Documente Academic
Documente Profesional
Documente Cultură
CONTROL AND ÜPTIMIZATION © 1995 Society for Industrial and Applied Mathematics
Vol. 33, No. 1, pp. 126-148, January 1995 006
Abstract. This paper gives a general formulation of the theory of nonholonomic control systems on a Riemannian
manifold modeled by second-order differential equations and using the unique Riemannian connection defined by
the metric. The main concern is to introduce a reduction scheme, replacing sorne of the second-order equations by
first-order equations. The authors show how constants of motion together with the nonholonomic constraints may be
combined to yield such a reduction. The theory is applied to a particular class of nonholonomic control systems that
may be thought of as modeling a generalized rolling ball. This class reduces to the classical example of a ball rolling
without slipping on a horizontal plane.
* Received by the editors August 3, 1992; accepted for publication (in revised form) September 28, 1993.
t Department of Mathematics, The Ohlo State University, Columbus, Ohio 432!0, and University of Michigan,
Ann Arbor, Michigan 48!09. This author's research was partially supported by National Science Foundation grant
PYI DMS-9157556, and Air Force Office of Scientific Research grant F49620-93-l-0037.
t Center for Systems Science and Engineering, Arizona State University, Tempe, Arizona 85287. This author's
research was partially supported by National Science Foundation grant DMS-9!011964.
126
NONHOLONOMIC CONTROL SYSTEMS ON RIEMANNIAN MANIFOLDS 127
Caplygin systems. Our framework incorporates this setting as a special case, even though it
is an important case.
We illustrate our general prescription by introducing an extended example. We generalize
the case of a ball rolling on a flat plane with extemal forces parallel to the plane to push the ball,
as studied by Brockett and Dai [1993]. The generalization incorporates the well-known model
of a generalized rigid body, using a compact semi-simple Lie group, together with velocity
constraints, linking the generalized rotational motion to a generalized translational motion.
We call this system the generalized rolling ball. In the case that the ball is not symmetric
(i.e., has an inertia tensor not equal to a multiple of the identity), the nonholonomic velocity
constraints are not symmetries, and our reduction procedure generalizes that of Koiller [1988].
We also consider sorne controllability and optimality properties of the generalized rolling
ball in the symmetric case. The controllability result generalizes the local argument of Bloch
et al. [1992b], using a general controllability result on principal bundles. We also formalize a
minimum force control problem as the higher-order analogue of the minimum energy control
problem. In the holonomic case the minimum force control problem for second-order New-
tonian systems was studied in Noakes et al. [1989] and Crouch and Leite [1991a], [199lb].
These works treat higher-order variational problems on Riemannian manifolds. We outline
the modifications necessary to treat the nonholonomic case.
The outline of the paper is as follows. In §2 we give a general formulation of holonomic
control systems on Riemannian manifolds, and in §3 we give the formulation of nonholonomic
systems. In §4 we introduce the concept of symmetries and describe the general reduction
procedure based on the nonholonomic constraints and the constants of motion derived from
the symmetries. In §5, we have four subsections dealing with the control of the generalized
rolling hall, in which we describe sorne preliminaries on Riemannian structures on Lie groups,
the generalized rigid body, the generalized rolling ball in both body fixed and inertial axes,
and finally controllability and optimality questions.
2. Holonomic control systems and optimal control. In this section we give a brief
formulation of mechanical systems, without any velocity constraints, under the influence of
externa} forces. We broadly follow the formulation given by Hermann [1982]. We consider
the case of systems with "nonintegrable" velocity constraints in the following section.
Mechanical systems, in which the velocity constraints are integrable, are referred to here
as holonomic mechanical systems. These integrable velocity constraints yield constraints on
the configuration variables only, and thereby determine a manifold in which the configuration
variables are constrained to evolve (see Arnold [1978]). This provides the motivation for
considering holonomic control systems in the generality discussed below.
We let M denote a smooth (infinitely differentiable), n-dimensional manifold with a
Riemannian metric denoted /C(., ·).TM will denote the tangent bundle to M. The norm of a
vector X p E TpM will be denoted by
M denotes the configuration space of a mechanical system, whereas TM denotes the phase
(or state) space. The notion of an inertia tensor will be modeled by a bundle mapping
J : TM TM ,
such that J is the identity on M. Thus for each p E M we have a linear mapping
Jp : TpM TpM .
(X, Y) = K( J X , Y )
for all vector fields X , Y on M. We refer to K, as the ambient metric and (·, ·) as the mechanical
metric. The norm of a vector X p E TpM with respect to the mechanical metric will be denoted
by
The mechanical metric determines aunique Riemannian connection on M,denoted V', and
thereby determines a covariant derivative, denoted D/ Ot. A mechanical system is determined
by its kinetic energy T : TM ---. R given by the expression
T _ / dq dq )
q - 2 \ dt , dt
and the potential energy U : M ---. R given by an arbitrary smooth function on M. Denote
the cotangent bundle to M by T* M. An extemal force is modeled as a covector field F on
r;
M , in general time varying. Thus, Fp E M for each p E M . We define the momentum P
as a covector field along the trajectories of the mechanical system on M by setting
(1) := F - dU.
For each smooth vector field W on M , we have along the motion of (1)
D ( P(W ) ) - D / dq
at - at \ dt '
w) - / at w) + /\
D2q
-\ 2 '
dq DW )
dt ' at
= DP (W ) p ( DW ) = DP (W ) / dq DW ) .
at + at at + \ dt '
at
We deduce that
(2)
From the definition of kinetic energy and potential energy, equation (1) yields
d
(3)
F (
dq )
= U ( q( t) ) + d Tq( t)·
dt dt dt
1 :::; i:::; N, be input or control functions (real valued functions of time). We then model the
force field F by setting
N
(
4) Fq( t) = L ui(t) (Xi(q(t)), ·).
i=l
Let J* : TM -r T*M denote the bundle isomorphism determined on fibers by
From equations (2) and (4) we may rewrite equation (1) in the form
D2q A *-1
(5) 8t2 = w UiXi( q) - Jq dUq.
i=l
Equation (5) now represents a general holonomic mechanical system with inputs. From now
on we shall ignore the potential term in this equation, but it may be added without any extra
= =
difficulty. When F O (and U O) equation (5) reduces to the geodesic equations on the
Riemannian manifold (M, (·, ·))
D2q
at2 = o.
This flow is known to be an extremal of the variational problem (Milnor [1963]),
We now introduce a natural optimal control problem for the system (5). First we define
a norm on fibers of T* M in the usual way:
IF(Wq)I
(7) llFqll = sup llW ll .
IW.lk10 q K
Note that we use the ambient metric in this definition. We introduce the minimum force
control problem as
(8)
or equivalently we may specify Pq(O) and Pq(T ). From (2) and (7) we obtain
(10)
NONHOLONOMIC CONTROL SYSTEMS ON RIEMANNIAN MANIFOLDS 131
lt is now natural to consider the formulation (5) of the holonomic control system. lt is
convenient to assume a little more structure for the force field F defined in (4). We modify
the definition as follows:
N
where Xi , 1 :o::; i < n is an orthonormal base of vector fields with respect to the ambient
metric
(12) 1 :o::; i, j :o::; n.
With the force field (11) the system (5) may be rewritten as
D2 N
(13) at'! = L
J;¡ 1xi( q)ui( t).
i=l
j D2q D2q ) N 2
\ 8t2 ' Jq at2 = ui ( t).
lt follows that for system (13) the minimum force control problem is defined by the cost
functional
{T l N
(14) mjn Jo l L u¡( t) dt
o i=l
(15)
(16)
where R is the curvature tensor associated with the connection 'V. lt follows that for N = n,
the minimum force control problem introduced above is a natural higher-order version of the
classical variational problem (6), which is often interpreted as a rninimum energy problem.
In the control literature another class of optimal control problem has received much
attention. lt may be characterized by the cost functionals (14) subject to systems of the form
where xi are independent vector fields on M and Vi are control functions. Although these
problems are govemed by first-order rather than second-order equations (13), the singular na-
ture of this optimal control problem, because of the fact that N < n, is of the same character as
that of the minimum force control problem when N < n. Brockett [1982] considered the con-
trol problem posed by (17) and (14) (see also Baillieul [1978]), and subsequently the analysis
has been taken up in the mathematics and control theory literature as sub-Riemannian geom-
etry. In the next section we reinterpret the class of systems (17) as kinematic nonholonomic
control systems (Remark 3).
3. Nonholonomic control systems. We now consider the formulation of controlled non-
holonomic control systems. Nonholonomic systems may be divided into two classes (see,
e.g., Vershik and Gershkovich [1988])-variational nonholonomic systems (dubbed vako-
nomic systems by Arnold [1988]) and classical (or mechanical) nonholonomic systems. In
either case the basic ingredients are "nonintegrable" constraints on phase space, defined by
m, m < n, one-forms on M , w 1 , . . • , Wm, such that their span, over the smooth functions on
M, contains no nontrivial exact forms and in particular none of the forms W¡ , . . . , Wm is exact.
These forms then define a smooth distribution H on M . For each p E M , Hv is the subspace
of TvM defined by
We also stipulate that the distribution H is nonsingular, so that the dimension of the subspace
Hv does not vary with p, although the significance of this is not fully understood.
Variational nonholonomic systems are obtained in our context as solutions of variational
problems of the form
subject to
If we consider the kinematic system (17), in which the vector fields xi are orthonormal
with respect to the mechanical metric
1 ::::; i, j ::::; m,
and we define n - m independent dual one-forms wk satisfying
132 A. M. BLOCH AND P. E. CROUCH
it follows that the variational problem described above coincides with the optimal control
problem posed by the cost functional (14), and subject to the kinematics (17). In this context
the reader should be aware of work by Kirshnaprasad and Yang [1991] and many others
including Brockett [1982], Baillieul [1978], Murray and Sastry [1990], and Lafferiere and
Sussmann [1991]. See also Remark 3 below.
Classical nonholonomic systems are not obtained from a variational principie in the usual
sense (see Remark 2 below), but from D'Alembert's principie. The equations may be written
in the form
(18)
for any vector field X on M . The multipliers Ai are also uniquely determined from the
equations (18) and (19) as shown in Remark 2 below.
After our formulation of a holonomic controlled mechanical system (5), we define a
nonholonomic controlled mechanical system to have the form
D2 m N
(22)
1 :=::; k :=::; m.
1 :=::; k :=::; m,
NONHOLONOMIC CONTROL SYSTEMS ON RIEMANNIAN MANIFOLDS 133
Thus the force field Fin a nonholonomic system (22) is not arbitrary but satisfies the constraints
(24). ltfollows that regardless of N,or the vector fields Xi in the nonholonomic control system
(20), after solving for the multipliers >.i , the resulting system has at most n - m independent
control directions (as specified by vector fields that multiply the control functions ui).
Remark 2. We note that although the system (20) and (21) may not arise from a variational
principle in the usual Lagrangian sense, it is a solution of an instantaneous variational problem
(Vershik [1984])
i
8t2
11 - t
i=l
ui Xi 11 subject to \ Wk, ) = O, 1 :::; k :::; m.
This is made clear by differentiating the constraints, as in (23), yielding m affine constraints
in D 2 q/ 8t 2 . The multipliers >.k in (20) are the solutions of the following system of equations
(26)
where Vi are also control functions and xi are vector fields spanning the distribution H. In
general xi differ from xi appearing in the dynamic equations (20), the particular choice govemed
by physical consideration of the control mechanism. However, the general principle goveming
the choice of X i and Vi is that Vi control all Of the independent COmponents of the
velocity 'f¡f . For controllability analysis of the system (26), the assumption is made that the
distribution H is completely nonholonomic (see Vershik and Gershkovich [1988]), that is, the
involutive closure HL of H satisfies
p E M.
This is also an important observation in Brockett [1982]. ltis shown in Bloch et al. [1992b] that
under certain circumstances, controllability of the associated dynamic nonholonomic control
system (20), (21) follows from the assumption above.
Remark 4. Computing the derivative of the kinetic energy along solutions of the non-
holonomic control system (20), (21) yields
Thus
d
dt Tq( t) = Fq( t ) ( dq)
dt '
134 A. M. BLOCH AND P. E. CROUCH
where F is defined by (4). Because the same equation holds for the holonomic system
(equation (3) with U = O), it is clear that the constraints do no work, as is well known (see,
e.g., Neimark and Fufaev [1972]).
Example 1. We consider a penny rolling on the x-y plane with rotation angle (} and
heading angle c/J, as in Bloch and McClamroch [1989] and Bloch et al. [1992b]. We have two
controls, one that rolls the penny about its center of mass and another that tums it about its
vertical axis. The constraints are given by
x = iJ cos c/J ,
y = iJ sin c/J.
Letting q = (x, y, (}, cjJ ) T, the dynamic equations of motion (20) may be written as
<j = V ¡X ¡ + V2X 2 ,
where .X-1 = (cos c/J , sin c/J, 1, O)T and X-2 = (O, O, O, l )T lie in the distribution H.
Example 2. We now consider the "Heisenberg" system (so called because its vector fields
generate the Heisenberg algebra)-see Brockett [1982] and the work of Vershik et al. (see
e.g., Vershik and Gershkovich [1988]). Here we have a system on R3 in the variables (x, y , z)
and subject to the constraint
z = yx - xiJ.
We have controls in the x and y directions and wish to control the system in R3 . Letting
q = ( x , y, z )r , the natural dynamic nonholonomic control system (20) may be written as
<j = v1 X 1 + v2X 2 ,
- _ T - _ T
where X1- (1, O, y ) ,X2- (O, 1, -x) .
NONHOLONOMIC CONTROL SYSTEMS ON RIEMANNIAN MANIFOLDS 135
(V'y Z, Y) = O
for ali vector fields Y. Further, a sufficient condition for (Z (q), i)
to be a constant of motion for
a geodesic ftow is that Z is a Killing vector field. For controlled nonholonomic systems we
have the following restatement of Theorem 6, p. 82 in Arnold [1988].
LEMMA l. Sufficient conditions for (Z, ¡) to be a constant of motionfor the controlled
nonholonomic system (20), (21) are
(i) Z E H,
(ii) Z E Span {X¡, ..., X N }..L ( with respect to (-,·) ),
(iii) Z is a Killing vectorfield.
Proof. As in equation (23) we have
.!!_ j z dq ) _ j D Z dq ) j z D1 q )
dt \ ' dt - \ at ' dt +\ ' at 2 •
The first term is zero by (iii), and the second is zero by the expression for D 2 q/ ot2 in (20)
and (i) and (ii). D
Note that when M = Rn and the metric ( ·, ·) is independent of the coordinate function
Xi , then a/ OX i is a Killing vector field.
Note also that if the uncontrolled nonholonomic system (18) is determined by constraints
(19) in which the vector fields Wi are indeed Killing vector fields, then equation (23) gives
(Wk , D 2 q/ 8t2 ) = O. lt then follows from (18) that >..k = O, 1 ::; k ::; m, so that the ftow
of such a nonholonomic system is a restriction of the geodesic ftow. This is illustrated in
Example 2.
lt is often convenient to introduce a bundle structure in M, 7í : M --+ B, with fiber F,
dim. B = r, and dim. F = n - r. This structure reftects the natural geometric structure
of the system induced by the constraints and must be compatible with the constraints in the
sense that
(27) \ Zi , ;) = ci = const.
are constants of the motion for (20), (21). We define a distribution ÍI0 on M by setting X E Ílo
if
(Wk , X ) = 0 1 S k S m,
(29)
( Zk , X ) = Ck 1 ks s s.
Itfollows that any vector field Y on M can be decornposed uniquely into cornponents
In general q_H and q_ F are not derivatives of functions qH and qF on M, although in rnany
applications one can indeed identify such functions. Frorn equations (21) and (29) we obtain
Note that f F is affine in q_H . Frorn the original controlled nonholonornic systern (20) we rnay
deduce equations of the form
(31) ;t =
D' H
Í H ( q, q, u).
Equations (30) and (31) provide areduction of the 2n first-order equations (20), with constraints
(21), to n+ r first-order equations, without constraints. Locally we can write q_H = ft qB for
sorne trajectory qB (t) E B. In sorne cases we rnay be able to rewrite equations (30) and (31)
globally in terrns of a trajectory q( t) = ( qB( t), qF ( t) ), qB E B, qF E F.
The class of Caplygin control systerns introduced in Bloch et al. [1992b] corresponds to
Assurnption 1, in the case where M is a product M = B x F, B = Rr , F = Rn-r , (·, ·) is
the Euclidean rnetric, and the whole dynarnics (30), (31) is invariant with respect to qF . A
global prescription for these systerns is given in the forrn
where f F is affine in q_ B.
NONHOLONOMIC CONTROL SYSTEMS ON RIEMANNIAN MANIFOLDS 137
Another class of systems in which a global reduction is possible is the class of controlled
nonabelian Caplygin systems. The uncontrolled systems of this type were discussed in Koiller
[1988]. We define a controlled nonabelian Caplygin system as a system (30), (31), in which M
is a principal G bundle M ( G,B) for a Lie group G, G acts by isometries on the mechanical
metric, H (in either Assumption 1 or 2) is invariant under G in the sense that
H 9 .p = g* HP , g E G, p E M ,
g* Xi( P ) = Xi( g . p ), g E G, p E M , 1 i N.
The reduced system of equations can be written (locally) (by analogy with the work of Koiller)
in the form
g E G,
(32) Di¡_B .
at=
fB( qB , qB , u), qB E B M / G,
x; = R .X ,
9
then
Ad h = </>h. l g=e·
138 A. M. BLOCH AND P. E. CROUCH
Thus,
(33)
For compact semi-simple Lie groups, Ad g is an isometry of Q, for every g E G, with respect
to the Killing form. Thus we have
then
d
dt Ad h(t )Y l t=O = [Z, Y] = ad Z(Y ).
Note that the definition of Lie bracket on g coincides here, under the identification of g with
TeG, with the standard definition of the Lie bracket
[W, V] 9 (f ) = W9 (V (f ) ) - V9 (W (f ) )
for all functions f and G and vector fields W and V on G.
Applying (34) to Ad h(t ) and differentiating, yields
(x;,Y
;) = ( (X, J(Y))), X ,Y E Q.
for all vector fields V and W on M . Corresponding to the right invariant metric (-, ·) there
exists a unique Riemannian connection \!.Explicit formulas for \! are given in Arnold [1978]
or Nomizu [1954]. Specifically, \! defines a bilinear form on g
For the purposes of our later analysis we prove the following result.
LEMMA 2.
X r ( fk )( g ) =:
/ k ( R9( h( t) ) ) \t=o'
5.2. Background on the generalized rigid body. In this section we briefl.y review ma-
terial on the motion of a generalized rigid body. This is modeled by geodesic equations on
a compact semi-simple Lie group G, with the right invariant metric defined by a positive
definite mapping J on g as described in the previous section. We have two representations
of the velocity defined by a right invariant frame and left invariant frame. If X 1, . . . , X L is a
basis for g , we set
L
dg l
(39) dt = L.,, wi(t )Xi ( g ),
i=I
L
(40) := L Vi(t)X [ ( g ).
i=I
T (g ) = \dg dg ) =
2 dt ' dt 2
((
L.,,
i=I
v·X
i " L.,,
i=I
v ·Jx-)).
i i
140 A. M.
A. M. BLOCH
BLOCH ANO
AND P. E.CROUCH
P.E. CROUCH
Thus the coordinates Vi refer to the velocity with respect to a frame moving in the body,
whereas the coordinates Wi refer to the velocity with respect to a frame fixed in inertial space.
(See also Arnold [1978, p. 323], but there the roles of left and right are reversed.) We have
L L
L viX [ (g ) = L wi X J (g ).
i=I i=I
Hence, if we now assume that the basis Xi of g is orthonormal
we obtain
L
Wk = L((Xi, Ad g- 1 Xk ))vi.
i=I
We now obtain two representations for the acceleration, based on the left and right invariant
settings. First we obtain the representation based on (40). Differentiating, we obtain
L L
D2
a/.
= L Vi(t)X [( g ) + L vi( t)v j( t)('V x;X [)( g ).
i=I j,i= I
Setting vt = ¿7= 1 vi( t)Xi and 8vt / 8t = ¿7= 1 vi(t)Xi as vectors in 9, using the expression
(36) gives the usual expression (see, for example, Hermano [1977])
(41)
Now turning to the representation (39) we have the following result in which we write
L L
(42) Wt = L wi( t)Xi , ªZt = L wi( t)Xi , J( g ) = Ad g J Ad g- 1 •
i=I i=I
LEMMA 3.
(43) D2g
at2 = ( 8t
8Wt + J( g ) -1 [Wt.J( g )Wt ]) tg .
D2
at;,
= LL wi( t)X J ( g ) + LL wi( t)w j( t)('V x; X J )( g ).
i=I ij= I
But 'Vx1 X f = 'V(Ad g-1x;ir Xf. so using (36) and Lemma (2)
J
1
'Vx; X J = 2{-[Ad g- 1 X j , Ad g- 1 Xi] + J-1 [Ad g- 1 X j , J Ad g- 1 Xi]
+ J-1 [Ad g- 1Xi , J Ad g- 1 Xj]} .
141 A. M. BLOCH AND P. E. CROUCH
Now using the definition of J(g) we obtain the stated result in equation (43). O
5.3. The generalized rolling hall. In this section we describe a generalization of a ball
rolling on a ftat table, in which we model the ball as a generalized rigid body described in the
preceding section. We take the configuration space to be given as M = G x RN, where G is
an L-dimensional compact semi-simple Lie group as in the previous sections with L > N.
We put two Riemannian structures on M ; the ambient structure is defined by setting
where (·, ·)E is the Euclidean structure on RN, and (X 9 , V), (Y9 , W) are vectors in T9 G x
TxRN. The mechanical metric is defined by setting
where ( ·, ·) is the right invariant mechanical metric on G,defined in terms of a positive definite
mapping J of Q. The mechanical metric on M determines a Riemannian connection on M,
but the product structure defined by the metric enables us to rely on the connections defined
by the metrics (·, ·) and (-, ·)E on G and RN, respectively.
We wish to define a nonholonomic control system on M according to the prescription
given by the equations (20) and (21). For this example we suppose that m = N, that is, the
number of independent control forces is equal to the number of kinematic constraints. The
nonholonomic control system is completely described by defining the input forces
(44)
F
= LN Ui d xi = LN Ui \ 8ax. , .)E
i=I i=I i
dg l
) \ dx a
)E \ dg -1 l ) \ dx a
)E
\ -1
This follows from setting the velocity of the point of contact of the ball equal to zero and using
the fact that a three-dimensional rotation has a unique axis of rotation. Because this is not true
NONHOLONOMIC CONTROL SYSTEMS ON RIEMANNIAN MANIFOLDS 141
in higher
Using thedimensions, the constraint
skew symmetry of the Lie(45) appears
bracket to be the natural generalization. The system
we get
(20) and (21) now becomes
(46)
. / dg -1 l ( ))
(47) Xk = \ dt ' Jg Xk g ' 1 :::; k :::; N.
THEOREM 2. The following are constants of motion for the controlled nonholonomic
system (46), (47).
It follows directly from this result that along the trajectories of (46), (47) we have for
suitable constants Ck ,
We may calculate the multipliers >.k in (46) as was done in Remark 2. Differentiating (47)
and substituting for the second derivatives given in (46), we obtain
143 A. M. BLOCH AND P. E. CROUCH
The assumed positive definiteness of J ensures that A is always invertible, so that (50) defines
the multipliers Ak uniquely in equations (46).
We now describe the two reductions of equations (46), (47), described in §4. We first
consider the reduction based on Assumption l. In this case we take the base B = G and the
fiber F = RN. The N-independent constraints (47) ensure that dim. H = (N + L) - N =
L = dim. B, as required. The reduction procedure described in §4 simply rewrites the second-
order equations on the fiber by the constraint equations (47). We may employ equation (50)
to eliminate the multipliers >.k ; however, we first employ a simple feedback control, defined
by
which defines Üj , 1 :::::: j :::::: N uniquely. lt follows that under this reduction and feedback the
system (46), (47) becomes
N
D2g
8t2
L J; 1 X k ( g )ük .
k= I
d [21 (dg
dt ' dt
dg ) + 2(x, E]
1 . .x) = 6 N
'""X.kuk .
dt
Thus the work done by the force control ü in (53) simply changes the generalized rotational
energy and not the generalized translational energy.
Finally we may rewrite the reduced system (53) in terms of an "inertial frame" (39), as
described in the previous section. Using Lemma 3 to express the acceleration in terms of the
velocity Wt in equation (42), we obtain
1 :::::: k :::::: N,
(54)
NONHOLONOMIC CONTROL SYSTEMS ON RIEMANNIAN MANIFOLDS 143
or
We now describe the reduction based on Assumption 2 in §4. In this case we take the base
B = RN and the fiber F = G. We employ the L constraints (47) and (49) to determine the
velocity '1if . Note that we employ ali N velocity constraints and L - N constants of motion.
To calculate '1if explicitly, it is useful to employ the expression for it in terms of an "inertial
frame" as given in equations (39). (47) and (49) can then be written as
X k = Wk , 1 sk sN
L
!fjf by substituting Wk =
It follows that we may rewrite the expression (39) for the velocity
s s s s
X k , 1 k N , Wk = fk ( x , g), N + 1 k L. The remaining equations of the reduced
system are those for X k in system (46). However, we employ feedback again, expressed in
vector form as
j dg D -1 l) 1 sk sN.
\dt ' at J 9 xk '
The control uis defined uniquely by the invertibility of the matrix A defined by (51). The
resulting reduced equations have the form
1 sk sN,
N L
(57)
= L
xk X k (g ) + L fk (± , g )X k ( g ).
k= I k=N+I
Note that for this system we have
N
1( . ' )E '"' .
ddt 2 X ,X =L., X kUk ·
A
k= I
u
Thus in this formulation the force control in (57) simply changes the generalized translational
energy and not the generalized rotational energy.
145 NONHOLONOMICA.CONTROL
M. BLOCH SYSTEMS ON RIEMANNIAN MANIFOLDS
AND P. E. CROUCH 145
5.4. Controllability and optimal control. In this section we make sorne comments
about the problems associated with the controllability and optimal control of the reduced
models (54) and (57). We first comment on the controllability aspects. Clearly the system
(54) as written cannot be controllable, because of the constants of motion (49), which can be
reexpressed in the form
Reduction of the equations (54) by these constraints would be complicated to analyze, and
the reduction has already been performed in the system (57). A necessary condition for
controllability is accessibility (or weak controllability) (see Isidori [1989] or Nijmeijer and
Van der Schaft [1990]). However, as stated, the Lie algebra associated with the system (57)
is also complicated to analyze. We therefore content ourselves to the case where J = I , in
which case it is easily deduced that both systems (54) and (57) reduce to the system
Xk = Vk,
(58) vk = úk ,
d N L
d = L vk X k ( g ) + L ck X k ( g ), 1 :::; k :::; N.
k=l k=N+I
To analyze the accessibility and controllability of this system, we introduce sorne subspaces
of the Lie algebra g. Let P be the subspace spanned by the vectors X 1 , • • . , X N , let g p be the
subalgebra of g generated by P and let I p denote the ideal of g p generated by the subspace,
[P, P], of gp.
THEOREM 3. Assume that g is a simple Lie algebra. Then the reduced system (58) is
controllable and accessible if and only ij gp = g and [P, P] -=J O.
Proof We first analyze accessibility of the system. It is sufficient to analyze the Lie
algebra of the system (58) as represented by the Lie algebra L on the vector space R2N x g
with generators
a
9k = 8 ,
Vk
Bcause the elements of L depend only on v E RN, accessibility of (58) is equivalent to the
fact that Lv = R2 N x g , where Lv is the subspace of R2 N x g spanned by elements of L
evaluated at v. We construct a subalgebra L e L with the property that none of its elements
depend on v and such that L = L0• Accessibility is therefore equivalent to L = R2N x g.
(Because L does not depend on v , we may identify it with a subspace of R2N x g.) Explicitly
we set L to be the subalgebra of L generated by the vectors { gk , [gk , f ]; 1 :::; k :::; N}. All
Lie brackets of generators 9k and f not in L vanish at v = O. It is easily verified that
L, = span {Ba
vk ; Ba
xk + X k , 1 :::; k :::; N ; lp } .
Indeed, span {[[f, gj ], [f , gi]]; 1 :::; j , i :::; N} = [P, P]. Now L = R2N x g if and only
if lp = g, and so accessibility is equivalent to lp = g. Consider the situation in which
lp e:;; gp = g. Thus lp is an ideal of a simple Lie algebra g, so that lp = O or lp = g.
Because lp = O ifand only if [P, P] = 0, accessibility is equivalentto gp = g and [P, P] -=J O.
We now turn to controllability. We appeal to the result in Theorem 1. Specifically, we treat
system (58) as a system on a principal bundle G x R2 N, in which the system is G invariant and