See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/265935804

An introduction to optimal control

Article · January 2005

CITATIONS READS
48 4,110

2 authors:

Ugo Boscain Benedetto Piccoli

Institut des Materiaux Jean Rouxel Rutgers, The State University of New Jersey
184 PUBLICATIONS 5,114 CITATIONS 381 PUBLICATIONS 10,769 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by Benedetto Piccoli on 08 November 2017.

The user has requested enhancement of the downloaded file.

 An Introduction to Optimal Control
Ugo Boscain Benetto Piccoli

The aim of these notes is to give an introduction to the Theory of Optimal Control
for finite dimensional systems and in particular to the use of the Pontryagin Maximum
Principle towards the construction of an Optimal Synthesis. In Section 1, we introduce
the definition of Optimal Control problem and give a simple example. In Section 2 we
recall some basics of geometric control theory as vector fields, Lie bracket and con-
trollability. In Section 3, that is the core of these notes, we introduce Optimal Control
as a generalization of Calculus of Variations and we discuss why, if we try to write
the problem in Hamiltonian form, the dynamics makes the Legendre transformation
not well defined in general. Then we briefly introduce the problem of existence of
minimizers and state the Pontryagin Maximum Principle. As an application we con-
sider a classical problem of Calculus of Variations and show how to derive the Euler
Lagrange equations and the Weierstraß condition. Then we discuss the difficulties in
finding a complete solution to an Optimal Control Problem and how to attack it with
geometric methods. In Section 4 we give a brief introduction to the theory of Time
Optimal Synthesis on two dimensional manifolds developed in [14]. We end with a
bibliographical note and some exercises.

1 Introduction
Control Theory deals with systems that can be controlled, i.e. whose evolution can
be influenced by some external agent. Here we consider control systems that can be
defined as a system of differential equations depending on some parameters u ∈ U ⊆
Rm :

ẋ = f (x, u), (1)

where x belongs to some n–dimensional smooth manifold or, in particular, to R n . For

each initial point x0 there are many trajectories depending on the choice of the control
parameters u.
One usually distinguishes two different ways of choosing the control:

• open loop. Choose u as function of time t,

• closed loop or Feedback. Choose u as function of space variable x.

19
20 U. B OSCAIN AND B. P ICCOLI

The first problem one faces is the study of the set of points that can be reached,
from x0 , using open loop controls. This is also known as the controllability problem.
If controllability to a final point x1 is granted, one can try to reach x1 minimizing
some cost, thus defining an Optimal Control Problem:
Z T
min L(x(t), u(t)) dt, x(0) = x0 , x(T ) = x1 , (2)
0

where L : R ×U → R is the Lagrangian or running cost. To have a precise definition

n

of the Optimal Control Problem one should specify further: the time T fixed or free,
the set of admissible controls and admissible trajectories, etc. Moreover one can fix
an initial (and/or a final) set, instead than the point x0 (and x1 ).
Fixing the initial point x0 and letting the final condition x1 vary in some domain of
Rn , we get a family of Optimal Control Problems. Similarly we can fix x1 and let x0
vary. One main issue is to introduce a concept of solution for this family of problems
and we choose that of Optimal Synthesis. Roughly speaking, an Optimal Synthesis is
a collection of optimal trajectories starting from x0 , one for each final condition x1 .
As explained later, building an Optimal Synthesis is in general extremely difficult, but
geometric techniques provide a systematic method to attack the problem.
In Section 3.1 Optimal Control is presented as a generalization of Calculus of
Variations subjects to nonholonomic constraints.

Example Assume to have a point of unitary mass moving on a one dimensional line
and to control an external bounded force. We get the control system:
ẍ = u, x ∈ R, |u| ≤ C,
where x is the position of the point, u is the control and C is a given positive constant.
Setting x1 = x, x2 = ẋ and, for simplicity, C = 1, in the phase space the system is
written as: 
ẋ1 = x2
ẋ2 = u.
One simple problem is to drive the point to the origin with zero velocity in minimum
time. From an initial position (x̄1 , x̄2 ) it is quite easy to see that the optimal strategy is
to accelerate towards the origin with maximum force on some interval [0, t] and then to
decelerate with maximum force to reach the origin at velocity zero. The set of optimal
trajectories is depicted in Figure 1.A: this is the simplest example of Optimal Synthesis
for two dimensional systems. Notice that this set of trajectories can be obtained using
the following feedback, see Figure 1.B. Define the curves ζ ± = {(x1 , x2 ) : ∓x2 >
0, x1 = ± x22 } and let ζ be defined as the union ζ ± ∪ {0}. We define A+ to be the
region below ζ and A− the one above. Then the feedback is given by:
 +1 if (x1 , x2 ) ∈ A+ ∪ ζ +


u(x) = −1 if (x1 , x2 ) ∈ A− ∪ ζ −
0 if (x1 , x2 ) = (0, 0).

A N I NTRODUCTION TO O PTIMAL C ONTROL 21

x2
x2
−(
ζ u=−1
)
u(x)=−1

x1 x1
u(x)=+1

(u=+
1)
+
A B ζ

Figure 1: The simplest example of Optimal Synthesis and corresponding feedback.

Notice that the feedback u is discontinuous.

2 Basic Facts on Geometric control

This Section provides some basic facts about Geometric Control Theory. This is a
brief introduction that is far from being complete: we illustrate some of the main
available results of the theory, with few sketches of proofs. For a more detailed treat-
ment of the subject, we refer the reader to the monographs [3, 29].
Consider a control system of type (1), where x takes values on some manifold M
and u ∈ U . Along these notes, to have simplified statements and proofs, we assume
more regularity on M and U :
(H0) M is a closed n-dimensional submanifold of RN for some N ≥ n. The set
U is a measurable subset of Rm and f is continuous, smooth with respect to x
with Jacobian, with respect to x, continuous in both variables on every chart of
M.
A point of view, very useful in geometric control, is to think a control system as a
family of assigned vector fields on a manifold:
F = {Fu (·) = f (·, u)}u∈U .
We always consider smooth vector fields, on a smooth manifold M , i.e. smooth map-
pings F : x ∈ M 7→ F (x) ∈ Tx M , where Tx M is the tangent space to M at x. A
22 U. B OSCAIN AND B. P ICCOLI

vector field can be seen as an operator from the set of smooth functions on M to R. If
x = (x1 , ..., xn ) is a local system of coordinates, we have:
n
X ∂
F (x) = Fi .
i=1
∂xi

The first definition we need is of the concept of control and of trajectory of a control
system.
Definition 1 A control is a bounded measurable function u(·) : [a, b] → U . A trajec-
tory of (1) corresponding to u(·) is a map γ(·) : [a, b] → M , Lipschitz continuous on
every chart, such that (1) is satisfied for almost every t ∈ [a, b]. We write Dom(γ),
Supp(γ) to indicate respectively the domain and the support of γ(·). The initial point
of γ is denoted by In(γ) = γ(a), while its terminal point T erm(γ) = γ(b)
Then we need the notion of reachable set from a point x0 ∈ M .
Definition 2 We call reachable set within time T > 0 the following set:
Rx0 (T ) := {x ∈ M : there exists t ∈ [0, T ] and a trajectory
γ : [0, t] → M of (1) such that γ(0) = x0 , γ(t) = x}. (3)
Computing the reachable set of a control system of the type (1) is one of the main
issues of control theory. In particular the problem of proving that R x0 (∞) coincide
with the whole space is the so called controllability problem. The corresponding local
property is formulated as:
Definition 3 (Local Controllability) A control system is said to be locally control-
lable at x0 if for every T > 0 the set Rx0 (T ) is a neighborhood of x0 .
Various results were proved about controllability and local controllability. We only
recall some definitions and theorems used in the sequel.
Most of the information about controllability is contained in the structure of the
Lie algebra generated by the family of vector fields. We start giving the definition of
Lie bracket of two vector fields.
Definition 4 (Lie Bracket) Given two smooth vector fields X, Y on a smooth mani-
fold M , the Lie bracket is the vector field given by:
[X, Y ](f ) := X(Y (f )) − Y (X(f )).
In local coordinates:
X  ∂Yj ∂Xj

[X, Y ]j = Xi − Yi .
i
∂xi ∂xi
 
In matrix notation, defining ∇Y := ∂Yj /∂xi (j row, i column) and thinking to
(j,i)
a vector field as a column vector we have [X, Y ] = ∇Y · X − ∇X · Y .
A N I NTRODUCTION TO O PTIMAL C ONTROL 23

Definition 5 (Lie Algebra of F) Let F be a family of smooth vector fields on a

smooth manifold M and denote by χ(M ) the set of all C ∞ vector fields on M . The Lie
algebra Lie(F) generated by F is the smallest Lie subalgebra of χ(M ) containing
F. Moreover for every x ∈ M we define:

Liex(F) := {X(x) : X ∈ Lie(F)}. (4)

Remark 1 In general Lie(F) is a infinite-dimensional subspace of χ(M ). On the

other side since all X(x) ∈ Tx M (in formula (4)) we have that Liex (F) ⊆ Tx M and
hence Liex(F) is finite dimensional.

Remark 2 Lie(F) is built in the following way. Define: D1 = Span{F}, D2 =

Span{D1 + [D1 , D1 ]}, · · · Dk = Span{Dk−1 + [Dk−1 , Dk−1 ]}. D1 is the so called
distribution generated by F and we have Lie(F) = ∪k≥1 Dk . Notice that Dk−1 ⊆
Dk . Moreover if [Dn , Dn ] ⊆ Dn for some n, then Dk = Dn for every k ≥ n.

A very important class of families of vector fields are the so called Lie bracket
generating (or completely nonholonomic) systems for which:

LiexF = Tx M, ∀ x ∈ M. (5)

For instance analytic systems (i.e. with M and F analytic) are always Lie bracket
generating on a suitable immersed analytic submanifold of M (the so called orbit of
F). This is the well know Hermann-Nagano Theorem (see for instance [29], pp. 48).
If the system is symmetric, that is F = −F (i.e. f ∈ F ⇒ −f ∈ F), then the
controllability problem is more simple. For instance condition (5) with M connected
implies complete controllability i.e. for each x0 ∈ M , Rx0 (∞) = M (this is a
corollary of the well know Chow Theorem, see for instance [3]).
On the other side, if the system is not symmetric (as for the problem treated in
Section 4), the controllability problem is more complicated and controllability is not
guaranteed in general (by (5) or other simple conditions), neither locally. Anyway,
important properties of the reachable set for Lie bracket generating systems are given
by the following theorem (see [37] and [3]):
Theorem 1 (Krener) Let F be a family of smooth vector fields on a smooth man-
ifold M . If F is Lie bracket generating, then, for every T ∈]0, +∞], Rx0 (T ) ⊆
Clos(Int(Rx0 (T )). Here Clos(·) and Int(·) are taken with respect to the topology
of M .
Krener Theorem implies that the reachable set for Lie bracket generating systems has
the following properties:
• It has nonempty interior: Int(Rx0 (T )) 6= ∅, ∀ T ∈]0, +∞].
• Typically it is a manifold with or without boundary of full dimension. The
boundary may be not smooth, e.g. have corners or cuspidal points.
24 U. B OSCAIN AND B. P ICCOLI

Figure 2: A prohibited reachable set for a Lie bracket generating systems.

In particular it is prohibited that reachable sets are collections of sets of different

dimensions as in Figure 2. These phenomena happen for non Lie bracket generating
systems, and it is not know if reachable sets may fail to be stratified sets (for generic
smooth systems) see [28, 29, 54].
Local controllability can be detached by linearization as shown by the following
important result (see [40], p. 366):

Theorem 2 Consider the control system ẋ = f (x, u) where x belongs to a smooth

manifold M of dimension n and let u ∈ U where U is a subset of Rm for some m,
containing an open neighborhood of u0 ∈ Rm . Assume f of class C 1 with respect to
x and u. If the following holds:

f (x0 , u0 ) = 0,
rank[B, AB, A2 B, ..., An−1 B] = n, (6)
where A = ∂f /∂x)(x0 , u0 ) and B = ∂f /∂u)(x0 , u0 ),

then the system is locally controllable at x0 .

Remark 3 Condition (6) is the well know Kalman condition that is a necessary and
sufficient condition for (global) controllability of linear systems:

ẋ = Ax + Bu, x ∈ Rn , A ∈ Rn×n , B ∈ Rn×m , u ∈ Rm .

In the local controllable case we get this further property of reachable sets:

Lemma 1 Consider the control system ẋ = f (x, u) where x belongs to a smooth

manifold M of dimension n and let u ∈ U where U is a subset of R m for some m.
A N I NTRODUCTION TO O PTIMAL C ONTROL 25

Assume f of class C 1 with respect to x and continuous with respect to u. If the control
system is locally controllable at x0 then for every T , ε > 0 one has:

Rx0 (T ) ⊆ Int(Rx0 (T + ε)). (7)

Proof. Consider x ∈ Rx0 (T ) and let ux : [0, T ] → U be such that the corresponding
trajectory starting from x0 reaches x at time T . Moreover, let Φt be the flux associated
to the time varying vector field f (·, u(t)) and notice that Φt is a diffeomorphism. By
local controllability at x0 , Rx0 (ε) is a neighborhood of x0 . Thus ΦT (Rx0 (ε)) is a
neighborhood of x and, using ΦT (Rx0 (ε)) ⊂ Rx0 (T + ε), we conclude.

3 Optimal Control
In this section we give an introduction to the theory of Optimal Control. Optimal Con-
trol can be seen as a generalization of the Classical Calculus of Variations
(min 0 L(x(t), ẋ(t))) to systems with nonholonomic constrains of the kind ẋ =
RT
f (x, u), u ∈ U .

Remark 4 We recall that a constraint on the velocity is said to be nonholonomic if it

cannot be obtained as consequence of a (holonomic) constraint on the position of the
kind:
ψi (x) = 0, i = 1, ..., n0 , n0 < n.
where the real functions ψi (x) are sufficiently regular to define a submanifold M 0 ⊂
M . Clearly since holonomic constraints can be eliminated simply by restricting the
problem to M 0 , the interesting case is the nonholonomic one. In the following
when we speak about nonholonomic constraints we always refer to nonholonomic
constraints of the kind ẋ = f (x, u), u ∈ U .

The most important and powerful tool to look for an explicit solution to an Optimal
Control Problem is the well known Pontryagin Maximum Principle (in the following
PMP, see for instance [3, 29, 49]) that give a first order necessary condition for opti-
mality. PMP is very powerful for the following reasons:

• it generalizes the Euler Lagrange equations and the Weierstraß condition of Cal-
culus of Variations to variational problems with nonholonomic constrains;
• it provides a pseudo-Hamiltonian formulation of the variational problem in the
case in which the standard Legendre transformation is not well defined (as in
the case of Optimal Control, see below).

Roughly speaking PMP says the following. If a trajectory of a control system is a

minimizer, then it has a lift to the cotangent bundle, formed by vector-covector pairs,
such that:
26 U. B OSCAIN AND B. P ICCOLI

• it is a solution of an pseudo-Hamiltonian system,

• the pseudo Hamiltonian satisfies a suitable maximization condition.
Here we speak of a pseudo-Hamiltonian system since the Hamiltonian depends on the
control (see below). In the regular cases the control is computed as function of the
state and costate using the maximization condition.
It is worth to mention that giving a complete solution to an optimization prob-
lem (that for us means to give an optimal synthesis, see Definition 7) in general is
extremely difficult for several reasons:
A the maximization condition not always provide a unique control. Moreover PMP
gives a two point boundary value problem with some boundary conditions given
at initial time (state) and some given at final time (state and covector);
B one is faced with the problem of integrating a pseudo–Hamiltonian system (that
generically is not integrable except for very special dynamics and costs);
C a key role is played by some special classes of extremals called abnormal (ex-
tremals independent from the cost) and singular (extremals that are singularity
of the End-Point Mapping). See Section 3.2.3;
D even if one is able to find all the solutions of the PMP it remains the problem of
selecting among them the optimal trajectories.
Usually A, B and C are very complicated problems and D may be even more difficult.
For this reason (out from the so called linear quadratic problem, see for instance
[3, 29]) one can hope to find a complete solution of an Optimal Control Problem only
in low dimensions, unless the system presents a lot of symmetries. For instance most
of the problems in dimension 3 are still open also for initial and final conditions close
one to the other.
In Section 3.1 we give a brief introduction to the PMP in Optimal Control as a
generalization of the classical first order conditions in Calculus of Variations.
In Section 3.2 we briefly discuss the problem of existence of minimizers, state
the PMP, define abnormal and singular trajectories. In Section 3.3 we show how to
get, in the case of the Calculus of Variations, the Euler Lagrange equations and the
Weierstraß condition. In Section 3.4, as an application we show in some detail how
to compute geodesics for a famous singular Riemannian problem using the PMP. In
Section 3.5 the problem of constructing an Optimal Synthesis with geometric methods
is treated.

3.1 Introduction
In this Section we first recall the Euler Lagrange equations, how to transform them in
an Hamiltonian form and discuss in which case this transformation is applicable (the
Legendre transformation must be invertible).
A N I NTRODUCTION TO O PTIMAL C ONTROL 27

In Section 3.1.2 it is shown that for nonholonomic constraints (i.e. for an Opti-
mal Control problem with nontrivial dynamics) the Euler Lagrange equation cannot
be written in standard form and the Legendre transformation is never well defined.
Finally we explicate some connection with the Lagrangian formulation of mechanical
systems.
Here, for simplicity, the state space is assumed to be Rn and not an arbitrary
manifold.

3.1.1 The Legendre Transformation

Consider a standard problem in Calculus of Variations:
 RT
 minimize 0
L(x(t), ẋ(t))dt,
x(0) = x0 , (8)
x(T ) = xT ,


where x = (x1 , ..., xn ) ∈ Rn and L : R2n → R is a C 2 function. It is a standard fact

that if a C 2 minimizer x(.) exists, then it must satisfy the Euler Lagrange equations:
d ∂L ∂L
= , (9)
dt ∂ ẋi ∂xi
that for our purpose is better to write more precisely as:
 !
d ∂L(x, u) ∂L(x, u)
= . (10)
dt ∂ui (x(t),ẋ(t)) ∂xi (x(t),ẋ(t))
(11)
Euler Lagrange equations are second order ODEs very difficult to solve in general,
also numerically. In fact, the theory of differential equation is much more developed
for first order than for second order differential equations. For this reason it is of-
ten convenient to transform equations (10) into a system of ODEs of a special form
(Hamiltonian equations) via the so called Legendre transformation.
The problem of finding solutions to a system of ODEs is simplified if the system
admits constants of the motion (in involution). Roughly speaking this permits to de-
couple the problem to the corresponding level sets. The most important advantage of
passing from the Lagrangian to the Hamiltonian formulation is that, in Hamiltonian
form, it is easier to recognize constants of the motion.
The Legendre transformation consists in the following.
• We first reduce the system (10) of n second order ODEs to a system of 2n first
order ODEs introducing the variable u := ẋ:

 d ∂L(x,u) = ∂L(x,u)
dt ∂ui
(x(t),u(t))
∂xi
(x(t),u(t)) (12)
ẋ(t) = u(t).

28 U. B OSCAIN AND B. P ICCOLI

• Then we make the change of coordinates in Rn :

∂L(x, u)
(x, u) → (x, p), where pi = Φi (x, u) := .
∂ui
This change of coordinates is well defined if it realizes a C 1 -diffeomorphism of
R2n into R2n , i.e. we must have:

0
 2
11
  
∂ L(x, u)
det ∂Φ(x,u) ∂Φ(x,u) = det 6= 0, (13)
∂x ∂u . ∂ui ∂uj

for every (x, u) ∈ R2n . If the condition (13) is satisfied, the Legendre trans-
formation is said invertible. In this case the inverse transformation is u =
Φ−1 (x, p) and the Lagrangian is called regular.

• Define the function (called Hamiltonian):

H(x, p) := pΦ−1 (x, p) − L(x, Φ−1 (x, p)). (14)

In the (x, u) coordinates the Hamiltonian takes the form ∂L

∂u u − L(x, u) and
usually one remembers it in the “mixed coordinates” form pu − L.
After the Legendre transformation (if it is invertible), the Euler Lagrange equa-
tions are written as the Hamiltonian equations:
∂H
(
ẋ = ∂p
(15)
ṗ = − ∂H
∂x

In fact using carefully the chain rule we have for the first:

∂
pΦ−1 (x, p) − L(x, Φ−1 (x, p))


ẋ =
∂p
∂Φ−1 (x, p) ∂L(x, u) ∂Φ−1 (x, p)
= Φ−1 (x, p) + p − ,
∂p ∂u (x,Φ−1 (x,p)) ∂p

and using the fact that u = Φ−1 (x, p) and p = ∂L

∂u we get the second of (12).
Similarly for the second of (15) we have:

∂
pΦ−1 (x, p) − L(x, Φ−1 (x, p))


ṗ = −
∂x !
∂Φ−1 (x, p) ∂L ∂L(x, u) ∂Φ−1 (x, p)
= − p − − .
∂x ∂x ∂u (x,Φ−1 (x,p)) ∂x

Again using the fact that p = ∂L

∂u we get the first of (12) and hence (10).
A N I NTRODUCTION TO O PTIMAL C ONTROL 29

3.1.2 Optimal Control as a Generalization of Calculus of Variations

An Optimal Control Problem can be thought as a generalization of a problem of Cal-
culus of Variations (8) in the case in which:

• a nonholonomic constraint is added (i.e. a dynamic ẋ = f (x, u), u ∈ U ⊂ R m ).

• the Lagrangian L is a function of (x, u) instead than function of (x, ẋ). Clearly
this is a generalization, since from (x, u) we can always find (x, ẋ) using the
dynamics ẋ = f (x, u).

Usually one considers also more general costs and targets (see below), but this is out
of the purpose of this introduction. The consequences of these generalizations are the
following:

• In the case of nonholonomic constraints the first order necessary condition can
not be written in the form (10). One can still write a Lagrange formulation of
the problem using Lagrange multipliers, but this procedure works only under
suitable regularity conditions (see the recent book [7] and references therein).
• One could try to write the problem in Hamiltonian form, but the dynamics ẋ =
f (x, u), u ∈ U (unless it is equivalent to the trivial one ẋ = u, u ∈ Rn ) renders
the Legendre transformation not well defined. This is simply consequence of the
fact that the ẋi ’s may be not independent (or equivalently f (x, U ) is a proper
subset of Rn ) and so operating with the Legendre transformation we may get
not enough pi ’s. In other words the Legendre transformation would be:

∂L(x, u)
(x, u) → (x, p), where pi = Φi (x, v) := , (16)
∂ u̇i
∂ 2 L(x,u)
but ∂ui ∂uj is not a n × n matrix (if U ⊂ Rm it is m × m) and so the Legendre
transformation can not be inverted.

There is a way of getting a well defined Legendre transformation also in the case
of Optimal Control, starting from the Lagrange equations written with Lagrange mul-
tipliers. But again this procedure works only under suitable regularity conditions, see
[7].
The Pontryagin Maximum Principle (that will be stated precisely in the next sec-
tion) permits to write the minimization problem directly in a pseudo-Hamiltonian form
(pseudo because the Hamiltonian still depends on the controls that must be determined
with a suitable maximization condition, see Definition 8, p. 33).

Remark 5 (nonholonomic mechanics) At this point it is worth to mention the fact

that Optimal Control Theory can not be used as a variational formulation of nonholo-
nomic mechanics. More precisely, any conservative mechanical problem, subject only
30 U. B OSCAIN AND B. P ICCOLI

Legendre Transf. connections

Lagrange Formulation Hamilt. Formulation with
invertible mechanics

2
Calulus of Variation In standard L(x,v) in standard form, systems
if =0 but possible only if with
form u u holonomic
geneneralization
the Legendre transf
is invertible constraints

the Legendre transf. in a pseudo Hamilt.

NOT form.
is NOT well defined NO
in standard but ALWAYS
Optimal Control
form possible

Figure 3:

to holonomic constraints, can be formulated as a variational problem in the following

sense. Writing L = T − V , where T is the kinetic energy and V the potential energy,
the equations the motion are the Euler Lagrange equations corresponding to L. On
the other side, if we have a mechanical system subjects to nonholonomic constraints
(e.g. a body rolling without slipping) then the equations of motion are not given by
the solution of the corresponding Optimal Control Problem. To get a variational for-
mulation of a nonholonomic mechanical problem, one should impose the constraints
after the variation, see [7].

In the Figure 3 we collect the ideas presented in this section.

3.2 The Theory of Optimal Control

An Optimal Control Problem in Bolza form for the system (1), p. 19, is a problem of
the following type:
 RT
 minimize 0
L(x(t), u(t))dt + ψ(x(T )),
x(0) = x0 , (17)
x(T ) ∈ TT ,


where L : M × U → R is the Lagrangian or running cost, ψ : M → R is the final

cost, x0 ∈ M is the initial condition and TT ⊂ M the target. The minimization is
taken on the set of all admissible trajectories of the control system (1) (the admissi-
bility conditions to be specified), that start at the initial condition and end at the target
A N I NTRODUCTION TO O PTIMAL C ONTROL 31

in finite time T that can be fixed or free, depending on the problem. Notice that this
Optimal Control Problem is autonomous, hence we can always assume trajectories to
be defined on some interval of the form [0, T ].

Of great importance for applications are the so called control affine systems:
m
(18)
X
ẋ = F0 + ui F i , ui ∈ R,
i=1

with quadratic cost:

Z m
T X
minimize u2i dt, (19)
0 i=1

or with cost equal to 1 (that is the minimum time problem) and bounded controls
|ui | ≤ 1. The term F0 in (18) is called drift term and, for the so called distributional
systems (or driftless), is equal to zero. Subriemannian problems are distributional
systems with quadratic cost. See also Exercise 3, p. 60. Single–input systems are
control affine systems with only one control, i.e. m = 1.

Definition 6 We call Px1 , x1 ∈ M , the Optimal Control Problem given by the dy-
namics (1) and the minimization (17) with TT = x1 .

Beside (H0), (see p. 21) we make the following basic assumptions on (1) and (17).

(H1) L is continuous, smooth with respect to x with Jacobian, with respect to x,

continuous in both variables on every chart of M ;

(H2) ψ is a C 1 function.

We are interested in solving the family of control problems {Px1 }x1 ∈M and for us a
solution is given by an Optimal Synthesis that is

Definition 7 (Optimal Synthesis) Given Ω ⊂ M , an optimal synthesis on Ω for the

family of Optimal Control Problems {Px1 }x1 ∈Ω is a collection {(γx1 , ux1 ) : x1 ∈ Ω}
of trajectory–control pairs such that (γx1 , ux1 ) provides a solution to Px1 .

Remark 6 One can invert the role of x0 and x1 letting x1 be fixed and x0 vary. More-
over, we can consider a generalization fixing an initial manifold SS, called source. In
this case the initial condition reads x0 ∈ SS.

Remark 7 In many cases it happens that an Optimal Synthesis, defined on R(T ), is

generated by a piecewise smooth feedback u(x), that is a map from R(T ) to U .
32 U. B OSCAIN AND B. P ICCOLI

3.2.1 Existence
In order to prove existence for Px1 , we first give the next:

Theorem 3 Assume f bounded, U compact and let the space of admissible controls
be the set of all measurable maps u(·) : [a, b] → U . If the set of velocities V (x) =
{f (x, u)}u∈U is convex, then the reachable set R(T ) is compact.

Remark 8 The fact that R(T ) is relatively compact is an immediate consequence of

the boundedness and continuity of f . One can also replace the boundedness with a
linear grows condition. The convexity of V (x) is the key property to guarantee that
R(T ) is closed. For a proof see [23].

From this theorem, we get immediately the following:

Theorem 4 Assume f bounded, U compact, and let the space of admissible controls
be the set of all measurable maps u(·) : [a, b] → U . Moreover assume that the set
{f (x, u), L(x, u)}u∈U is convex. If L ≥ C > 0 and x1 ∈ Rx0 (∞), then the problem
Px1 has a solution.

Sketch of the Proof. There exists T > 0 such that every trajectory ending at x 1 after
T is not optimal. Consider the augmented system obtained adding the extra variable
y such that:
ẏ = L(x, u).
Since f is bounded, all trajectories defined on [0, T ] have bounded costs. Thus apply-
ing the previous Theorem to the augmented system, we get that R (x0 ,0) (T ), the reach-
able set for the augmented system of (x, y), is compact. Therefore there exists a min-
imum for the function (x, y) → y + ψ(x) on the set R(x0 ,0) (T ) ∩ {(x, y) : x = x1 }.
The trajectory, reaching such a minimum, is optimal. 

3.2.2 Pontryagin Maximum Principle

The standard tool to determine optimal trajectories is the well known Pontryagin Max-
imum Principle, see [3, 29, 49], that gives a first order condition for optimality.
Pontryagin Maximum Principle can be stated in several forms depending on the fol-
lowing:

i) the final time is fixed or free (for fixed final time see Exercise 4, p. 61),

ii) dimension and regularity of the source and of the target,

iii) the cost contains only the running part, only the final part, or both,

iv) the source and/or the target depend on time.

A N I NTRODUCTION TO O PTIMAL C ONTROL 33

Here we state a version in which: i) the final time is free, ii) the source is zero dimen-
sional and the target TT is a smooth submanifold of M of any dimension, iii) there are
both running and final cost, iv) the source and the target do not depend on time. Let
us introduce some notation.
For every (x, p, λ0 , u) ∈ T ∗ M × R × U we define:

H(x, p, λ0 , u) = hp, f (x, u)i + λ0 L(x, u), (20)

and
H(x, p, λ0 ) = max{H(x, p, λ0 , u) : u ∈ U }.

Definition 8 (Extremal Trajectories) Consider the Optimal Control Problem (1),

(17) and assume (H0), (H1) and (H2). Let u : [0, T ] → U be a control and γ a cor-
responding trajectory. We say that γ is an extremal trajectory if there exist a Lipschitz
continuous map called covector λ : t ∈ [0, T ] 7→ λ(t) ∈ Tγ(t) ∗
M and a constant
λ0 ≤ 0, with (λ(t), λ0 ) 6= (0, 0) (for all t ∈ [0, T ]), that satisfy:
(PMP1) for a.e. t ∈ [0, T ], in a local system of coordinates, we have
∂H
λ̇ = − (γ(t), λ(t), λ0 , u(t));
∂x

(PMP2) for a.e. t ∈ [0, T ], we have H(γ(t), λ(t), λ0 ) = H(γ(t), λ(t), λ0 , u(t)) = 0;
(PMP3) for every v ∈ Tγ(T ) TT , we have hλ(T ), vi = λ0 h∇ψ(γ(T )), vi (transversal-
ity condition).
In this case we say that (γ, λ) is an extremal pair.

Pontryagin Maximum Principle (briefly PMP) states the following:

Theorem 5 (Pontryagin Maximum Principle) Consider the Optimal Control Prob-
lem (1), (17) and assume (H0), (H1) and (H2). If u(·) is a control and γ a correspond-
ing trajectory that is optimal, then γ is extremal.

Remark 9 Notice that the dynamic ẋ = f (x, u) and equation (PMP1) can be written
in the pseudo-Hamiltonian form:
∂H
ẋ(t) = (x(t), p(t), λ0 , u(t)),
∂p
∂H
ṗ(t) = − (x(t), p(t), λ0 , u(t)).
∂x
Remark 10 Notice that the couple (λ, λ0 ) is defined up to a positive multiplicative fac-
tor, in the sense that if the triple (γ, λ, λ0 ) represents an extremal, than the same hap-
pens for the triple (γ, αλ, αλ0 ), α > 0. If λ0 6= 0, usually one normalizes (λ(·), λ0 )
by λ0 = −1/2 or λ0 = −1.
34 U. B OSCAIN AND B. P ICCOLI

Remark 11 If the target TT is a point then the transversality condition (PMP3) is

empty.

Remark 12 In the case in which we have also a source: γ(0) ∈ SS, we have also a
transversality condition at the source:

(PMP3’) For every v ∈ Tγ(0) SS, hλ(0), vi = 0.

Remark 13 The maximized Hamiltonian H(x, p) can be written as

H(x, p) = H(x, p, u(x, p)),

where u(x, p) is one of the values of the control realizing the maximum. Now if the
set of controls U is an open subset of Rm and H is smooth with respect to u then
u(x, p) is one of the solutions of:

∂H(x, p, u)
= 0. (21)
∂u
A useful fact is the following. If moreover (21) permits to find in a unique way u
as a smooth function of (x, p), then we can write the Hamiltonian equations with H
instead of H getting a true Hamiltonian system with no dependence on the control. In
fact in this case:
∂H(x, p) ∂H(x, p, u(x, p)) ∂H(x, p, u)
= =
∂x ∂x ∂x (x,p,u(x,p))
∂H(x, p, u) ∂u(x, p) ∂H(x, p, u)
+ =
∂u (x,p,u(x,p)) ∂x ∂x (x,p,u(x,p))
∂H(x, p) ∂H(x, p, u(x, p)) ∂H(x, p, u)
= =
∂p ∂p ∂p (x,p,u(x,p))
∂H(x, p, u) ∂u(x, p) ∂H(x, p, u)
+ =
∂u (x,p,u(x,p)) ∂p ∂p (x,p,u(x,p))

where the last equality follows from (21). Then one gets the controls plugging the
solution of this Hamiltonian system in the expression u(x, p). This nice P situation
happens always for normal extremals for distributional systems (ẋ = i ui Fi (x))
with quadratic cost (min 0 i ui dt). See Section 3.4, p. 38 for an example.
RT P 2

3.2.3 Abnormal Extremals and Endpoint Singular Extremals

Assume now that f is differentiable also with respect to u. The following sets of
trajectories have very special features:
A N I NTRODUCTION TO O PTIMAL C ONTROL 35

Definition 9 (Endpoint Singular Trajectories and Endpoint Singular Extremals)

We call endpoint singular trajectories, solutions to the following equations:

∂ H̄
ẋ(t) = (x(t), p(t), u(t)),
∂p
∂ H̄
ṗ(t) = − (x(t), p(t), u(t)),
∂x
∂ H̄
(x(t), p(t), u(t)) = 0,
∂u
where H̄(x, p, u) := hp, f (x, u)i, p(t) 6= 0 and the constraint u ∈ U ∈ Rm is
changed in u ∈ Int(U ) 6= ∅. Endpoint singular trajectories that are also extremals
are called endpoint singular extremals.

Remark 14 Notice that, although endpoint singular trajectories do not depend on the
cost and on the constraints on the control set, endpoint singular extremals (that for our
minimization problem are the interesting ones) do depend.
The name endpoint singular trajectories comes from the fact that they are singu-
larities of the endpoint mapping that, fixed an initial point and a time T , associates to
a control function u(·), defined on [0, T ], the end point of the corresponding trajectory
γ:
E x0 ,T : u(·) 7→ γ(T ).
By singularity of the end point mapping, we mean a control at which the Fréchet
derivative of E x0 ,T is not surjective. For more details see [11]. Roughly speaking,
this means that the reachable set, locally, around the trajectory, does not contain a
neighborhood of the end point γ(T ).
In the case of a minimum time problem for a control affine system (18) with
|ui | ≤ 1, endpoint singular trajectories satisfy < p(t), Fi (γ(t)) >= 0. Endpoint sin-
gular extremals, are just endpoint singular trajectories for which there exists λ 0 ≤ 0
satisfying < p(t), F0 > +λ0 = 0 and corresponding to admissible controls.
In Section 4, we consider control systems of the form ẋ = F (x) + uG(x) where
|u| ≤ 1. In this case, under generic conditions, endpoint singular extremals are arcs
of extremal trajectories corresponding to controls not constantly equal to +1 or −1.

Definition 10 (Abnormal Extremals) We call abnormal extremals extremal trajecto-

ries for which λ0 = 0.

Remark 15 Abnormal extremals do not depend on the cost, but depend on the con-
straint u ∈ U .

Remark 16 In some problems, like in subriemannian geometry or in distributional

problems for the minimum time with bounded controls, endpoint singular extremals
and abnormal extremals coincide, see Exercise 7, p. 61.
36 U. B OSCAIN AND B. P ICCOLI

In other problems (e.g. minimum time for control affine systems) the two defi-
nitions are different, but coincide for some very special class of trajectories. These
trajectories are usually called singular exceptional (see [13]).
Remark 17 (High Order Conditions) PMP is used in Synthesis Theory to attempt a
finite dimensional reduction of the minimization problem, as explained in Step 2 of
Section 3.5, p. 41. Of course high order conditions can be very useful for further
restriction of candidate optimal trajectories.
There are several possible higher order variations. For instance the high order
principle of Krener [38] and the envelope theory developed by Sussmann [52, 53].
Other high order conditions are based on different techniques: symplectic geometric
methods, conjugate points, Generalized Index Theory. For a list of references see [14].
High order conditions are out of the purpose of this notes. See for instance [14]
and references therein.

3.3 Calculus of Variations

The classical case of Calculus of Variations can be formulated as an Optimal Control
Problem for the dynamics ẋ = u, u ∈ Tx M , fixed final time T , ψ = 0 and TT = x1 ∈
M:

 ẋ = u
 RT
min 0 L(x(t), u(t)) dt

. (22)

 x(0) = x0 ,
x(T ) = xT .


Here, together with (H0) and (H1), we assume also L(x, u) ∈ C 1 and we are looking
for Lipschitz minimizers. For this problem there are not abnormal extremals and end-
point singular trajectories. Moreover one can deduce from PMP the Euler Lagrange
equations and the Weierstraß condition.
Indeed from (22) the Hamiltonian reads:
H(x, p, λ0 , u) = hp, ui + λ0 L(x, u). (23)
In this case H̄(x, p, u) = hp, ui thus ∂ H̄/∂u implies p ≡ 0. Hence there are not
endpoint singular trajectories (cfr. Definition 9, p. 35). Moreover (PMP2) with λ 0 = 0
implies (cfr. Remark 10) p = 0 contradicting the fact that (p, λ0 ) 6= 0. Therefore
there are not abnormal extremals and it is convenient to normalize λ 0 = −1 (in the
following we drop the dependence from λ0 ):
H(x, p, u) = hp, ui − L(x, u). (24)
Since there are is no constraint on the control set (u ∈ Rn ) and L is differentiable with
respect to u, the maximization condition (PMP2) implies a.e.:
∂H ∂L
(x(t), p(t), u(t)) = 0 ⇒ p(t) = (x(t), u(t)), (25)
∂u ∂u
A N I NTRODUCTION TO O PTIMAL C ONTROL 37

while the Hamiltonian equations are:

ẋ(t) = ∂H

∂p (x(t), p(t), u(t)) = u(t), a.e., (26)
ṗ(t) = − ∂H ∂L
∂x (x(t), p(t), u(t)) = + ∂x (x(t), u(t)), a.e.

The first equation is the dynamics, while the second together with (25) give the Euler
Lagrange equations in the case of a Lipschitz minimizer. (We recall that in general
(∂L/∂u)(x(t), u(t)) is not absolutely continuous, but it coincides a.e. with an abso-
lutely continuous function, whose derivative is a.e. equal to (∂L/∂x)(x(t), u(t))). If
x(.) ∈ C 2 (that implies u(.) ∈ C 1 ) and ∂L/∂u is C 1 , then p(t) is C 1 and one gets the
Euler Lagrange equations in standard form:
d ∂L ∂L
(x(t), u(t)) = (x(t), u(t)),
dt ∂u ∂x
but it is well known that they hold everywhere also for x(.) ∈ C 1 and for a less regular
L (cfr. for instance [22, 26]).
Remark 18 Notice that, in the case of Calculus of Variations, the condition to get a
true Hamiltonian system in the sense of Remark 13 (i.e. ∂H(x, p, u)/∂u solvable with
respect to u) is the same condition of invertibility of the Legendre transformation:
 2 
∂ L(x, u)
det 6= 0.
∂ui ∂uj
From PMP one gets also the stronger condition known as Weierstrass condition. From
the maximization condition:

H(x(t), p(t), u(t)) = maxn H(x(t), p(t), v),

v∈R

we get for every v ∈ Rn :

hp(t), u(t)i − L(x(t), u(t)) ≥ hp(t), vi − L(x(t), v),

and using the second of (26):

∂L
L(x(t), v) − L(x(t), u(t)) ≥ h (x(t), u(t)), (v − u(t))i. (27)
∂u
Condition (27) is known as Weierstraß condition. Weierstraß condition is stronger
(in a sense it is “more global”) than Euler Lagrange equations, because it is a neces-
sary conditions for a minimizer to be a strong local minimum (while Euler Lagrange
equations are necessary conditions for a weak local minimum). We recall that given
the minimization problem (22), x(.) ∈ C 1 is a strong local minimizer (respectively a
weak local minimizer) if there exists δ > 0 such that for every x̄(.) ∈ C 1 satisfying the
boundary conditions and kx − x̄kL∞ ≤ δ (respectively kx − x̄kL∞ + kẋ − x̄˙ kL∞ ≤ δ)
38 U. B OSCAIN AND B. P ICCOLI

we have 0 L(x(t), ẋ(t)) dt ≤ 0 L(x̄(t), x̄˙ (t)) dt. A classical example of a weak
RT RT

local minimizer that is not a strong local minimizer is x(.) ≡ 0 for the minimization
problem:  R1
minx(.)∈C 1 [0,1] 0 (ẋ2 − ẋ4 ) dt
x(0) = 0, x(1) = 0.
For more details on the Weierstraß condition see any book of Calculus of Variations,
for instance [22, 26].

3.4 An Detailed Application: the Grusin’s Metric

A nice and simple application of the PMP is the computation of geodesics for the
so called Grusin’s metric. The Grusin’s metric is a Riemannian metric on the plane
(x1 , x2 ) having singularities along the x2 -axis.
The problem of computing geodesics for a Riemannian metric on a manifold M of
dimension n can be presented as an optimal control problem with n controls belonging
to R (in fact it is a problem of calculus of variations):
n
(28)
X
ẋ = uj Fj (x)
j=1
v
T u n
Z uX
min t uj (t)2 dt, (29)
u(.) 0 j=1

where {Fj (x)} is an orthonormal basis for the given metric gx : Tx M × Tx M →

R, that is gx (Fj (x), Fk (x)) = δj,k (we recall that giving a metric on a manifold is
equivalent to give an orthonormal base).
One immediately verify that the cost (29) is invariant by a smooth reparameteriza-
Pn of the controls.
tion Moreover every minimizer parametrized by constant speed (i.e.
j=1 u j (t) 2
= const) is also a minimizer of the so called “energy cost”:
Z T n
(30)
X
min uj (t)2 dt.
u(.) 0 j=1

This is a simple consequence (exercise) of the Cauchy-Schwartz inequality (that holds

for two integrable functions f, g : [a, b] → R):
!2
Z b Z b Z b
f (t)g(t) dt ≤ f 2 (t)dt g 2 (t)dt
a a a
!2
Z b Z b Z b
f (t)g(t) dt = f 2 (t)dt g 2 (t)dt ⇔ f (t) = c g(t) (31)
a a a
A N I NTRODUCTION TO O PTIMAL C ONTROL 39

a.e. in [a, b], for some c ∈ R. The cost (30) is “nicer” than the cost (29) (is smooth and
quadratic) and usually, in Riemannian geometry and in control theory, one deals with
(30). The only thing to keep in mind is that, in (30), the final time must be fixed other-
wise there is no minimum (cfr. Exercise 3). Thus if we want to find minimizers via the
PMP (in this case one can use also Euler-Lagrange equations), then the Hamiltonian
must be equal to a (strictly) positive constant (cfr. Exercises 4).

If the vector fields {Fj (x)} are not always linearly independent, then the corre-
sponding metric gx presents singularities at points x where span{Fj (x)} is a proper
subset of Tx M . In this case the problem can not be treated as a problem of calculus
of variations, since there are nonholonomic constraints at the singular points.
An example is given by Grusin’s metric on the plane (x1 , x2 ), that is the metric
corresponding to the frame F1 = (1, 0)T and F2 = (0, x1 )T . In the language of
control theory, this translates into the following problem:

ẋ = u1 F1 (x) + u2 F2 (x),
     
x1 1 0
x= ∈ R2 , F1 (x) = , F2 (x) = .
x2 0 x1
Z T
u1 (t)2 + u2 (t)2 dt.


min
u(.) 0

The vector field F2 is zero on the x2 axis. In other words, on the x2 axis, every
trajectory of the system has velocity parallel to the x1 axis. Therefore the Grusin’s
metric explodes when one is approaching the x2 axis (see Figure 4).

Let us compute the geodesics starting from the origin using PMP. In this case the
Hamiltonian is (cfr. equation 20, p. 33):

H(x, p, λ0 , u) = hp, f (x, u)i + λ0 L(x, u) = hp, u1 F1 + u2 F2 i + λ0 (u21 + u22 ). (32)

There are not non trivial abnormal extremals. In fact let x(.) be an extremal trajectory
and p(.) the corresponding covector. If λ0 = 0, then it is possible to maximize (32)
with respect to the control only if

hp(t), F1 (x(t)i = hp(t), F2 (x(t))i = 0, (33)

(otherwise the supremum is infinity). Being λ0 = 0, p(t) can not vanish. It follows
that (33) can hold only in the region where the two vector fields are linearly dependent
(i.e. on the axis x1 = 0). But the only admissible trajectories, whose support belongs
to this set, are trivial (i.e. γ(t) = (0, const)). We then normalize λ0 = −1/2.
In this case, since the controls belong to R, the maximization condition:

H(x(t), p(t), λ0 , u(t)) = max H(x(t), p(t), λ0 , v)

v
40 U. B OSCAIN AND B. P ICCOLI

possible velocities
x2 F1

F1
F2 F2

x1
F1 F1

possible velocities

Figure 4: The Grusin’s orthonormal frame

can be rewritten as (cfr. Remark 13, p. 34):

∂H
= 0. (34)
∂u
From this condition we get controls as functions of (x, p):

u1 = p1 ,
u2 = p 2 x1 ,

where p = (p1 , p2 ). In this way the maximized Hamiltonian becomes:

1 2
H= (p + p22 x21 ).
2 1
Since the maximization condition define u uniquely as a smooth function of x and
p, we can write the Hamiltonian equations for H instead than H (cfr. again Remark
(13)). We have:
∂H
ẋ1 = = p1 , (35)
∂p1
∂H
ẋ2 = = p2 x21 , (36)
∂p2
A N I NTRODUCTION TO O PTIMAL C ONTROL 41

∂H
ṗ1 = − = −p22 x1 , (37)
∂x1
∂H
ṗ2 = − = 0. (38)
∂x2
From (38), we have p2 = a ∈ R. Equations (35),(37) are the equation for an harmonic
oscillator whose solution is:
x1 = A sin(at + φ0 ),
p1 = aA cos(at + φ0 ).
Since we are looking for geodesics starting from the origin we have φ 0 = 0 and
integrating (36), with x2 (0) = 0, we have:
aA2 A2
x2 (t) = t− sin(2at). (39)
2 4
In the expressions of x1 (t), x2 (t), p2 (t) there are still the two real constants A and a.
Requiring that the geodesics are parameterized by arclength (u21 (t) + u22 (t) = 1), that
corresponds to fix the level 1/2 of the Hamiltonian, we have:
1 1 1
= H = (p21 (t) + p22 (t)x21 (t)) = a2 A2 .
2 2 2
This means that A = ±1/a. Finally the formulas for extremal trajectories and controls
are given by the following two families (parametrized by a ∈ R):
x1 (t) = ± a1 sin(at)


1
t − 4a12 sin(2at)

x2 (t) = 2a

(40)
 u1 (t) = p1 (t) = ± cos(at)

u2 (t) = ax1 (t) = ± sin(at)


For a = 0, the definition is obtained by taking limits.

Due to the symmetries of the problem, and checking when extremal trajectories
self intersects, one can verify that each trajectory of (40) is optimal exactly up to time
t = π/|a|. In Figure 5A geodesics for some values of a are portrayed. While Figure
5B, illustrates some geodesics and the set of points reached in time t = 1. Notice that,
opposed to the Riemannian case, this set is not smooth.

3.5 Geometric Control Approach to Synthesis

Geometric control provides a standard method toward the construction of an Optimal
Synthesis for the family of problems {Px1 }x1 ∈M , of Definition 6, p. 31.
The approach is illustrated for systems with a compact control set as for the prob-
lem treated in Section 4. The following scheme, consisting of four steps, elucidates
the procedure for building an Optimal Synthesis.
42 U. B OSCAIN AND B. P ICCOLI

a=2.8 a=2 a=1.3

(family with −) a=1.3 a=1.3 (family with +) t=1
0.3
0.75
a=.5
0.2
a=2 a=2
0.5

a=3 0.25 a=3 0.1 a=4

a=0
-0.75 -0.5 -0.25 0.25 0.5 0.75
-1 -0.5 0.5 1
a=−3-0.25 a=−3
-0.1
a=−2 -0.5 a=−2
-0.2
-0.75

-0.3
a=−1.3 a=−1.3
A B

Figure 5: Geodesics and minimum time front (for t = 1) for the Grusin’s metric

Step 1. Use PMP and high order conditions to study the properties of optimal trajec-
tories.

Step 2. Use Step 1. to obtain a finite dimensional family of candidate optimal trajec-
tories.

Step 3. Construct a synthesis, formed of extremal trajectories, with some regularity

properties.

Step 4. Prove that the regular extremal synthesis is indeed optimal.

Let us describe in more detail each Step.

Step 1.
We stated the PMP that in some cases gives many information on optimal trajec-
tories. Beside high order maximum principle and envelopes, there are various higher
order conditions that can be used to discard some extremal trajectories that are not lo-
cally optimal. These conditions come from: symplectic geometric methods, conjugate
points, degree theory, etc.

Step 2. (finite dimensional reduction)

The family of trajectories and controls, on which the minimization is taken, is
clearly an infinite dimensional space. Thanks to the analysis of Step 1, in some cases
A N I NTRODUCTION TO O PTIMAL C ONTROL 43

it is possible to narrow the class of candidate optimal trajectories to a finite dimen-

sional family. This clearly reduces drastically the difficulty of the problem and allows
a deeper analysis. More precisely, one individuates a finite number of smooth con-
trols ui (x), such that every optimal trajectory is a finite concatenation of trajectories
corresponding to the vector fields f (x, ui (x)).

Step 3.
Once a finite dimensional reduction is obtained, one may construct a synthesis in
the following way. Assume that on the compact set R(τ ), there is a bound on the
number of arcs (corresponding to controls ui ) that may form, by concatenation, an
optimal trajectory. Then one can construct, by induction on n, trajectories that are
concatenations of n arcs and cut the not optimal ones. The latter operation produces
some special sets, usually called cut loci or overlaps, reached optimally by more than
one trajectory.
The above procedure is done on the base space M , however extremals admit lifts
to T ∗ M . Thus another possibility is to construct the set of extremals in the cotan-
gent bundle and project it on the base space. In this case, projection singularities are
responsible for singularities in the synthesis.

Step 4.
Even if a finite dimensional reduction is not obtained, one can still fix a finite
dimensional family of extremal trajectories and construct a synthesis on some part of
the state space. If the synthesis is regular enough then there are sufficiency theorems
ensuring the optimality of the synthesis.
These sufficiency theorems fit well also inside the framework of viscosity solution
to the corresponding Hamilton–Jacobi–Bellman equation.

The above approach is quite powerful, but in many cases it is not known how to reach
the end, namely to produce an Optimal Synthesis. For the problem treated in Section
4, we are able not only to construct an Optimal Synthesis under generic assumptions,
but also to give a topological classification of singularities of the syntheses and of the
syntheses themselves, see [14].

Remark 19 For general problems (not necessarily with bound on the control, e.g. with
quadratic cost), Steps 1 and 4 are still used, while Step 3 is not based on a finite
dimensional reduction.

3.6 Fuller Phenomenon and Dubins’ Car With Angular Accelera-

tion.
A major problem in using the Geometric Control Approach for the construction of
Optimal Syntheses is that the finite dimensional reduction, of Step 2., may fail to
44 U. B OSCAIN AND B. P ICCOLI

exits. The most famous example is that given by Fuller in [25]. The author considered
the infinite horizon Optimal Control problem:

ẋ1 = x2 , ẋ2 = u, |u| ≤ 1,

Z +∞
min x1 (t)2 dt.
0
Optimal trajectories reach the origin in finite time after an infinite number of switch-
ings between the control +1 and the control −1. Thus they are optimal also for the
finite horizon case.
The genericity of Fuller phenomenon, that is the presence of optimal trajectories
with an infinite number of switchings, was extensively analyzed in [39] and [61].
For the problem considered by Fuller, we have anyway the existence of an Opti-
mal Synthesis that is quite regular, see [48]. We give now an example of an optimal
control problems, coming from car–like robot models, for which Fuller phenomenon
is present and it is not known how to construct an Optimal Synthesis.
Dubins’ car with angular acceleration. One of the simplest model for a car–like
robot is the one known as Dubins’ car. In this model, the system state is represented
by a pair ((x, y), θ) where (x, y) ∈ R2 is the position of the center of the car and
θ ∈ S 1 is the angle between its axis and the positive x-axis. The car is assumed
to travel with constant (in norm) speed, the control is on the angular velocity of the
steering and is assumed to be bounded, thus we obtain the system:

 ẋ = cos(θ)
ẏ = sin(θ)
θ̇ = u


where |u| ≤ C (usually for simplicity one assumes C = 1).

A modified version of the Dubins’ car, in which the control is on the angular acceler-
ation of the steering, is given by:


 ẋ = cos(θ)
 ẏ = sin(θ)
(41)

 θ̇ = ω
ω̇ = u.


with |u| ≤ 1, (x, y) ∈ R2 , θ ∈ S 1 and ω ∈ R. We use the notation x = ((x, y), θ, ω)

and M = R2 × S 1 × R.
Pontryagin Maximum Principle conditions are now written as:

(PMP1) ṗ1 = 0, ṗ2 = 0, ṗ3 = p1 sin(θ) − p2 cos(θ), ṗ4 = −p3 ,

(PMP2) H(x(t), p(t), λ0 ) = H(x(t), p(t), λ0 , u(t)) = 0, that is p4 (t)u(t) = |p4 (t)|,
where p= (p1 , ..., p4 ).
A N I NTRODUCTION TO O PTIMAL C ONTROL 45

The function p4 (.) is the so called switching function (cfr. Section 4.1, p. 46). In fact
from (PMP2) it follows that:

• if p4 (t) > 0 (resp < 0) for every t ∈ [a, b], then u ≡ 1 (resp. u ≡ −1) on
[a, b]. In this case the corresponding trajectory x(.)|[a,b] is an arc of clothoid in
the (x, y) space.

• if p4 (t) ≡ 0 for every t ∈ [a, b], then u ≡ 0 in [a, b]. In this case the trajectory
x(.)|[a,b] is a straight line in the (x, y) space.

The main feature of this highly nongeneric problem is that an optimal trajectory can-
not contain points where the control jumps from ±1 to 0 or from 0 to ±1. In other
words a singular arc cannot be preceded or followed by a bang arc.
In [58] it is proved the presence of Fuller phenomenon. More precisely, there exist
extremal trajectories x(.) defined on some interval [a, b], which are singular on some
interval [c, b], where a < c < b, and such that p4 (.) does not vanish identically on any
subinterval of [a, c]. Moreover the set of zeros of p4 (.) consists of c together with an
increasing sequence {tj }∞ j=1 of points converging to c. At each tj , x(.) is switches the
control from +1 to −1 or viceversa.
An optimal path can thus have at most a finite number of switchings only if it is
a finite concatenation of clothoids, with no singular arc. Existence of optimal trajec-
tories (not only extremal) presenting chattering was proved by Kostov and Kostova in
[36]. More precisely if the distance between the initial and final point is big enough,
then the shortest path can not be a finite concatenation of clothoids.

4 The Minimum Time Problem for Planar Control

Affine Systems
In this section we give an introduction the theory of “optimal synthesis for planar
control affine systems” developed by Sussmann, Bressan, and the authors in [15, 16,
17, 18, 19, 46, 47, 55, 57], and recently collected in [14]. The main ingredients of that
theory are described next.
Consider a control system on R2 :

ẋ = F (x) + uG(x), x ∈ R2 , |u| ≤ 1, (42)

and the problem of reaching every point of the plane in minimum time starting from
the origin.
To guarantee completeness of the vector fields and to be able to compute deriva-
tives of arbitrary order, we assume the following:

(H) F, G are C ∞ vector fields, uniformly bounded with all their partial derivatives on
R2 .
46 U. B OSCAIN AND B. P ICCOLI

The PMP takes a particularly simple form for (42) and one can easily see that (un-
der generic conditions) controls are always either bang-bang, that is corresponding to
constant control u = +1 or u = −1, or singular (i.e. corresponding to trajectories
whose switching functions vanishes, see Definition 12 below and Remark 22). The
latter can happen only on the set where the vector field G and the Lie bracket [F, G]
are parallel (this generically is a one dimensional submanifold of the plane).
Moreover one can also predict (see Section 4.2) which kind of “switchings” (See
Definition 12) can happen on each region of the plane. We first need to state the PMP
for the system (42) and introduce some notations.

4.1 Pontryagin Maximum Principle and Switching Functions

The Pontryagin Maximum Principle (see Section 3.2.2, p. 32) in this special case
states the following.
Remark 20 (Notation in the following version of the PMP)
• Tx∗ R2 = (R2 )∗ denotes the set of row vectors.
• the duality product < . , . > is now simply the matrix product and it is
indicated by “ · ”.
• the Hamiltonian does not include the cost factor λ0 L, (that in this case is just
λ0 ). The condition H = 0 become then H + λ0 = 0. With this new definition
of Hamiltonian, abnormal extremals are the zero levels of the H.

Theorem 6 (PMP for the minimum time problem for planar control affine sys-
tems)
Define for every (x, p, u) ∈ R2 × (R2 )∗ × [−1, 1]:

H(x, p, u) = p · F (x) + u p · G(x)

and:

H(x, p) = max{p · F (x) + u p · G(x) : u ∈ [−1, 1]}. (43)

If γ : [0, a] → R2 is a (time) optimal trajectory corresponding to a control u :

[0, a] → [−1, 1], then there exist a nontrivial field of covectors along γ, that is a
Lipschitz function λ : [0, a] → (R2 )∗ never vanishing, and a constant λ0 ≤ 0 such
that for a.e. t ∈ Dom(γ):
i) λ̇(t) = −λ(t) · (∇F + u(t)∇G)(γ(t)),
ii) H(γ(t), λ(t), u(t)) + λ0 = 0,
iii) H(γ(t), λ(t), u(t)) = H(γ(t), λ(t)).
A N I NTRODUCTION TO O PTIMAL C ONTROL 47

Remark 21 Notice that, since the Lagrangian cost is constantly equal to 1, the condi-
tion (λ(t), λ0 ) 6= (0, 0), for all t ∈ [0, T ], given in Definition 8, p. 33, now becomes
λ(t) 6= 0 for all t. In fact λ(t) ≡ 0 with the condition (PMP2) of Definition 8 implies
λ0 = 0. To see what happens for vanishing Lagrangians cfr. Exercise 8, p. 61).

Definition 11 (Switching Function) Let (γ, λ) : [0, τ ] → R2 ×(R2 )∗ be an extremal

pair. The corresponding switching function is defined as φ(t) := λ(t)·G(γ(t)). Notice
that φ(·) is absolutely continuous.

From the PMP one immediately get that the switching function describes when the
control switches from +1 to -1 and viceversa. In fact we have:
Lemma 2 Let (γ, λ) : [0, τ ] → R2 × (R2 )∗ be an extremal pair and φ(·) the corre-
sponding switching function. If φ(t) 6= 0 for some t ∈]0, τ [, then there exists ε > 0
such that γ corresponds to a constant control u = sgn(φ) on ]t − ε, t + ε[.
Proof. There exists ε > 0 such that φ does not vanish on ]t − ε, t + ε[. Then from
condition iii) of PMP we get u = sgn(φ). 

Reasoning as in Lemma 2 one immediately has:

Lemma 3 Assume that φ has a zero at t, φ̇(t) is strictly greater than zero (resp.
smaller than zero) then there exists ε > 0 such that γ corresponds to constant control
u = −1 on ]t − ε, t[ and to constant control u = +1 on ]t, t + ε[ (resp. to constant
control u = +1 on ]t − ε, t[ and to constant control u = −1 on ]t, t + ε[).
We are then interested in differentiating φ:
Lemma 4 Let (γ, λ) : [0, τ ] → R2 × (R2 )∗ be an extremal pair and φ the corre-
sponding switching function. Then φ(·) is continuously differentiable and it holds:

φ̇(t) = λ(t) · [F, G](γ(t)).

Proof. Using the PMP we have for a.e. t:

d
φ̇(t) = (λ(t) · G(γ(t))) = λ̇(t) · G(γ(t)) + λ · Ġ(γ(t))
dt
= −λ(t)(∇F + u(t)∇G)(γ(t)) · G(γ(t)) + λ · ∇G(γ(t))(F + u(t)G)(γ(t))
= λ(t) · [F, G](γ(t)).

Since φ(·) is absolutely continuous and λ(t) · [F, G](γ(t)) is continuous, we deduce
that φ is C 1 . 

Notice that if φ(·) has no zeros then u is a.e. constantly equal to +1 or −1. Next
we are interested in determining when the control may change sign or may assume
values in ] − 1, +1[. For this purpose we give first the following definition:
48 U. B OSCAIN AND B. P ICCOLI

Definition 12 Let u(.) : [a, b] → [−1, 1] be a control for the control system (42).

• u(.) is said to be a bang control if for almost every t ∈ [a, b], u(t) is constant
and belongs to a point of the set {+1, −1}.

• A switching time of u(.) is a time t ∈ [a, b] such that for every ε > 0, u(.) is not
bang on (t − ε, t + ε) ∩ [a, b].

• If uA : [a1 , a2 ] → [−1, 1] and uB : [a2 , a3 ] → [−1, 1] are controls, their

concatenation uB ∗ uA is the control:

uA (t) for t ∈ [a1 , a2 ]
(uB ∗ uA )(t) :=
uB (t) for t ∈]a2 , a3 ].

The control u(.) is called bang-bang if it is a finite concatenation of bang arcs.

• A trajectory of (42) is a bang trajectory, (resp. bang-bang trajectory), if it cor-

responds to a bang control, (resp. bang-bang control).

• An extremal trajectory γ defined on [c, d] is said to be singular if the switching

function φ vanishes on [c, d].

Remark 22 A singular extremal in the sense above is also an endpoint singular ex-
tremal in the sense of Section 3.2.3, p. 34. In fact for these trajectories the Hamil-
tonian is independent from the control. In the following we use the term singular
trajectories with the same meaning of endpoint singular extremal.

Remark 23 On any interval where φ has no zeroes (respectively finitely many zeroes)
the corresponding control is bang (respectively bang-bang).
Singular trajectories are studied in details below. In Figure 6 we give an example
to illustrate the relationship between an extremal trajectory and the corresponding
switching function. The control ϕ, corresponding to the singular arc, is computed
below, see Lemma 6, p. 51.

4.2 Singular Trajectories and Predicting Switchings

In this section we give more details on singular curves and explain how to predict
which kind of switching can happen, using properties of the vector fields F and G.
A key role is played by the following functions defined on R2 :

∆A (x) := Det(F (x), G(x)) = F1 (x)G2 (x) − F2 (x)G1 (x), (44)

∆B (x) := Det(G(x), [F, G](x)) = G1 (x)[F, G]2 (x) − G2 (x)[F, G]1 (x).(45)
A N I NTRODUCTION TO O PTIMAL C ONTROL 49

φ( t)

bang (u=+1) bang(u=−1) singular (u=ϕ ) bang (u=+1)

Figure 6: The switching function

The set of zeros ∆−1

A (0), ∆B (0), of these two functions, are respectively the set
−1

of points where F and G are parallel and the set of points where G is parallel to
[F, G]. These loci are fundamental in the construction of the optimal synthesis. In
fact, assuming that they are smooth one dimensional submanifolds of R 2 , we have the
following:
• on each connected component of M \ (∆−1 A (0) ∪ ∆B (0)), every extremal
−1

trajectory is bang-bang with at most one switching. Moreover, at a switching

time, the value of the control switches from −1 to +1 if fS := −∆−1 B (0) \
∆−1
A (0) > 0 and from +1 to −1 if f S < 0;
• the support of singular trajectories is always contained in the set ∆−1
B (0).

The synthesis is built by induction on the number of switchings of extremal trajecto-

ries, following the classical idea of canceling, at each step, the non optimal trajectories.
We refer to Chapter 2 of [14], for more details about this construction.

First we study points out of the regions ∆−1 −1

A (0) ∪ ∆B (0).

Definition 13 A point x ∈ R2 is called an ordinary point if x ∈/ ∆−1 −1

A (0) ∪ ∆B (0).
2
If x is an ordinary point, then F (x), G(x) form a basis of R and we define the
scalar functions f, g to be the coefficients of the linear combination: [F, G](x) =
f (x)F (x) + g(x)G(x).
The function f is crucial in studying which kind of switchings can happen near ordi-
nary points. But first we need a relation between f , ∆A and ∆B :
50 U. B OSCAIN AND B. P ICCOLI

Lemma 5 Let x an ordinary point then:

∆B (x)
f (x) = − . (46)
∆A (x)

Proof. We have:


∆B (x) = Det G(x), [F, G](x) = Det(G(x), f (x)F (x) + g(x)G(x))


= f (x)Det G(x), F (x) = −f (x)∆A (x).

On a set of ordinary points the structure of optimal trajectories is particularly sim-

ple:
Theorem 7 Let Ω ∈ R2 be an open set such that every x ∈ Ω is an ordinary point.
Then, in Ω, all extremal trajectories γ are bang-bang with at most one switching.
Moreover if f > 0 (resp. f < 0) throughout Ω then γ corresponds to control +1, −1
or has a −1 → +1 switching (resp. has a +1 → −1 switching).
Proof. Assume f > 0 in Ω, the opposite case being similar. Let (γ, λ) be an extremal
pair such that γ is contained in Ω. Let t̄ be such that λ(t̄) · G(γ(t̄)) = 0, then:

φ̇(t̄) = λ(t̄) · [F, G](γ(t̄)) = λ(t̄) · (f F + gG)(γ(t̄)) = f (γ(t̄)) λ(t̄) · F (γ(t̄)).

Now from PMP, we have H(γ(t̄), λ(t̄)) = λ(t̄) · F (γ(t̄)) ≥ 0. Hence φ̇ > 0, since
F (γ(t̄)), and G(γ(t̄)) are independent. This proves that φ has at most one zero with
positive derivative at the switching time and gives the desired conclusion. 

We are now interested in understanding what happens at points that are not ordinary.
Definition 14 A point x ∈ R2 is called an non ordinary point if x ∈ ∆−1
A (0) ∪
∆−1
B (0).
In the following we study some properties of singular trajectories in relation to non
ordinary points on which ∆B = 0.
Definition 15 • An non ordinary arc is a C 2 one-dimensional connected embedded
submanifold S of R2 with the property that every x ∈ S is a non ordinary point.

• A non ordinary arc is said isolated if there exists a set Ω satisfying the following
conditions:
(C1) Ω is an open connected subset of R2 .
(C2) S is a relatively closed subset of Ω.
A N I NTRODUCTION TO O PTIMAL C ONTROL 51

(C3) If x ∈ Ω \ S then x is an ordinary point.

(C4) The set Ω \ S has exactly two connected components.
• A turnpike (resp. anti-turnpike) is an isolated non ordinary arc that satisfies the
following conditions:
(S1) For every x ∈ S the vectors X(x) and Y (x) are not tangent to S and point to
opposite sides of S.
(S2) For every x ∈ S one has ∆B (x) = 0 and ∆A (x) 6= 0.
(S3) Let Ω be an open set which satisfies (C1)–(C4) above and ∆A 6= 0 on Ω. If
Ω− and Ω+ are the connected components of Ω \ S labeled in such a way
F (x)−G(x) points into Ω− and F (x)+G(x) points into Ω+ , then the function
f satisfies

f (x) > 0 (resp. f (x) < 0) on Ω+

f (x) < 0 (resp. f (x) > 0) on Ω−

The following Lemmas describes the relation between turnpikes, anti-turnpikes and
singular trajectories. In Lemma 6 we compute the control corresponding to a trajec-
tory whose support is a turnpike or an anti-turnpike. In Lemma 7 we prove that if this
control is admissible (that is the turnpike or the anti-turnpike is regular, see Defini-
tion 16) then the corresponding trajectory is extremal and singular. In Lemma 8 we
show that anti-turnpike are locally not optimal.
Lemma 6 Let S be a turnpike or an anti-turnpike and γ : [c, d] → R2 a trajectory of
(42) such that γ(c) = x0 ∈ S. Then γ(t) ∈ S for every t ∈ [c, d] iff γ corresponds to
the feedback control (called singular):
∇∆B (x) · F (x)
ϕ(x) = − , (47)
∇∆B (x) · G(x)
Proof. Assume that γ([c, d])) ⊂ S and let u be the corresponding control, that is
γ̇(t) = F (γ(t)) + u(t)G(γ(t)), for almost every t. From ∆B (γ(t)) = 0, for a.e. t we
have:
d
0 = ∆B (γ(t)) = ∇∆B · (F (γ(t)) + u(t)G(γ(t))).
dt
This means that at the point x = γ(t) we have to use control ϕ(x) given by (47). 
Definition 16 (regular turnpike or anti-turnpike) We say that a turnpike or anti-
turnpike S is regular if |ϕ(x)| < 1 for every x ∈ S.

Lemma 7 Let (γ, λ) : [0, t̄] → R2 be an extremal pair that verifies γ(t̄) = x, x ∈ S
where S is a turnpike or an anti-turnpike, and λ(t̄) · G(γ(t̄)) = 0. Moreover let
γ 0 : [0, t0 ] → R2 (t0 > t̄) be a trajectory such that:
52 U. B OSCAIN AND B. P ICCOLI

• γ 0 |[0,t̄] = γ,

• γ 0 ([t̄, t0 ]) ⊂ S.

Then γ 0 is extremal. Moreover if φ0 is the switching function corresponding to γ 0 then

φ0 |[t̄,t0 ] ≡ 0.

For the proof see [14].

Lemma 8 Let S be an anti-turnpike and γ : [c, d] → R2 be an extremal trajectory

such that γ([c, d]) ⊂ S. Then γ is not optimal.

Proof. Choose an open set Ω containing γ([c, d]) such that ∆A 6= 0 on Ω and define
the differential form ω on Ω by ω(F ) = 1, ω(G) = 0. Let γ1 : [c, d1 ] → Ω be any
trajectory such that γR1 (c) = γ(c), γ1R(d1 ) = γ(d), γ1 (t) ∈
/ SR for every t ∈]c, d1 [.
Notice that d − c = γ ω, d1 − c = γ1 ω. Hence d − d1 = γ −1 ∗γ ω where γ1−1
1
is γ1 run backward. Assume that γ1−1 ∗ γ is oriented counterclockwise, being similar
the opposite case. Then by Stokes’ Theorem, d − d1 = A dω where A is the region
R

enclosed by γ1−1 ∗ γ. Now dω(F, G) = F · ∇ω(G) − G · ∇ω(F ) − ω([F, G]) = −f ,

by definition of ω. Since dω(F, G) = ∆A dω(∂x , ∂y ), we get
Z  
f
d − d1 = − dx dy.
A ∆A

If ∆A is positive (resp. negative) then F + G (resp. F − G) points to the side of S

where γ1 is contained and, by definition of anti-turnpike, f is negative (resp. positive)
on A. We conclude that d > d1 and γ is non optimal. 
One finally gets:

Theorem 8 Let γ : [0, t̄] → R2 be an optimal trajectory that it is singular on some

interval [c, d] ⊂ [0, t̄]. Then, under generic conditions, Supp(γ|[c,d]) is contained in
a regular turnpike S.

Proof. From φ ≡ 0 on [c, d] it follows φ̇ ≡ 0 on [c, d]. By Lemma 4, Supp(γ|[c,d]) ⊂

B (0). Under generic conditions, ∆B (0) ∩ R(t̄) is formed by a finite number
∆−1 −1

of turnpikes, anti-turnpikes and isolated points (at intersections with ∆ −1

A (0)). By
Lemma 8 we conclude. 

Figure 7 illustrates the possible switchings in the connected regions of R 2 \(∆−1

A (0)∪
B (0)) in relation with the sign of f , and an example of extremal trajectory contain-
∆−1
ing a singular arc.
A N I NTRODUCTION TO O PTIMAL C ONTROL 53

−1
∆ Α(0)

f<0 (+1 −1)

f>0 (−1 +1)

f>0 (−1 +1)

f<0 (+1 −1)

+1
sing.
−1
−1 ∆Β(0)

Figure 7: Possible switchings in the connected regions of R2 \ (∆−1 A (0) ∪ ∆B (0)) in

−1

relation with the sign of f , and an example of extremal trajectory containing a singular
arc.

4.3 The Optimal Synthesis: A Brief Summary of Results

Consider the control system (42) assume (H), and focus on the problem of reaching
every point of the plane in minimum time starting from the origin x0 = (0, 0), under
the additional hypothesis that F (x0 ) = 0.

Remark 24 The hypothesis F (x0 ) = 0, i.e. x0 is a stable point for F , is very natural.
In fact, under generic assumptions, it guarantees local controllability. Moreover if we
reverse the time, we obtain the problem of stabilizing in minimum time all the points
of M to x0 . For this time-reversed problem, the stability of x0 guarantees that once
reached the origin, it is possible to stay.

In [46, 47] (see also [14]) it was proved that this problem with F and G generic,
admits an optimal synthesis in finite time T . More precisely, if we define the reachable
set in time T :

R(T ) := {x ∈ R2 : ∃ bx ∈ [0, T ] and a trajectory

γx : [0, bx ] → M of (42) such that γx (0) = x0 , γx (bx ) = x}, (48)

an optimal synthesis in time T , for the control system (42) under the additional hy-
pothesis F (x0 ) = 0, is a family of time optimal trajectories Γ = {γx : [0, bx ] 7→ R2 ,
54 U. B OSCAIN AND B. P ICCOLI

FCs of kind Y FC of kind C FC of kind S FC of kind K

(X similar)

Figure 8:

x ∈ R(T ) : γ(0) = x0 , γ(bx ) = x}. Moreover it was also proved that there exists a
stratification of R(T ) (roughly speaking a partition of R(T ) in manifolds of different
dimensions, see [14] for more details) such that the optimal synthesis can be obtained
from a feedback u(x) satisfying:
• on strata of dimension 2, u(x) = ±1,
• on strata of dimension 1, called frame curves (FC for short), u(x) = ±1 or
u(x) = ϕ(x), where ϕ(x) is defined by (47).
The strata of dimension 0 are called frame points (FP). Every FP is an intersection of
two FCs. A FP x, which is the intersection of two frame curves F1 and F2 is called an
(F1 , F2 ) Frame Point. In [47] (see also [14]), it is provided a complete classification
of all types of FPs and FCs, under generic conditions. All the possible FCs are:
• FCs of kind Y (resp. X), corresponding to subsets of the trajectories γ + (resp.
γ − ) defined as the trajectory exiting x0 with constant control +1 (resp. con-
stant control −1);
• FCs of kind C, called switching curves, i.e. curves made of switching points;
• FCs of kind S, i.e. singular trajectories;
• FCs of kind K, called overlaps and reached optimally by two trajectories com-
ing from different directions;
• FCs which are arcs of optimal trajectories starting at FPs. These trajectories
“transport” special information.
The FCs of kind Y, C, S, K are depicted in Fig. 8.
There are eighteen topological equivalence classes of FPs. They are showed in Figure
9. A detailed description can be found in [14, 47].
A N I NTRODUCTION TO O PTIMAL C ONTROL 55

(X,Y) (Y,C) 1 (Y,C)

2 γ γ
Y 0 or A
C C

Y Y
X

(Y,C) (Y,S) (Y,K)1

3 C
γ S K
0

Y
Y Y
γ γ
0 or A

(Y,K) 2 (Y,K) (C,C) 1

γ 3 C
k
C
Y
Y γ
0

K K
(C,C) (C,S) 1 (C,S) 2 γ
2 0
γ S
C γ γ 0
0 or A
γ
C 0

C S C
(C,K) 1 (C,K) 2 (S,K) 1
K
K K S

γ
0

C γ C
k
(S,K) (S,K) 3 K
(K,K)
γ 2
K
k

γ
k

S K S K

Figure 9: The FPs of the Optimal Synthesis.

56 U. B OSCAIN AND B. P ICCOLI

4.4 Some Examples

As first example, let us compute the time optimal synthesis for the problem given in
the introduction. 
ẋ1 = x2
, |u| ≤ 1
ẋ2 = u.
With respect to the introduction, here we reverse the time and we consider the problem
of reaching every point of the plane starting from x0 = (0, 0). Writing the system as
ẋ = F + uG we have:
     
x2 0 −1
F = , G= , [F, G] = .
0 1 0

Hence:

∆A (x) = det(F, G) = x2 , ∆B (x) = det(G, [F, G]) = 1,

f (x) = −∆B /∆A = −1/x2 . (49)

Since ∆−1B (0) = ∅ there are not singular trajectories (cfr. Theorem 8, p. 52). More-
over, in the region x2 > 0 (resp. x2 < 0), the only switchings admitted are from
control +1 to control −1 (resp. from control −1 to control +1, cfr. Theorem 7).
The trajectories corresponding to control +1 (resp. −1) and starting from the point
(x10 , x20 ) are the two parabolas:

x1 = x10 + x20 t ± 12 t2

(50)
x2 = x20 ± t.

To compute the switching times, we have to solve the equation for the covector ṗ =
−p(∇F + u∇G):

p1 = p10
(51)
p2 = p20 − p10 t

The switching function is then φ(t) = p(t)·G(x(t)) = p20 −p10 t. It is clear that every
time optimal trajectory has at most one switching because, for every initial conditions
p10 and p20 , the function φ has at most one zero. After normalizing p210 + p220 = 1, the
switchings are parametrized by an angle belonging to the interval [0, 2π]. The time
optimal synthesis is showed in Figure 10 A. In Figure 10 B it is presented the time
optimal synthesis for the same control system, but starting from a ball of radius 1/2.
Computing the explicit expression of the switching curves is an exercise for the reader
(see Exercise 9, p. 61).
A N I NTRODUCTION TO O PTIMAL C ONTROL 57

x2 x2
+1 +1

x1 x1

−1
−1
A B

Figure 10:

A more rich example presenting the FPs of kind (X, Y ), (Y, K)1 , (Y, S) is the
following: 
ẋ1 = u
.
ẋ2 = x1 + 12 x21
Let us build the time optimal synthesis for a fixed time T > 2. The X– and Y –
trajectories can be described giving x2 as a function of x1 and are, respectively, cubic
polynomials of the following type:
x31 x2
x2 = − − 1 +α α∈R
6 2
x31 x2
x2 = + 1 +α α ∈ R. (52)
6 2
With a straightforward computation we obtain:
 
0
[F, G] =
−1 − x1
then the system is locally controllable and:
 
1 0
∆B (x) = det = −1 − x1 . (53)
0 −1 − x1
From equation (53) it follows that every turnpike is subset of {(x1 , x2 ) ∈ R2 : x1 =
−1}. Indeed, the synthesis contains the turnpike:
 
1
S = (x1 , x2 ) : x1 = −1, x2 ≤ − .
3
58 U. B OSCAIN AND B. P ICCOLI

Given b, consider the trajectories γ1 : [0, b] 7→ R2 for which there exists t0 ∈ [0, b]
such that γ1 |[0, t0 ] is a Y –trajectory and γ1 |[t0 , b] is an X–trajectory, and the trajecto-
ries γ2 : [0, b] 7→ R2 , b > 2, for which there exists t1 ∈ [2, b] such that γ2 |[0, t1 ] is an
X–trajectory and γ2 |[t1 , b] is a Y –trajectory. For every b > 2, these trajectories cross
each other in the region of the plane above the cubic (52) with α = 0 and determine
an overlap curve K that originates from the point (−2, − 32 ). We use the symbols
x+− (b, t0 ) and x−+ (b, t1 ) to indicate, respectively, the terminal points of γ1 and γ2
above. Explicitly we have:
(2t0 − b)3 (2t0 − b)2 t3
x+−
1 = 2t0 − b x+−
2 =− − + t20 + 0 (54)
6 2 3

(b − 2t1 )3 (b − 2t1 )2 t3
x−+
1 = b − 2t1 x−+
2 = + − t21 + 1 . (55)
6 2 3
Now the equation:

x+− (b, t0 ) = x−+ (b, t1 ), (56)

as b varies in [2, +∞[, describes the set K. From (54), (55) and (56) it follows:
 
t0 = b − t 1 t1 − 2t21 + (2 + 3b)t1 + (−b2 − 2b) = 0.

Solving for t1 we obtain three solutions:

b
t1 = 0, t1 = b, t1 = 1 + . (57)
2
The first two of (57) are trivial, while the third determines a point of K, so that:
 
2
K = (x1 , x2 ) : x1 = −2, x2 ≥ − .
3
The set R(T ) is portrayed in Fig. 11.

Bibliographical Note
For the problem of controllability, classical results are found in the papers by Krener
[37], and Lobry [41]. For the problem of controllability on Lie groups see the papers
by Jurdjevic-Kupka [30, 31], Jurdjevic-Sussmann [32], Gauthier-Bornard [27], and
Sachkov [60].
The issue of existence for Optimal Control as well as for Calculus of Variations is
a long standing research field. For a review of classical available results we refer to
[22, 23, 26].
A N I NTRODUCTION TO O PTIMAL C ONTROL 59

Y
K

X
A

B
S
A=(−1,−1/3)
B=(−2,−2/3)

Figure 11: Example 1, (X, Y ), (Y, K)1 , (Y, S) Frame Points

The terms Abnormal Extremals and Singular Extremals (in our case called End-
point Singular Extremals) are used with several different meanings. For some results
see [1, 2, 11, 12, 13, 44, 45].
The first sufficiency theorem, for extremal synthesis, was given by Boltyanskii in
[10]. Then various generalization appeared, [20, 21, 48]. In particular the result of
[48], differently from the previous ones, can be applied to systems presenting Fuller
phenomena.
In Step 3 of Section 3.5, an alternative method for the construction of the Optimal
Synthesis is indicated. Namely, to construct all extremals in the cotangent bundle.
This second method is more involved, but induces a more clear understanding of the
fact that projection singularities are responsible for singularities in the Optimal Syn-
thesis. This approach was used in [15, 16, 33, 34].
For sufficiency theorems, in the framework of viscosity solutions to Hamilton–
Jacobi–Bellman equation (mentioned in Step 4 of Section 3.5), see [6].
Dubins’ car problem was originally introduced by Markov in [43] and studied by
Dubins in [24]. In particular Dubins proved that every minimum time trajectory is con-
catenation of at most three arcs, each of which is either an arc of circle or a straight
line. If we consider the possibility of non constant speed and admit also backward
motion, then we obtain the model proposed by Reed and Shepp [50]. A family of
time optimal trajectories, that are sufficiently reach to join optimally any two points,
was given in [59]. Now the situation is more complicated since there are 46 possible
60 U. B OSCAIN AND B. P ICCOLI

combination of straight lines and arcs of circles. Then a time optimal synthesis was
built by Soueres and Laumond in [51].
Time optimal trajectories for the system (41) were studied mainly by Boissonnat,
Cerezo, Kostov, Kostova, Leblond and Sussmann, see [8, 9, 35, 36, 58].
The concept of synthesis was introduced in the pioneering work of Boltianski [10].
Properties of extremal trajectories for single input affine control systems on the plane
were first studied by Baitman in [4, 5] and later by Sussmann in [55, 56].
The existence of an optimal synthesis for analytic systems and under generic con-
ditions for C ∞ systems was proved by Sussmann in [57] and by Piccoli in [46] respec-
tively.
Finally, a complete classification of generic singularities and generic optimal syn-
thesis is given by Bressan and Piccoli in [19, 47].

Exercises
Exercise 1 Consider the control system on the plane:
     
ẋ sin(x) G1 (x, y)
= +u , u ∈ R, (58)
ẏ 0 G2 (x, y)

where G1 , G2 ∈ C 1 (R2 , R) and G1 (x, y), G2 (x, y) > 0 for every (x, y) ∈ R2 . Find
the set of points where Theorem 2, p. 24 permits to conclude that there is local con-
trollability. Is it true that in all other points the system is not locally controllable?

Exercise 2 Consider the control system ẋ = F (x) + uG(x), u ∈ R where:

 
x1
x =  x2  , x21 + x22 + x23 = 1,
x3
   
x2 cos(α) 0
F (x) =  −x1 cos(α)  , G(x) =  x3 sin(α)  , α ∈]0, π/2[.
0 −x2 sin(α)

Find where the system is locally controllable and prove that it is (globally) control-
lable.
[Hint: to prove global controllability, find a trajectory connecting each couple of
points.]

Exercise 3 Prove that, if the controls take values in the whole Rm , then there exists
no minimum for the minimum time problem of system (18). Moreover, show that, for
a distributional problem with quadratic cost (the so called subriemannian problem), if
we do not fix the final time, then there exists no minimum.
A N I NTRODUCTION TO O PTIMAL C ONTROL 61

Exercise 4 Prove that if we fix the final time in the PMP, then condition (PMP2)
becomes
(PMP2bis) For a.e. t ∈ [0, T ],

H(γ(t), λ(t), λ0 ) = H(γ(t), λ(t), λ0 , u(t)) = const ≥ 0;

and find the value of the constant for a given γ.

Exercise 5 Consider the distributional system ẋ = u1 F1 (x) + u2 F2 (x), x ∈ M ,

u1 , u2 ∈ R, with quadratic cost (19) and fixed final time T . Prove that (λ, λ0 ) can be
chosen up to a multiplicative positive constant and if H(γ(t), λ(t), λ0 , u(t)) = 21 with
λ0 = − 21 (see (PMP2bis) of the previous Exercise), then the extremal is parametrized
by arclength (i.e. u21 (t) + u22 (t) = 1).
[Hint: Use the condition ∂H ∂u = 0 implied by the condition (PMP2bis).]

Exercise 6 Consider the control system ẋ = f (x, u), x ∈ M and u ∈ U = Rm , and

the problem (17) with TT = {x1 }, x1 ∈ M . Prove that if γ is an endpoint singular
trajectory and, for every t ∈ Dom(γ), the function u → hp(t), f (γ(t), u)i is strictly
convex, then γ is an abnormal extremal.

Exercise 7 Consider a distributional system ẋ = i ui Fi (x). ProvePthat for the

P
problem (17), with ψ ≡ 0, TT = {x1 } and the quadratic cost L = i ui , every
2

endpoint singular extremal is an abnormal extremal and viceversa (what happens if

x0 = x1 ?). Prove the same for minimum time with bounded control.

Exercise 8 Consider the system ẋ = F (x) + uG(x), u ∈ R, x = (x1 , x2 ) ∈ R2 and

the optimal control problem with Lagrangian L(x) = x22 , ψ = 0, initial point (0, 0)
and terminal point (c, 0). Assume that, for every x, we have F1 (x) > 0, G2 (x) > 0,
∆A (x) := F1 (x)G2 (x) − F2 (x)G1 (x) > 0. Prove that there exists a monotone
function φ(t), such that γ(t) = (φ(t), 0) is the unique optimal trajectory and show
that it is extremal for every covector λ ≡ 0 and λ0 < 0.
[Hint: find a control such that the corresponding trajectory has zero cost.]

Exercise 9 Compute the time optimal synthesis for the control system on the plane
ẋ1 = x2 , ẋ2 = u, |u| ≤ 1 with source the unit closed ball.

References
[1] A.A. Agrachev, A. Sarychev, “Abnormal Sub-Riemannian Geodesics: Morse
Index and Rigidity,” Ann. Inst. H. Poincaré Anal. Non Linéaire Vol. 13, No. 6
(1996), pp. 635-690.
[2] A.A. Agrachev, A. Sarychev, “ On Abnormal Extremals for Lagrange Variational
Problems”, J. Math. Systems Estim. Control, Vol. 8, No. 1 (1998), pp. 87-118.
62 U. B OSCAIN AND B. P ICCOLI

[3] A.A. Agrachev, Yu.L. Sachkov, “Lectures on Geometric Control Theory”,

Preprint SISSA 38/2001/M, May 2001, SISSA, Trieste. .

[4] M. M. Baitmann, “Controllability Regions on the Plane”, Differential Equations,

Vol.14 (1978), pp. 407-417.

[5] M. M. Baitmann, “Switching Lines in the Plane”, Differential Equations, Vol.14

(1978), pp. 1093-1101.

[6] M. Bardi and I. Capuzzo-Dolcetta, “Optimal control and viscosity solutions of

Hamilton-Jacobi-Bellman equations”, Systems & Control: Foundations & Ap-
plications, Birkhäuser, Boston, 1997.
[7] A. Bloch, J. Baillieul, J. Marsden, “Nonholonomic Mechanics and Control”
Springer Series: Interdisciplinary Applied Mathematics, Vol. 24, 2003.
[8] J.D. Boissonnat, A. Cerezo and J. Leblond, “A note on shortest path in the plane
subject to a constraint on the derivative of the curvature”, Research Report 2160,
INRIA, 1998.

[9] J.D. Boissonnat, A. Cerezo, E. Degtiariova-Kostova, V. Kostov, and J. Leblond

“Shortest plane paths with bounded derivative of the curvature.” C. R. Acad. Sci.,
329-I, pp.613–618, 1999.
[10] V. Boltyanskii, “Sufficient Condition for Optimality and the Justification of the
Dynamic Programming Principle”, SIAM J. Control and Optimization, Vol.4,
(1966), pp. 326-361.

[11] B. Bonnard, M. Chyba, ”The role of singular trajectories in control theory”,

Springer SMAI, Vol.40.

[12] B. Bonnard, M. Chyba, E. Trelat, “Sub-Riemannian geometry, one-parameter

deformation of the Martinet flat case.” J. Dynam. Control Systems 4 (1998), no.
1, 59–76.

[13] B. Bonnard, I. Kupka, “Théorie des singularites de l’application entree/sortie et

optimalite des singulieres dans le probleme du temps minimal”, Forum Mathe-
maticum Vol.5 (1993), pp. 111-159.
[14] U. Boscain, B, Piccoli, Optimal Synthesis for Control Systems on 2-D Manifolds,
Springer, SMAI, Vol.43.

[15] U. Boscain and B. Piccoli, “Extremal syntheses for generic planar systems”,
Journal of Dynamical and Control Systems Vol.7 No.2 (2001), pp. 209-258.
[16] U. Boscain and B. Piccoli, “On automaton recognizability of abnormal ex-
tremals”, SIAM J. Control and Opt., Vol. 40 No.5 pp. 1333-1357. (2002).
A N I NTRODUCTION TO O PTIMAL C ONTROL 63

[17] U. Boscain and B. Piccoli, “Morse Properties for the Minimum Time Function
on 2-D Manifolds”, Journal of Dynamical and Control Systems Vol. 7, No. 3,
pp. 385-423, (2001).
[18] A. Bressan and B. Piccoli, “Structural Stability for Time–Optimal Planar Syn-
theses”, Dynamics of Continuous, Discrete and Impulsive Systems, No.3 (1997),
pp. 335-371.
[19] A. Bressan and B. Piccoli, “A Generic Classification of Time Optimal Planar
Stabilizing Feedbacks”, SIAM J. Control and Optimization, Vol.36 No.1 (1998),
pp. 12-32.
[20] P. Brunovsky, “Existence of Regular Syntheses for General Problems,” J. Diff.
Equations 38 (1980), pp. 317-343.
[21] P. Brunovsky, “Every Normal Linear System Has a Regular Time–Optimal Syn-
thesis”, Math. Slovaca 28, 1978, pp. 81-100.
[22] G. Buttazzo, M. Giaquinta, S. Hildebrandt, “One-dimensional variational prob-
lems : an introduction” Oxford University Press, 1998.
[23] L. Cesari, “Optimization-theory and applications: problems with ordinary dif-
ferential equations”. New York : Springer-Verlag, c1983.
[24] Dubins L.E. “On curves of minimal length with a constraint on average curvature
and with prescribed initial and terminal position and tangents”, Amer. J. Math,
V.79 (1957), pp.497-516.
[25] A.T. Fuller, “Relay Control Systems Optimized for Various Performance Crite-
ria”, Automatic and remote control, Proc. first world congress, IFAC Moscow,
vol. 1, Butterworths, 1961, pp. 510-519.
[26] M. Giaquinta, S. Hildebrandt, “Calculus of Variations” Berlin, Springer-Verlag,
(1996).
[27] J.P. Gauthier and G. Bornard. “Controlabilite des sytemes bilineaires”, SIAM J.
Control and Optimization, 20:377-384, 1982.
[28] A. Isidori. “Nonlinear control systems: an introduction”, Springer-Verlag,
(1985).
[29] V. Jurdjevic, “Geometric Control Theory”, Cambridge University Press, (1997).
[30] V. Jurdjevic and I.K. Kupka, “Control Systems subordinated to a group action:
accessibility”, J. Diff. Eq. Vol.39, pp. 186–211.
[31] V. Jurdjevic and I.K. Kupka, “Control Systems on Semisimple Lie Groups and
Their Homogeneous Spaces”, Ann. Inst. Fourier, Vol.31, pp. 151–179.
64 U. B OSCAIN AND B. P ICCOLI

[32] V. Jurdjevic and H.J. Sussmann, “Controllability of Non-Linear systems”, J.

Diff. Eq. Vol.12, pp. 95–116.

[33] M. Kiefer and H. Schattler, “Cut-Loci, and Cusp Singularities in Parameterized

Families of Extremals,” Optimal Control (Gainesville, FL, 1997), pp. 250-277,
Appl. Optim., 15, Kluwer Acad. Publ., Dordrecht, (1998).

[34] M. Kiefer and H. Schattler, “Parametrized Families of Extremals and Singulari-

ties in Solution to the Hamilton Jacobi Bellman Equation” SIAM J. Control and
Opt., 37 (1999), pp. 1346-1371.

[35] V. Kostov and E. Degtiariova-Kostova “Suboptimal paths in the problem of a

planar motion with bounded derivative of the curvature.” Comptes rendus de
l’acad?mie des sciences 321, 1995, pp. 1441–1447.

[36] V. Kostov and E. Degtiariova-Kostova, “Irregularity of Optimal Trajectories in a

Control Problem for a Car-like Robot”, Research Report 3411, INRIA, 1998.

[37] A.J. Krener, “A generalization of Chow’s and the bang bang theorem to non-
linear control problems”. SIAM J. Control Optimization 12 (1974), pp. 43-51.

[38] A.J. Krener,“The high order maximal principle and its application to singular
extremals”, SIAM J. Control Optimization 15 (1977), no. 2, pp. 256–293.

[39] I.A.K. Kupka, “The ubiquity of Fuller’s phenomenon”, in Nonlinear controlla-

bility and optimal control, 313–350, Monogr. Textbooks Pure Appl. Math., 133,
Dekker, New York, 1990.

[40] M.M. Lee, L. Markus, “Foundations of Optimal Control Theory”, Whiley, New
York, 1967.

[41] C. Lobry, “Controllability of non-linear systems on compact manifolds”, SIAM

J. Control Optim., Vol.1, (1974) pp.1–4.

[42] C. Marchal, “Chattering arcs and chattering controls”, J. Optimization Theory

Appl. 11 (1973), 441–468.

[43] A.A. Markov, “Some examples of the solution of a special kind of problem in
greatest and least quantities,” (in Russian) Soobshch. Karkovsk. Mat. Obshch. 1,
1887, pp. 250–276.

[44] R. Montgomery, “A survey of singular curves in sub-Riemannian geometry”,J.

Dynam. Control Systems 1 (1995), no. 1, pp. 49–90.

[45] R. Montgomery, “A Tour of Subriemannian Geometries, Their Geodesics and

Applications”, American Mathematical Society, (2001).
A N I NTRODUCTION TO O PTIMAL C ONTROL 65

[46] B. Piccoli, “Regular Time–Optimal Syntheses for Smooth Planar Systems,”

Rend. Sem Mat. Univ. Padova, Vol.95 (1996), pp. 59-79.

[47] B. Piccoli, “Classifications of Generic Singularities for the Planar Time-Optimal

Synthesis”, SIAM J. Control and Optimization, Vol.34 No.6 (December 1996),
pp. 1914-1946.

[48] B. Piccoli and H.J. Sussmann, “Regular Synthesis and Sufficiency Conditions
for Optimality”, SIAM J. Control and Optimization, Vol. 39 No. 2 pp. 359-410,
(2000).

[49] L.S. Pontryagin, V. Boltianski, R. Gamkrelidze and E. Mitchtchenko, “The

Mathematical Theory of Optimal Processes”, John Wiley and Sons, Inc, 1961.

[50] J.A. Reeds and L.A. Shepp, “Optimal Path for a car that goes both forwards and
backwards”, Pacific J. Math, Vol. 145, pp. 367-393, (1990).

[51] P. Soueres and J.P. Laumond, Jean-Paul, “Shortest paths synthesis for a car-like
robot”, IEEE Trans. Automat. Control 41 (1996), no. 5, pp. 672–688.

[52] H.J. Sussmann, “Envelopes, conjugate points, and optimal bang-bang ex-
tremals”, in Algebraic and geometric methods in nonlinear control theory, 325–
346, Math. Appl., 29, Reidel, Dordrecht, (1986).

[53] H.J. Sussmann, “Envelopes, higher–order optimality conditions and Lie Brack-
ets”, in Proc. 1989 I.E.E.E, Conf. Decision and Control.

[54] S. Lojasiewicz, H.J. Sussmann, “Some examples of reachable sets and optimal
cost functions that fail to be subanalytic”. SIAM J. Control Optim. Vol. 23, no.
4, pp. 584–598 (1985).

[55] H.J. Sussmann, “The structure of time-optimal trajectories for single-input sys-
tems in the plane: the general real analytic case”, SIAM J. Control Optim. 25
No. 4 (1987), pp. 868–904

[56] H.J. Sussmann, “The Structure of Time-Optimal Trajectories for Single-Input

Systems in the Plane: the C ∞ Nonsingular Case,” SIAM J. Control Optim. 25,
No.2 (1987), pp. 433-465.

[57] H.J. Sussmann, “Regular synthesis for time optimal control of single–input real–
analytic systems in the plane,” SIAM J. Control and Opt., 25 (1987), pp.1145-
1162.

[58] H.J. Sussmann, “The Markov-Dubins Problem with Angular Acceleration Con-
trol”, Proceedings of the 36th IEEE, Conference on Decision and Control, San
Diego, CA, Dec. 1997. IEEE Publications, New York, 1997, pp. 2639-2643.
 66 U. B OSCAIN AND B. P ICCOLI

[59] H.J. Sussmann, G. Q. Tang, “Shortest Paths for the Reeds-Shepp Car: A Worked
Out Example of the Use of Geometric Techniques in Nonlinear Optimal Con-
trol” Rutgers Center for Systems and Control Technical Report 91-10, Septem-
ber 1991, to appear in SIAM J. Control.
[60] Y. Sachkov, “Controllability of Invariant Systems on Lie Groups and Homoge-
neous Spaces,” J. Math. Sci., Vol.100, n.4, pp.2355-2427, (2000).
[61] M.I. Zelikin and V.F. Borisov, “Theory of chattering control. With applications to
astronautics, robotics, economics, and engineering”, Systems & Control: Foun-
dations & Applications, Birkhäuser, Boston, 1994.

U. B. B. P.
SISSA, IAC-CNR,
via Beirut 2-4 viale Del Policlinico 137,
34014 Trieste, Italy 00161 Roma, Italy
[email protected] [email protected]

View publication stats