|
Summary
Basically, what I tried to do in my Ph.D. is to take inspiration from
the way we, humans, do to move our limbs around to grasp objects and
manipulate them, to control the limbs of a humanoid robot, to enable him
to do the same thing.
Although, we don't usually think about it, it is quite amazing how good we are
at finding our way through space, dealing with our bodies, with objects and
manipulating them. Because I am very clumsy, I am maybe more aware of it than
others, which probably lead me to this research topic.
My hypothesis is that our dexterity is made possible by some
"representation" of the space surrounding us. Or, if you don't like the word
representation, our perception of space is structured in some way, which is
intrinsically linked to our motor abilities. That, which structures our
representation of space, has been called the body schema. The body schema has
long been the topic of debates and investigations among neurophysiologists,
psychologists and phenomenologists and its definition is still disputed. In
the first part of my thesis, I suggest a computational model of some
aspects of the body schema suitable for a humanoid robot.
In the second part of my thesis, I suggest a reaching controller based on
what we know about human control of reaching movements. Of course the
human arms are very different from robotic arms, so similarity can only
be present at the level of underlying principles, and this is what I tried
to do. The underlying principles I tried to implement in the robot controller
are multi-modal control on one hand and dynamical system control on the other
hand.
In the third part of my thesis, I tried to tackle somewhat more complex
behavior, building on my reaching controller. I try to give the robot the
ability to perform very simple manipulation behavior.
Movies
Here are a few movies illustrating my work.
|
Learning the body schema
A simulation showing the learning of the body schema. Starting from
scratch, the simulated robot learns its shape (light gray) and motor rotation axes (dark gray) online by
looking at its hands and feet.
The algorithm used is described in
this paper .
|
|
avi file (3 Mb)
|
Learning the head and eye rotation axes
The iCub robot learns how to visually track a target by learning the rotation axes of the motor in its head. This is done by analyzing the visual flow resulting from its head and eye motions as
described in my thesis.
|
|
avi file (7 Mb)
|
|
Adaptive reaching movements
The iCub robot reaching a visually tracked target using the multi-referential dynamical system control, while learning its body schema at the same time. First the reaching in imprecise because the robot is badly calibrated. Then it becomes more precise as the robot learns its body schema.
.
| |
avi file (10 Mb)
|
Adaptive reaching movements
The iCub robot reaching a visually tracked target using the multi-referential dynamical system control, while learning its body schema at the same time. The robot can adapt from reaching with one finger to reaching with another finger if the color marker indicating the reaching finger is displaced.
.
|
|
avi file (13 Mb)
|
|
The control of reaching movements
The Hoap2 robot reaching a visually observed target using the multi-referential dynamical system control described in
this paper .
|
|
avi file (10.6 Mb)
|
|
Learning to put an object into a box
The Hoap3 robot learning to put an object into a box in a way robust to initial conditions and online perturbation as described in this paper.
.
|
|
avi file (13 Mb)
|
Grasping a chess piece
The Hoap3 robot reproducing the chess task (grabbing a chess piece) described in this paper.
.
|
|
mpg file (14 Mb)
|
Bibliography
The .bib file I used for my thesis is available here and I have an electronic version of most of those papers in this big file [1GB] .
|
