The Wayback Machine - https://web.archive.org/web/20050113151122/http://www.etca.fr/CTA/gip/Publis/Luzeaux/Index.html
Dominique Luzeaux
Dominique Luzeaux was born in Paris, France, in 1965.
He received the Dipl.
Ing. degree from the Ecole Polytechnique, Palaiseau, France, in 1987,
the Dipl. Ing. from the Ecole Nationale des Techniques Avancées (Ensta),
Paris, in 1989, and the Doctorat d'Université (Ph.D.) degree from the
University Paris XI, Orsay, in 1991.
He held a visiting position at the University of California, Berkeley from
1991 to 1992 in the EECS department. He is currently technical Director of the
of the Centre Technique d'Arcueil and held until 2000 the positions of head of
the Department Géographie-Imagerie-Perception and of the lab
Perception for Robotics at the Centre Technique d'Arcueil
(DGA/DCE/CTA).
His research interests include intelligent control, mobile robots,
artificial intelligence, grammars and chaos theory.
Hobbies: foreign languages, karate, car driving, violin playing, hard rock
and great food tasting.
Mpeg Movies
You can have a glance at some of our robotics activities:
We discuss the research developed in our lab these past years,
leading to autonomous robots evolving in non-cooperative, even hostile,
outdoor environments.
Our research deals with the design and experiment of a control architecture
for an autonomous outdoor mobile robot which uses mainly vision for perception.
In the case of a single robot, we have designed a hybrid architecture with an
attention mechanism that allows dynamic selection of perception processes.
Building on this work, we have developed an open multi-agent architecture,
for standard multi-task operating system, using the C++ programming
language and Posix threads.
Our implementation features efficient and fully generic messages between
agents, automatic acknowledgement receipts and built-in synchronization
capabilities.
Knowledge is distributed among robots according to a collaborative scheme :
every robot builds its own representation of the world and shares it with
others. Pieces of information are exchanged when decisions have to be made.
Experiments are to be led with two outdoor ActiveMedia Pioneer AT mobile robots.
Distributed perception, using mainly vision but also ultrasound, will serve as p
roof of concept.
Building on our previous research, which deals with learning autonomous
robots evolving in unknown environments, we investigate various categorization
issues and discuss how models drawn from catastrophe theory may help
to solve the symbol grounding problem for such robots.
Dans le cadre de la logique mathématique, nous
cherchons à voir les liens entre l'interprétation et les
multiples véeritées auxquelles elle donne accès, les
rapports qu'elle établit entre le signe et le référent,
de quelle manière le signe peut guider l'interprétation, et enfin
la relativité de l'interprétation dans son contexte le plus
général de l'établissement d'assertions
mathématiques (localité universelle des topoi).
Our research on the modelization of the loop action-perception for an
autonomous robot have led us naturally towards a tree-like representation
of that interaction, which allows to state rigorously the problems of
viable control in terms of games.
Furthermore it is necessary to describe these tree sets as simply as possible,
in order to decide for the existence of a controller, and in the best
case to synthetize such a controller. The goal of this paper is to show
how this approach is closely related to hybrid systems: we study the most
representative class of decidable hybrid system, the timed automata, from the
point of view of regular tree grammars.
In this paper, we are interested in the design and the experiment
of a control architecture for an autonomous outdoor mobile robot which
mainly uses vision. We focus on the design of a mechanism that permits
the dynamic selection and firing of perception processes. We propose an
hybrid architecture that uses an attention mechanism which controls the
robot environment awareness while managing the computational resources
and allowing a fair reactivity. We describe its implementation and
experimentation on a robot in an outdoor environment.
After a brief review of the various control architectures used in mobile
robotics, we present a hybrid architecture whose core is an attention manager.
Then we illustrate that architecture with experiments performed on a
real-world robot and conclude by showing how our design of control laws
solves partially the symbol grounding problem.
Starting from a system theory of digitally controlled systems, we
define computable control laws. After recalling classical undecidability
results, we focus on the notion of ``cell decomposition'', defined as an
oracle partitioning the space of input-output trajectories and show how
this notion solves the previous undecidability. The key issue is that the
cell decomposition induces a natural hybrid system model of the initial
system.
Learning robots are faced with two major issues: identification of the
dynamics of the robot and identification of the environment as well as
its interaction with the robot.
We discuss in this paper a way to acquire representations
of both these concepts through an iterative learning procedure. Furthermore we
will concentrate on qualitative representations, since we are not
necessarily interested in the precise equations of the dynamics or the exact
location of potential limits of the viability domain.
The key notion of our representations is the phase space: it will be
used to express the dynamics of the controlled robot, as well as the landmark
values, corresponding to the environment, that refer either to
control goals or obstacles to be avoided. The important issue is that the
exposed learning procedure provides a way to acquire these various knowledges.
Nous allons principalement nous intéresser dans cet article aux relations
entre modèle et langage, en tâchant de cerner
précisément ce qu'est un
modèle, comment il est exprimé (d'où le rôle
prépondérant que va jouer le langage sous-jacent), et en quoi le
langage d'expression va limiter sa portée. L'étude sommaire de
cette problématique permettra ensuite de répondre aux diverses
questions que l'on peut se poser, quant aux relations que l'on peut entretenir
avec les modèles. Nous développerons le point de vue de la
logique mathématique, en essayant de ne pas rentrer sur un plan
technique, puis nous nous tournerons vers le monde de la physique
théorique, ou plus exactement vers le monde de la théorie de la
physique.
Nous montrons comment une approche entrées/sorties en théorie des
systèmes peut se révéler utile pour évaluer la
complexité de certains problèmes
relatifs à la classe des systèmes hybrides. La méthode
exposée est appliquée en particulier à une sous-classe
des automates temporisés.
In this paper, we discuss two major issues in the modeling of digital complex
systems: the first one deals with the strutural aspects, while the second deals
with the transformational aspects. In order to express these aspects with a
convenient mathematical framework, we will use category theory and will
describe the general models while keeping the discussion at a rather low
technical level.
In this paper, we discuss two major issues in the modeling of digital complex
systems: the first deals with the structural aspects for systems, while the
second deals with the transformational aspects. In order to express these
aspects with a convenient mathematical framework, we will use category theory
and will describe the general models while keeping the discussion at a rather
low technical level.
Dans cet article, nous nous intéressons à l'intelligence dans les
systèmes d'armes, et notamment au traitement d'images et à la
fusion de donnéees. Après avoir brossé le tableau
des besoins, nous faisons un panorama du vaste domaine du traitement d'images,
puis esquissons les évolutions à venir dans ce domaine.
In this paper we discuss the importance of considering together sensory
and motor information in order to learn to control a robot. We also
discuss of the possibility of incorporating in a low-level control loop
higher-level representations of the environment.
In this paper, we address the problem of tracking a moving target with
a mobile camera. We discuss the whole perception-decision-action process,
including the image processing and the control loop. Then we turn to the
theoretical proof of the stability of our control law in closed loop with
the image processing.
This paper is dedicated to the formal definition of a dynamical system
controlled by a digital device. Starting from general considerations,
we introduce the ``natural'' state space and turn then to its simplification
using mathematical arguments. Finally, we relate this state space to
the common real space and show how the digital system theory proposed
here encompasses the usual control theory.
In this paper, we present first a brief survey of some techniques in regular
grammar inference. The core of the discussion is then to try to classify these
different methods within a unique framework. A first step in that
direction has introduced a general methodology based on
letter-to-letter morphisms. We pursue that work and propose successively
paradigms based on rewriting morphisms, substitutions, and transductions.
We address the problem of learning to control an unknown system evolving in an
unknown world, as a problem of knowledge and metaknowledge acquisition. After
discussing these notions, we illustrate the various issues involved with our
learning program Candide and conclude on a real-world example taken from
mobile robotics.
Let a be a real number strictly greater than 1 and let D be a finite
interval of the set of rationals Z containing 0.
In this paper we give necessary and sufficient conditions that guarantee
the existence, for any real, of an expansion in base a with
coefficients in D. Then we turn to balanced expansions, for which the
sums of the digits of any initial segment are uniformly bounded.
These issues yield necessary and sufficient conditions for
stabilizing stationary linear discrete systems with a particular family
of control laws.
Nous présentons dans ce papier une représentation permettant de
prendre en compte simultanément dans un système artificiel,
comme un robot mobile, un modèle de la dynamique propre du
système et un modèle topologique de l'environnement.
Cette unicité de la représentation permet d'envisager
de manière constructive de réels comportements sensori-moteurs
depuis le simple réflexe jusqu'à des niveaux supérieurs
(navigation, réalisation de buts)
In this paper we discuss a special family of control laws, rule-based
incremental control, and present two approaches (a time-varying state feedback
stabilization and a motion planner) to the control of discrete nonholonomic
systems with an application to a car-like robot in simulation and real-world.
We compare both approaches and see that despite their very different appearance
they induce similar behaviors on a car-like robot.
The aim of this tutorial is to present the necessary interactions between
machine learning and control theory.
First we recall the basic definitions of control theory and machine learning.
After a brief review of the various approaches in machine learning -
distinguishing between supervised and unsupervised learning -, we
discuss the major methods used in intelligent control. Then we expose
another approach based on qualitative physics and rule-based incremental
control.
In this paper we discuss the design of a rule-based incremental controller
in the control of a nonholonomic robot equipped with a CCD
video camera. We address here the problem of a car platoon. Our
experimentation illustrates the paradigm perception-decision-action by
integrating an active vision process which allows for efficiency and
robustness and introduces a natural decoupling of the controls of
the robot. We show how a naive and intuitive translation into rules
of the qualitative observation of the system yields a controller, the
performance of which will be analyzed and justified from a theoretical
point of view in the rest of the paper.
This paper discusses a general approach to learning to control a
system using a black-box approach. It can be dived into: qualitative modeling
based on a grammatical representation of the input-output data, automatic
building of a representation of the viability domain, production of a
controller allowing local control inside regions of that domain, automatic
generation of a plan in order to reach preassigned goals. This methodology is
illustrated by a mobile robot application.
In this paper we discuss a special family of control laws,
rule-based incremental control, and
we address the problem of the parking maneuver for a car, first
in simulation and then with a small mobile robot we have built.
This paper presents our methodology for learning to control a
black-box system. That is, we have no {\em a priori}
knowledge about the system we want to control.
The controller is learnt stage by stage, each stage corresponding
to the construction of a level in a hierarchy of behaviors,
from basic viability behaviors to cognitive goal-reaching
behaviors.
Our approach is based on a hierarchical distinction
between a {\em general purpose reactive control learning algorithm} that builds
{\em basic behaviors} for the agent, and a higher level {\em state space
partitioning algorithm} that provides the lower level with a small
number of {\em local goals} bound to each element of the partition.
The advantages of this approach over reinforcement learning
methods for control are discussed, and the corresponding control
learning program, {\sc Candide},
is exemplified through a simulated car driving experiment.
We discuss in this paper a learning program, {\sc Candide}, which learns to
control a system in order to guarantee its viability. Without a priori
knowledge about that system, the program
observes random initial evolutions and acquires a qualitative
model. Monotonic or derivative relationships between inputs and outputs are
recognized, then a rule-based incremental controller is deduced from this model
after a close study of the phase space. We apply the learning methodology to
several mobile robot applications both in simulation and with real world
experiments.
Given a real number a>1 and a finite interval D of Z containing 0,
we discuss the existence of an expansion of any real into a series of
negative powers of a with coefficients in D.
First we look for intervals D minimal in length;
then we look for balanced expansions, i.e. such that the sums of the digits
of all initial segments of the expansion are uniformly bounded.
1994
Computability in control [125 Kbytes]
3rd International Symposium on
Artificial Intelligence and Mathematics, Fort Lauderdale, FL, USA, January 1994.
We show how the computability assumption in a control law can drastically
influence the controllability of a system. There are for instance systems
that can be controlled (in the usual sense) but not by any computable control
law.
We give here some stabilization and robustness properties of our family of
control laws, and show how these laws deal with the parking of multibody mobile
robots.
This chapter of a book dedicated to grammar inference unifies the main
approaches to regular grammar inference found in literature and extends the
corresponding paradigm.
1993
Steps or stages for incremental learning [175 Kbytes]
AAAI-93 Spring Symposium Series, Symposium on training issues
in incremental learning, Stanford University, CA, USA, March 1993.
During an incremental learning process, data are delivered packetwise.
The learner may or may not propose then a new hypothesis. We investigate here
the relationship between the flow of data and the hypotheses proposed by the
learner. This leads to the notion of learning stages rather than learning
steps. We formulate a result which shows precisely why the concept of stage
is an important one.
Our main concern in learning models is the presence of the environment
in the information provided to any learner; in other words, the experience
the learner acquires is not a stream of data generated ex nihilo by some
superior teacher, but is generated by the constructive interaction
between the learner and its environment.
This paper presents Candide, a program that learns to control a system from
scratch.
Intelligent control, ergodicity and chaos [328 Kbytes]
International Simulation Technology Multiconference, Simtech-Wnn-Fnn93,
San Francisco, USA, November 1993. (nominated paper)
We discuss some properties of the shift map on different spaces and relate
them to control issues.
We discuss several distances on the free monoid and the induced topologies.
We introduce a new distance weighted by a real sequence that has the following
fancy property: the resulting topological structure is complete iff the
generating sequence is not Cauchy.
This report discusses some of our research done at the Robotics Lab at the
University of California at Berkeley. We deal here with the control of mobile
robots and compare the different approaches to that problem. We apply our
theory (rule-based incremental control) to the parking maneuvers of multibody
mobile robots.
This report details another part of the research done at the Robotics Lab at
the University of California at Berkeley. Various control-theoretic properties
of rule-based incremental control, like stability or reference tracking for
linear and non linear systems are discussed.
Autonomous systems are nowadays a challenge to many classical techniques:
they are either ill-modeled or evolve in changing environments which perturbate
them deeply. It is not possible to apply the well-known mathematical tools and
new methods have to be found to control them. These new methods are
usually known as intelligent control and include neural networks, fuzzy control,
qualitative control, path-planning methods and rule-based reasoning.
In this paper we first introduce incremental rule-based control laws and
then we discuss a learning program, called Candide, which allows a
large class of systems to synthesize such control laws in order to perform a
given task.
The following problem is addressed here: learning to control a system. We give
the theoretical framework that can be used to solve convergence issues for
that general problem.
We give some general results concerning incremental control in
order to understand how this type of control works for linear systems.
Then we address more complex problems as can be encountered in
robot motion: controllability of car-like robots and parking maneuvers.
The interesting point in this new type of control is furthermore
its flexibility and relationship with learning, as we will see in the last
section.
We discuss some properties of the expansion of a real number in a base
that can be integer or real. For particular bases, well-known fractal curves
can be obtained.
We show how one can abstract symbolic concepts and then symbolic operators
from quantitative data and operators defined on quantitative data. A lattice
structure can easily be built on the symbolic concepts. We conclude by building
a lattice of these lattices.
Starting from a data base corresponding to the values in time of various
parameters, we show how to extract monotonous and derivative relationships
between these parameters. This yields a qualitative model (same underlying
language as QSIM) of the system that has generated the data.
We introduce natural control, which we use for controlling
physical discrete-time processes. After a rapid presentation of this type of
control, we present some theoretical results, concerning reference
signal following and stability. Then we discuss learning of such control.
Given a finite number of strings, find a regular grammar that "generalizes"
them (grammar inference). One advantage of that algorithm is robustness in
the following sense: given a set S1 of strings {s1,...,sn} and a set S2 of
strings {t1,...,tn} such that the Levenstein distance between si and ti is
under a given bound, there exists a simple transformation between the
regular grammar inferred from S1 and the one inferred from S2. In other
words the inference map has a property similar to continuity.
Dominique Luzeaux was born in Paris, France, in 1965.
He received the Dipl.
Ing. degree from the Ecole Polytechnique, Palaiseau, France, in 1987,
the Dipl. Ing. from the Ecole Nationale des Techniques Avancées (Ensta),
Paris, in 1989, and the Doctorat d'Université (Ph.D.) degree from the
University Paris XI, Orsay, in 1991.
You can have a glance at some of our robotics activities:
Parking maneuver of the car-like mobile robot MORBIDE1 using an external
camera.
Autonomous small robots for military applications
[3532 Kbytes]
SPIE Aerosense'00, Conference on Unmanned Ground Vehicle Technology,
Orlando, FL, USA, April 2000.
