Nottale, Matthieu (2008) Ancrage d'un lexique partagé entre robots autonomes dans un environnement non-contraint. PhD thesis Robotique Cognitive , Laboratoire Electronique et Informatique (ENSTA - UEI) p.115.
Full text available as:
|
|
Alternative Locations: http://cogrob.ensta.fr/papers/nottale_these.pdf
Abstract
The Talking Heads is a developmental robotics experiment conducted by Luc Steels and Frederic Kaplan in 1995, at Sony CSL lab. In this experiment, agents acquire a shared grounded lexicon of words associated with visual properties of objects in their environnement through language game interactions. The grounding of symbols in the agent's perception is ensured by the way the interactions are designed, thus providing an answer to the grounding problem formulated by Brooks and others, illustrated by the Chinese room experiment.
The Talking Heads are also capable of building a shared lexicon with a population of thousands of agents in an unsupervised way. They implement a model of how a language could emerge in a population of agents capable of communication.
This experiment has been extended in multiple areas : the lexicon has been extended with a grammar, and the static environnement has been modified to include scenes with moving objects. Yet the limited environnement is now limiting some interesting extensions of these experiments.
The aim of this thesis is to reimplement the Talking Heads using autonomous robots and using an unconstrained environment: our laboratory. This would first show that the Talking Heads models are able to scale to a more complex environment, and allow the development of future experiments with more complex cognitive models giving more autonomy to the agents and that would require this richer environnement.
We first focus on the new problems introduced by mobile robots: the detection of an other robot, the problem of positionning a robot, and the problem of pointing a region of the environnement to another robot.
We then explore a first model using image segmentation algorithms to stay very close to the original Talking Heads, and show why this model is failing to obtain lexicon convergence between the agents.
We finally introduce a second model using the state of the art in the object detection field to build object models based on the detection of recurring patterns in the environnement. We first show that this model can successfully be used for classification tasks on object databases, then apply it to the Talking Heads setup. Our results show that the agents are able to exchange symbols associated with regions of their environnement, although the overall game success rate stays low.
| Item Type: | PhD Thesis (PhD) |
|---|---|
| PhD Supervisor: | Ganascia, Jean-Gabriel |
| Date: | 2008 |
| Board of examiners: | Steels, Luc and Baillie, Jean-Christophe and Sigaud, Olivier and Oudeyer, Pierre-Yves and Gaussier, Philippe and Dominey, Peter Ford |
| Ecole Doctorale: | ED 130 INFORMATIQUE, TELECOMMUNICATIONS ET ELECTRONIQUE (EDITE) |
| Discipline: | Robotique Cognitive |
| Collection (Fonds): | ENSTA ParisTech |
| Institution: | UPMC |
| Department: | Laboratoire Electronique et Informatique (ENSTA - UEI) |
| Subjects: | 2. Information and Communication Sciences and Technologies |
| Uncontrolled Keywords: | Robotique développementale, Émergence du langage, Vision artificielle, Ancrage, Symbol grounding, Developmental robotics, Computer vision |
| ID Code: | 4260 |
| Deposited By: | Sophie Chouaf |
| Deposited On: | 21 November 2008 |
Table of content
1 Introduction - 4
2 Les Talking Heads - 7
2.1 Présentation générale - 7
2.2 Les agents et leur environnement - 9
2.3 Perception - 10
2.4 Signifiant, arbres de discrimination - 11
2.5 Lexique - 12
2.6 Initialisation : le jeu de discrimination, le jeu de nommage - 12
2.7 Le Guessing Game - 14
2.8 Résultats, extensions et travaux connexes - 16
3 Les Talking Robots - 18
3.1 Cadre - 18
3.2 Objectifs - 19
3.2.1 Vers des agents mobiles, autonomes et incarnés - 19
3.2.2 Vers un environnement peu restreint - 20
3.3 Déroulement - 20
4 Talking Robots 1.0, segmentation a priori - 22
4.1 Présentation générale - 22
4.2 Mécanismes auxiliaires - 22
4.2.1 Synchronisation - 22
4.2.2 Synthèse et reconnaissance vocale - 23
4.2.3 Pointage - 23
4.2.3.1 Présentation du problème - 23
4.2.3.2 Premier algorithme : modélisation et agrégation - 25
4.2.3.3 Second algorithme : vitesse - 26
4.2.3.4 Performances et jeu de pointage - 27
4.2.4 Localisation des aibos - 29
4.2.4.1 SPOMF et FMI-SPOMF - 29
4.2.4.2 Marquage des aibos - 41
4.3 M´ecanismes fondamentaux - 43
4.3.1 Référents - 43
4.3.2 Signifiants et canaux perceptifs - 50
4.3.3 Arbres de discrimination - 51
4.3.4 Symbole et combinaison de nœuds - 51
4.3.5 Mots, lexique - 52
4.4 Jeu de discrimination - 52
4.5 Guessing Game - 53
4.6 Implantation - 54
4.7 Résultats - 54
4.8 Analyse - 59
5 Talking Robots 2.0, modèle d’objet - 61
5.1 Présentation - 61
5.2 La reconnaissance d’objets - 63
5.3 Algorithmes de détection de points caractéristiques - 66
5.3.1 Histogramme et distance EMD - 67
5.3.2 Corrélogramme - 67
5.3.3 SIFT - 67
5.3.4 K-Adjacent Segment - 68
5.4 Graphes de discrimination - 68
5.4.1 Premier dictionnaire : taille de rayon fixe - 69
5.4.2 Second dictionnaire : nombre de fils par nœud fixe - 71
5.5 Les objets - 73
5.5.1 Définition - 73
5.5.2 Comparaison - 74
5.5.2.1 Normes L1 et L2, TF-IDF - 74
5.5.2.2 Calcul probabiliste - 74
5.5.2.3 Vote - 75
5.5.2.4 Pondération entropique - 76
5.5.3 Reconnaissance, construction non supervisée - 76
5.5.4 Agrégation de multiples détecteurs de caractéristiques - 77
5.5.5 Fenêtrage - 78
5.6 Mots et lexique - 79
5.7 Initialisation : jeu de discrimination - 80
5.8 Guessing Game - 80
5.9 Implantation - 81
5.10 Résultats - 81
5.10.1 Paramétrage - 82
5.10.2 Catégorisation sur la base d’images GRAZ-02 - 83
5.10.3 Persistance des modèles d’objets - 85
5.10.4 Apprentissage supervisé d’associations entre sons et
objets - 87
5.10.5 Limites du modèle : suppression de fond - 89
5.10.6 Guessing Game - 90
5.11 Différence avec les Talking Heads - 91
5.12 Futurs développements - 92
6 Conclusions - 95
A Génération de graphes de comportement en C++ - 97
Repository Staff Only: edit this item