Signal Processing - May 2016 - 9

brain, or as the concerted activity of
neuronal populations of different sizes,
depending on the position of the electrodes-either implanted in the brain,
on the surface of the brain or outside
the scalp." He notes that a combination
of these approaches may be necessary
to achieve the ultimate goal of controlling a neuroprosthesis as easily and
precisely as able-bodied people control
their natural limbs.
In tests involving nine disabled and
ten healthy people in Italy, Germany,
and Switzerland, participants wore an
electrode-studded hat that detected their
brain signals. The individuals then
instructed a robot to move in various
ways. "Each of the nine subjects with
disabilities managed to remotely control the robot with ease after fewer than
ten days of training," Millán says. The
tests ultimately revealed no difference
in piloting ability between healthy and
disabled participants.
The researchers also believe that at
least some degree of robot autonomy
should be available to supplement and
complement user control. A robotic
mobility device, for example, should be
able to avoid obstacles by itself, even
when it is not told to, Millán notes. To
avoid becoming overly tired, the user
should be able to take a break from giving instructions, allowing the robot to
continue on its current path until it
receives an order to stop or change
course. "In this way, control over the
robot is shared between the human and
the computer, allowing the pilot to rest
while navigating," Millán says.
Signal processing is a critical component in several areas of BCI research.
"The first is to increase the signal-tonoise ratio of the recorded signals, which
have very low amplitude on the order of
microvolts," Millán says. "Typical tools
at this stage are filtering in the frequency
and spatial domain, as we record from
many electrodes at a relatively high sampling rate." Signal processing also helps
to extract features that may reflect different neuronal processes, each associated
to an aspect of the user's intent. "For
instance, imagination and execution of a
movement gives rise to rhythmic activity
in different frequencies in the sensorimo-

Figure 1. A robotic wheelchair equipped with a BCI that records neural signals and decodes them in
order to transform user's intentions into appropriate commands. (Photo courtesy of Ecole Polytechnique Fédérale de Lausanne's Defitech Foundation Chair in Brain-Machine Interfaces.)

tor cortex of the corresponding body
part," Millán explains. "Out of many
potential candidates, machine-learning
techniques select those features that
improve decoding."
Millán says that one of the biggest
signal processing-related challenges that
researchers face is the intrinsic variability of brain signals, which makes it difficult to decode a user's intentions from
within a short time window. "We deal
with this issue by using statistical
approaches that combine evidence accumulated over time to robustify the final
decision," he says.
Millán is confident that the BCI will
ultimately be used by robotic mobility
system manufacturers worldwide, but
not for at least several more years. "It
will require large trials to demonstrate
the robustness and reliability of the technology," he says. "This is not only timeconsuming, but will also require
substantial financial resources to cover all
the necessary personnel to run the trials."

Finger control
Researchers at the Georgia Institute of
Technology (Georgia Tech) believe that
businesses and consumers will soon
have the ability to control entire fleets
IEEE SIgnal ProcESSIng MagazInE

May 2016

of robots with just the flick of a finger.
Their new tablet-based system is
designed to be used by almost anyone,
including people with no technical
training. To make a swarm of robots do
his or her bidding, the user simply taps
the tablet display to control where a
beam of red light appears on a floor.
The robots then will then move toward
the illuminated area, constantly communicating with each other and deciding how to evenly cover the lit space.
If the user swipes a finger across the
tablet display to drag the light across the
floor, the robots will obediently follow.
If the operator places two fingers in different locations on the tablet, the robots
will begin splitting into teams and repeat
the even-covering process within the
two specified areas. "Basically the
robots move around so as to balance
how much light they are responsible for
in the sense that they should end up with
the same amount of light in the different
areas of responsibility," says Magnus
Egerstedt, Schlumberger Professor in
Georgia Tech's School of Electrical and
Computer Engineering (Figure 2).
"A few years ago, we ran user studies
in the lab and found that people were
generally quite bad at controlling large
9

Table of Contents for the Digital Edition of Signal Processing - May 2016