o: x ~ z is the Goursat change of coordinates. Performing this calculating,


the nilpotent control system has the form
--COSX3~
x*
22 = [ - s i n xa(x 2 COSlX3 -- xl sin X3) /

I1 23
X4 L --(X 2 COS Xa -- X1 sin x3) '
]
vl +

Here v2 corresponds to the negative of the angular velocity of the penny about the
v2. (16)

plane's normal and v, is a linear combination of v2 and the forward velocity of the
penny.

2.4. Systems with more than Two Inputs


Partial extensions of the results presented above to systems with m o r e than two
inputs are available using the recent work of Gardner and Shadwick [GS]. Gardner
and Shadwick have proposed an algorithm for M I M O feedback linearization which
uses an extension of the Goursat theorem presented above.

Theorem 6 (Extended Goursat Normal Form). Let U be an open subset o f ~n and


let I = {@, ..., ~]~:j = 1. . . . . m - 1} be a collection of smooth, linearly independent
one-forms defined on U. Let I (I~ denote the j t h derived system of I. I f a single, exact
one-form n v~ 0 rood I exists such that, for all j,
d~j = _~j+i /~ n rood I (~'-0, i = 1. . . . . sj -- 1,
(17)
d~t]J ~ 0 mod I,
then a set of coordinates x exists such that
I = (dx~+2 - x~.+, d x , . . . . ,dx~ - x~ dx,: j = 1 , . . . , m - 1}.

Remark 2. A set of vector fields which is perpendicular to the ideal I is given by


a a ~
01- - -Z+ + .-- + xJn-1 ~o x nJ' gj+l = ~x~' j = 1..... m- 1. (18)

These vector fields are a special case of the multi-input chained form vector fields
described in [MS].
70 R.M. Murray

The exterior differential system Ij = {@, ..., e/} is called the jth tower of the
complete system. Theorem 6 is considerably more restrictive than the original
Goursat theorem since it requires a very limited coupling between the different
towers. This is enforced by the independence of the mod operations in Theorem 6
and the use of a s i n g l e one-form 7~which connects the towers. In the case of M I M O
feedback linearization, n becomes dt, the differential of time.
In the two-input case Theorem 3 was used to show that the Goursat conditions
could always be satisfied if a certain dimension count was satisfied. This is not true
when more than two inputs are present, since the growth of control Lie algebra can
be much more complex.
An example of a system with more than two inputs is the "fire-truck" example
presented in [BTS].

3. Application: Trajectory Generation

The results of the previous section can be used to generate a feedback transforma-
tion which renders a nonholonomic control system nilpotent. Recent work by
Lafferriere and Sussmann [LS2] and Jacob [J1], [J2] has given insights into how
to generate trajectories for nilpotent control systems. We review those results here
and combine them with some of the steering methods presented in [MS] to obtain
a technique for trajectory generation for chained systems. The methods presented
here do not rely on finding the coordinate transformation which inputs the system
explicitly into chained form, but only on the input transformation which renders
the system nilpotent.
Following the results of the previous section, we consider control systems of the
form

Yr = g , ( x ) u ~ + ' " + 9,.(x)u.,,

where the vector fields

gl,

g2, adol g2, ad~ g2 . . . . . ad~ 92,

g,,, ado1 gin, ad21 g,,, ---, ad~]" gm


span T , , M , n = 1 + n2 + " " + n,,, at each point x ~ U, and all other Lie products
between the vector fields are identically zero. These are precisely the class of systems
which are generated by the chained-form theorems of the previous section and hence
constructive, sufficient conditions for finding a change of basis (precompensator)
which renders the system nilpotent are available.
To steer nilpotent, nonhotonomic control systems with a chained structure, we
use exponential coordinates. Let (0[ represent the flow along a vector field f for h
units of time. If f i s smooth and everywhere nonzero, q~[ exists for h sufficiently
small and (p[: •" ~ R is a diffeomorphism of the state space onto itself. The flow
Nilpotent Bases for a Class of NonintegrableDistributions 71

satisfies the differential equation

Now let {BI . . . . , Br} be a Hall basis of order k generated by the symbols 9i- We let
bi denote a basis element Bi which has'been evaluated on the vector fields 91, .--, gin.
Concatenating the flows of the Bi we obtain a map ~: h ~ x,
9 (h) = ~b,h . . . . ~b.h~(xO). (19)

This map is surjective in a neighborhood of h = 0 and is a local diffeomorphism in


the case that r = n. By an abuse of terminology, we call h the exponential coordinates
adapted to the basis {Bi} (even though 9 is not necessarily injective).
Our goal is to find an h 1 such that x 1 = ~(h 1) and then solve the problem in h
coordinates. This entails finding a u such that the original system is steered to the
desired h coordinates. The power of this approach is that we never have to solve
explicitly for the mapping (I). It suffices to consider the differential equation which
h must satisfy and solve the problem in that context. Computationally, this can be
done in such a way as to require nothing more than simple integration and the
solution of algebraic (typically polynomial) equations. The foundations of this
technique can be found in the work of Sussmann [LS2], [S2] and have been used
in much the same form as given here in the recent work of Jacob [J2]. We present
only the case where the Lie algebra generated by 91 and 92 has the-chained structure
derived in the previous section, although the general case follows in essentially the
same way.
To compute the value h(1) which corresponds to the desired final state of the
original system, we make use of the "extended system" defined by Lafferriere and
Sussmann:
2 = bl(x)vl + b2(x)v2 + " " + br(x)v,. (20)
By assumption, the vector fields b l , . . . , br are full rank. Given any path x(t), we can
solve for some v(t) which satisfies (20). This corresponds to choosing how to project
the velocity of the system onto the Hall basis elements; ifr = n this choice is unique,
otherwise not.
We can find the exponential coordinates corresponding to this motion by differ-
entiating (19). This yields a set of differential equations for hl, as follows:

h2 ~ V2~

[~a = h2vl + v3, (21)


h4 = 2gh2vl
2 -ff h2va + v4,

f~. _ _ _ 1 h~-2vl + " " + h2v._ 1 + v..


(n - 2)!
72 R.M. Murray

A complete derivation of these equations can be found in [$2]. Since v(t) is known,
this equation can be integrated to find h(1).
The task is now to find a feasible path which steers the system to the same point
in exponential coordinates. Since every feasible path corresponds to one in which
vj = 0, j > m, the problem reduces to finding v~, ..., v,, which steers (21) from
h(0) = 0 to h(1) = h ~. The simple form of (21) allows us to choose v(t) as a para-
metrized function (e.g., a Fourier or Taylor series in t), integrate the differential
equation to obtain a parametrized final state, and equate the final state to the desired
state. For Fourier and Taylor series parametrizations of the inputs, the final solution
can be reduced to finding the roots of a polynomial. We illustrate this in an example
below.
The system (21) can be shown to be diffeomorphic to a system in chained form
[TMW]. Hence, the choice of v(t) can be transformed into the problem of solving
the trajectory-generation problem for a system in chained form. Specific algorithms
for steering systems in chained form are presented in [MS].
The algorithm outlined here can be extended to the case of an arbitrary nilpotent
Lie algebra instead of the special chained form used here. In this case we again
proceed by using exponential coordinates and solving a polynomial control prob-
lem relative to that basis. The general case is discussed in [J2], [LS1], and [LS2].

Example 3. Consider again the rolling penny from the previous example. Let
x ~ (0, 0, 0, 0) be the initial configuration and let x 1 = (1, 2, -re/2, zr/4) be the
desired final configuration. Motion between x ~ and x 1 consists of a net translation
in the x and y directions, combined with a 90 ~ rotation of the penny relative to a
fixed line on the plane and a 45 ~ rotation of a fixed line on the penny with respect
to the vertical. We assume that system is nilpotent, as in (16).
We begin by constructing the exponential coordinates corresponding to x 1. Since
the Hall basis evaluated on the system vector fields has exactly four nonzero
elements, the extended system is uniquely defined as
Y: = g l v l + 9217,2 -b (adg 1 g2)/)3 "q- (ad~l g2)v4.
Choosing x ( t) = x ~ + t(x I - x ~ we can solve for v(t) and integrate (21) to determine
the exponential coordinates corresponding to x 1. Performing this calculation yields
h 1 ~ ( - 1.57, 1, - 3.57, 2.59).
We now construct an input which steers the system in exponential coordinates
from h ~ to h 1. Using the ideas developed in [MS], we choose
v 1 = a o sin rot + al sin 27ct,
(22)
v2 = b0 sin 7rt + bl cos 2~t,
which can be integrated symbolically (from the origin) to obtain
2
hi(l) = - a o ,
7Z
(23)
h2(1) = 2b0,
Nilpotent Bases for a Class of Nonintegrable Distributions 73

x2 xl

0 , , , , , I , , , I
1
x2

-i -! ,
i I . . . . I , , . , I
1 2 -1 I , , , I , . . . . . . . . . . . I
xl
x3

vl

-2.0 , , , I , , , l

X4
v2

-1
-20 I , , , I , , , I , , 9 I , , , I , , , I

0.0 0.2 0.4 0.6 0.8 1-0 0.0 0.2 0.4 0,6 0.~ 1.0
time time

Fig. 2. Motion of a rolling penny from x ~ = (0, 0, 0, 0) to x ~ = (2, 1, n/2, n/4). The upper left figure
shows the trajectory of the penny on the plane; the remaining plots show the states and inputs as a
function of time.

8
h3(1) = ~aobo + 3 ~ a l b o "+ l~--~2a2bo - ~_~albl,

h4(1) = a~b o + aoalbo + 1---~-~2aibo - a~b 1 - -4-~2aoalbl - ~ a~bl"

This is a p o l y n o m i a l equation in the input parameters. A solution of the equation


h(1) = h 1 (obtained using Mathematica) is given by
vt = - 2 . 4 6 7 sin :~r -- 2.273 sin 2rot,
(24)
v2 = 1.571 sin nt - 12.73 cos 2nt.

The m o t i o n generated by this set of inputs is s h o w n in Fig. 2.

4. Conclusions

In this paper w e have presented a set of necessary and sufficient conditions for
converting a two-input, n o n h o l o n o m i c control system into a special canonical form.
This result also gives constructive sufficient conditions for finding a nitpotent basis
74 R.M. Murray

for a nonholonomic distribution. The application of these results to steering a


control system with a nilpotent basis was also given.
Further research is needed to achieve an understanding of the conditions under
which a nonholonomic system with more than two inputs can be made nilpotent
by use of a nonlinear precompensator. The extended Goursat normal form theorem
used by Gardner and Shadwick gives one criterion but there are many other
possibilities. As an example, all systems which are in chained form (in its full
generality, as described in [MS]) are nilpotent, but 0nly a very special subset
satisfies the extended G_oursat normal form conditions. They correspond to systems
in which controllability is "generated" by a single vector field (which plays the dual
role of the one-form re).
As noted by Jacob [J21 path-planning techniques based on exponential coordi-
nates generated by a nilpotent basis do not necessarily rely on the system having
no drift vector field. Formally, the same techniques can be applied to systems of the
form
Yc = f ( x ) + gl(x)ul + ' " + gm(x)um
if the vector fields are nilpotent. However, the polynomial equations that result from
choosing a symbolic set of inputs as in Example 3 become much more difficult to
solve. Also, the problem of finding a nilpotent basis for a nonlinear control system
which is only affine in the inputs is considerably more difficult, since the drift vector
field f cannot be scaled. It is interesting to note that the Brunovsky canonical form
used in feedback linearization is nilpotent, and hence the MIMO linearization
conditions are sufficient conditions for finding a nilpotent basis for a nonlinear
system which is affine in the inputs.

Acknowledgments. The author would like to thank R. Gardner and W. Sluis of


the Fields Institute for Mathematical Sciences, Ontario, Canada, for many fruitful
conversations on the use of exterior differential systems for studying chained sys-
tems and for pointing out the connections between the Goursat normal form and
chained systems. In addition, Sameer Jalnapurkar and Michiel van Nieuwstadt
provided invaluable assistance in debugging the proof for Theorem 3 and the
anonymous reviewers provided many valuable comments that helped improve the
quality of the paper.

References

[BRM] A. M. Bloch, M. Reyhanoglu, and N. H. McClamroch. Control and stabilization of non-


holonomic dynamic systems. I EEE Transactions on Automatic Control, 37(11):1746-1757,
1992.
[B] R. W. Brockett. Control theory and singular Riemannian geometry. In P. J. Hilton and G. S.
Young, editors, New Directions in Applied Mathematics, pages 11-27. Springer-Verlag, New
York, 1981.
[BCG ~] R. L. Bryant, S. S. Chern, R. B. Gardner, H. L. Goldschmidt, and P_ A. Griffiths. Exterior
Differential Systems. Springer-Verlag, 1991.
[BTS] L. Bushnell, D. Tilbury, and S. Sastry. Steering three-input chained form nonholonomic
systems using sinusoids. Proceedings of the European Control Conference, pages 1432-1437,1993.
Nilpotent Bases for a Class of Nonintegrable Distributions 75

[c] W.-L. Chow, Uber systeme yon linearen partiellen differentialgleichungen erster ordnung.
M athematische Annalen, 117:98-105, 1940-194i.
[GS] R. B. Gardner and W. F. Shadwick. The GS algorithm for exact linearization to Brunovsky
normal form. IEEE Transactions on Automatic Control, 37(2):224-230, 1992.
[GKR] A Giaro, A. Kumpera, and C. Ruiz. Sur la lecture correcte d'un r6sultat d'l~li Cartan. Comptes
Rendus des S~ances de l'Acaddmie des Sciences, S~rie A, 287:241-244, 1978.
[HI H. Hermes. Distributions and the Lie algebras their bases can generate. Proceedings of the
American Mathematical Society, 106(2):555-565, 1989.
[HLS] H. Hermes, A. Lundell, and D. Sullivan. Nilpotent bases for distributions and control systems.
Journal of Differential Equations, 55:385-400, 1984.
[J1] G. Jacob. Lyndon discretization and exact motion planning. Proceedings of the European
Control Conference, 1991.
[J2] G. Jacob. Motion planning by piecewise constant or polynomial inputs. Proceedings of the
IFAC Symposium on Nonlinear Control Systems Design (NOLCOS), pages 628-633, 1992.
[LSl] G. Lafferriere and H. J. Sussmann. Motion Planning for Controllable Systems Without Drift:
A Preliminary Report. Technical Report SYCON-90-04, Rutgers Center for Systems and
Control, 1990.
[LS2] G. Lafferriere and H. J. Sussmann. Motion planning for controllable systems without drift.
Proceedings of the IEEE International Conference on Robotics and Automation, pages 1148-
1153, 1991.
[LJTM] J.-P. Laumond, P. Jacobs, M. Taix, and R. M. Murray. A motion planner for nonholonomic
mobile robots. 1EEE Transactions on Robotics and Automation, to appear.
[MS] R. M. Murray and S. S. Sastry. Nonholonomic motion planning: Steering using sinusoids.
IEEE Transactions on Automatic Control, 38(5):700-716, 1993.
[Sl] O. J. Sordalen. Conversion of the kinematics of a car with N trailers into a chained form.
Proceedings of the IEEE International Conference on Robotics and Automation, pages 382-387,
1993.
[S2] H. J. Sussmann. A product expansion for the Chen series. In C. Byrnes and A. Lindquist,
editors, Theory and Applications of Nonlinear Control Systems, pages 323-335. North-Holland,
Amsterdam, 1986.
[TMW] A. Teel, R. M. Murray, and G. Walsh. Nonholonomic control systems: From steering
to stabilization with sinusoids. Proceedings of the 1EEE Control and Decision Conference,
pages 1603-1609, 1992.