Imitation learning of motor skills for synthetic humanoids [Elektronische Ressource] / von Heni Ben Amor

Heni Ben Amor

Sujets

• Informatik

Informations

• Publié par: technische_universitat_bergakademie_freiberg
• Publié le: 01 janvier 2010
• Nombre de lectures: 75
• Langue: Deutsch
• Poids de l'ouvrage: 18 Mo

Extrait

Imitation Learning of Motor Skills for
Synthetic Humanoids
Von der Fakultät für Mathematik und Informatik
der Technischen Universität Bergakademie Freiberg
genehmigte
DISSERTATION
zur Erlangung des akademischen Grades
Doktor-Ingenieur
Dr.-Ing.
vorgelegt
von Dipl.-Inform. Heni Ben Amor
geboren am 01.02.1982 in Dernbach
Gutachter:
Prof. Dr.-Ing. Bernhard Jung, Freiberg
Prof. Dr. rer. nat. Ulrich Furbach, Koblenz
Tag der Verleihung: 12.11.2010Versicherung
Hiermit versichere ich, daß ich die vorliegende Arbeit ohne unzulässige Hilfe Dritter
und ohne Benutzung anderer als der angegebenen Hilfsmittel angefertigt habe; die aus
fremden Quellen direkt oder indirekt übernommenen Gedanken sind als solche kenntlich
gemacht.
BeiderAuswahlundAuswertungdesMaterialssowiebeiderHerstellungdesManuskripts
habe ich keine Unterstützungsleistungen Dritter, wie Promotionsberater, in Anspruch
genommen. Weitere Personen haben von mir keine geldwerten Leistungen für Arbeiten
erhalten, die nicht als solche kenntlich gemacht worden sind.
Die Arbeit wurde bisher weder im Inland noch im Ausland in gleicher oder ähnlicher
Form einer anderen Prüfungsbehörde vorgelegt.
Freiberg, den 01. Juni 2010
Heni Ben Amor
iiiAcknowledgements
You can dream, create, design and build
the most wonderful place in the world, but
it requires people to make the dream a reality.
(Walt Disney)
Although the focus of this thesis are virtual and robotic humanoids, it is the interac-
tion with humans of flesh and blood that played the most important part in its comple-
tion. During the last four years, many people at the Technical University Bergakademie
Freiberg and the University of Osaka supported me in the work leading to this thesis.
First, I would like to thank my advisor Bernhard Jung for giving me the opportunity
to work on exciting scientific projects, and for guiding and educating me during my
years in Freiberg. Bernhard is a very inspiring person who always managed to create
a motivating atmosphere in our working group. Thank you Bernhard for all the ideas,
your help and patience! I would also like to thank my coworkers at the virtual reality
group Guido Heumer, Arnd Vitzthum and Matthias Weber for the stimulating scientific
(and sometimes non-scientific) discussions and their great support.
I am grateful to Ulrich Furbach for initiating me to the world of artificial intelligence
and for being available as a reviewer of this thesis.
I am indebted to Hiroshi Ishiguro for making me part of the Intelligent Robotics Lab
at the University of Osaka in Japan. Ishiguro-sensei was a great mentor and constantly
supported me in my research on android robots.
My special thanks go to Konrad Froitzheim for the numerous hints onhow to improve
the thesis andforhis support inacquiring thesmall humanoid robotsthat were essential
for carrying out the experiments presented here.
IwouldliketothankmyclosecollaboratorShuheiIkemotoforhisconstanthelpaswell
ashissupportduringmystaysinJapan. Althoughattimeswewereseperatedbyseveral
thousands of kilometers, we always managed to continue our successful collaboration. I
had a great time working with Shuhei and I will always remember the funny discussions
at 3:00 a.m. in the morning on the way home from the lab.
IthankOliverObstforintroducingmetorobotics,machinelearning,scientificwriting
and many other things that were important for my later career.
I am deeply thankful to all of my students who supported and helped me in vari-
ous ways, especially Erik Berger, David Vogt, Henry Lehmann, Edith Kegel, Christian
Schlegel, Ralf Müller, Maik Deininger, Stefanie Höfig, Peter Scheicher and Eric Kunze.
iiiWithout the delicious couscous of Moncef Bouaziz, his constant support, as well as
his great humor, the writing of this thesis would have taken a significantly longer time.
A big Thank You! also goes to Martin Wiesenmayer for proofreading this thesis and
for keeping me up-to-date about recent cinematic activities.
I thank my sisters Hazar and Houyem for being an endless source of inspiration and
for cheering me up during the hard times of solitary writing and experimentation.
Finally, I want to thank my brother Heikel, my sister Hounaida and my parents Hedi
and Mesaouda Ben Amor for their moral support and for always being there when I
need them.
Freiberg, 2010
ivContents
1. Introduction 1
1.1. Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2. Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.3. Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.4. Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5. Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.6. Overview of the Main Chapters . . . . . . . . . . . . . . . . . . . . . . . 7
2. Related Work on Imitation Learning 9
2.1. Imitation in Humans and Animals . . . . . . . . . . . . . . . . . . . . . . 9
2.2. Programming by Demonstration . . . . . . . . . . . . . . . . . . . . . . . 11
2.3. Computer Animation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15
2.4. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
3. Mathematical Foundations 21
3.1. Articulated Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1.1. Kinematic Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
3.1.2. Rotation Representation . . . . . . . . . . . . . . . . . . . . . . . 23
3.2. Dimensionality Reduction . . . . . . . . . . . . . . . . . . . . . . . . . . 24
3.2.1. Basic Concepts . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25
3.2.2. Dimensionality Reduction Methods . . . . . . . . . . . . . . . . . 26
3.2.3. Comparison of Methods . . . . . . . . . . . . . . . . . . . . . . . 38
3.2.4. Estimation of the Intrinsic Dimensionality . . . . . . . . . . . . . 42
3.3. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4. An Imitation Learning Approach:
Probabilistic Low-Dimensional Posture Models 45
4.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.2. The Three Steps in Imitation . . . . . . . . . . . . . . . . . . . . . . . . 47
4.3. Probabilistic Low-Dimensional Posture Models . . . . . . . . . . . . . . . 49
4.4. Example: Pressing a Button . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.4.1. Motion Recording . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
4.4.2. Model Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.4.3. Posture Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 60
vContents
4.5. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5. Learning to Imitate Natural Human Grasping 71
5.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71
5.1.1. Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.2. Modeling the Human Hand . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.2.1. Grasp Taxonomies . . . . . . . . . . . . . . . . . . . . . . . . . . 74
5.2.2. The Kinematic Model . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2.3. The Sensor Model . . . . . . . . . . . . . . . . . . . . . . . . . . . 77
5.2.4. Grasp Parametrization . . . . . . . . . . . . . . . . . . . . . . . . 78
5.3. Grasp Quality Measure . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
5.3.1. Finger Distance . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81
5.3.2. Anatomical Plausibility . . . . . . . . . . . . . . . . . . . . . . . . 82
5.3.3. Stability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
5.4. Grasp Synthesis Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.4.1. Initialization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
5.4.2. Optimization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
5.4.3. Computational Speedups . . . . . . . . . . . . . . . . . . . . . . . 91
5.5. Evaluation and Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
5.5.1. E1: Data Acquisition and Analysis . . . . . . . . . . . . . . . . . 93
5.5.2. E2: Simple optimization setting . . . . . . . . . . . . . . . . . . . 95
5.5.3. E3: Optimization with different Rotation Representations . . . . 97
5.5.4. E4: Optimization using different DR techniques . . . . . . . . . . 98
5.6. Other Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
5.7. Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102
6. Learning to Imitate and Adapt Full-Body Motions 105
6.1. Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
6.2. Imitation from Motion Capture Data . . . . . . . . . . . . . . . . . . . . 106
6.2.1. Kinematic Modeling of Virtual Humans . . . . . . . . . . . . . . . 107
6.2.2. Motion Recording . . . . . . . . . . . . . . . . . . . . . . . . . . . 108
6.2.3. Motion Synthesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 115
6.3. Programming Robots by Demonstration . . . . . . . . . . . . . . . . . . 119
6.3.1. Kinesthetic Bootstrapping . . . . . . . . . . . . . . . . . . . . . . 119