Daily life of hand amputees
can be poor compared to what it was before the amputation. The state of
the art in hand prosthetics, at the time of writing, does not offer more
than 2-3 degrees of freedom and a very coarse control of the force, as
there is no haptic feedback. Patients interface with the prothesis via
surface electromyography (sEMG), recorded using surface electrodes.
Learning how to control the device through many input sEMG channels is a
long and difficult process for most patients, that therefore settles
for limited and very simplified movements (open/close). This contrasts
with recent advances in mechatronics, thanks to which mechanical hands
gifted with many degrees-of-freedom and force control are being built.
There is a need for prosthetic hands able to naturally reproduce a wide
amount of movements and forces, while at the same time requiring a lower
effort in learning how to control hand postures. This goes beyond
mechatronic dexterity: the real challenge is how to provide patients
with a cheap, easy and natural way of controlling the prosthesis.
The goal of this project is to develop a family of algorithms able to
significantly augment the dexterity, and reduce the training time, for
sEMG controlled prosthesis. By testing our findings on a very large
collection of data, this project will pave the way for a new generation
of prosthetic hands. The work will be organized along the following four
themes.
Theme 1: Data Acquisition and Analysis.
The goal of this theme is to develop a reproducible protocol to acquire
large data sets for healthy patients performing certain movements and
amputated patients also making complex movements, while analyzing and
assessing the data as they become avaliable. The data acquisition
includes the acquisition of signal data and the calibration of the
sensors to limit the noise in the data. Relevant clinical data will be
acquired at the same time such as age, gender, height, weight and for
amputated patients also the exact place of the amputation and the time
between amputation and tests performed. The data acquisition and
analysis will proceed in close connection with the other themes.
Theme 2: Augmented Dexterity: Posture Classification. The
objective of this theme is to push the current state of the art in
prosthetic hand posture classification from handling a maximum of 12
postures up to 40-50. We will design and implement state of the art
machine learning algorithms within the multi kernel learning framework,
using the sEMG signals separated instead of concatenated, as it is the
mainstream practice today. We will then proceed to extend the algorithm
so to exploit the intrinsic hierarchical structure of hand postures. The
outcome of this theme will offer patients a much wider dexterity
compared to the current state of the art.
Theme 3: Augmented Dexterity: Natural Control. This
research theme is about pushing the envelope of sEMG control: extending
it to a quasi-perfect prediction of force, by independently modeling
and controlling single degrees-of-motion. The overall aim is then
to augment the dexterity that an sEMG-controlled prosthesis could
potentially achieve mimicking the way a human hand works. Results in
this theme will be periodically benchmarked against those achieved in
Theme 2.
Theme 4: Adaptive Learning. The
goal of this theme is to develop learning algorithms to better interpret
the sEMG signals acquired from the patients, with the ultimate goal of
boosting the learning process necessary for the patient to effectively
use the prosthesis. We will build pre-trained models of various data
postures, on the data acquired in theme 1, and we will adapt these
general models to the needs of individual users as new data will became
available using adaptive online learning methods.
