Page 166 - Handbook of Biomechatronics

Biomechatronic Applications of Brain-Computer Interfaces 163

response to erroneous performance of a robotic arm (Kreilinger et al., 2012),
a mobile robot (Chavarriaga and Milla ´n, 2010), or a virtual avatar (Pavone
et al., 2016) as well as in response to erroneous predictions made by a sim-
ulated intelligent car (Zhang et al., 2013), suggesting many potential appli-
cations in biomechatronics, for example, identifying when a robotic arm
prosthesis has performed an undesired action or identifying when an
in-car navigation system has provided the wrong directions to the driver.
However, it is still not clear how to respond to ERPs in real-world situations
where errors may have multiple possible causes and many possible corrective
actions can be taken.
One issue with error-driven learning is that, while ERPs have the poten-
tial to detect errors in machine behavior, the ERPs themselves may also be
misclassified, for example, a correct BCI action may be misinterpreted as an
error. In such cases, error-driven learning will actually increase the proba-
bility of future errors by incorrectly retraining the BCI. One possible way to
address this would be through probabilistic classifiers: the BCI calculates the
probability that an ERP (or lack of ERP) has been detected, and only
retrains its algorithms based on this new data if it is sufficiently certain
(e.g., above 90%) that it is correct. Such methods have been proposed in
the literature (Artusi et al., 2011; Llera et al., 2012), but have primarily been
tested with simulated BCIs where prerecorded data are used as a stand-in for
actual signal acquisition from a user. Thus, further testing of this approach is
needed in natural settings with actual users.
To summarize, BCIs are most commonly used for control of assistive and
rehabilitation devices by people with disabilities (e.g., wheelchairs, spellers,
prostheses), but can also monitor users’ brain activities in a passive fashion
and use this information to adapt a mechatronic device—by changing the
amount of assistance provided, by changing the difficulty of a task, or by
responding to potential errors. Particularly, assistive devices have already been
shown to be quite effective, and extensive work is being done to improve the
performance and acceptance of BCIs in many biomechatronic applications.
However, several challenges do remain, as discussed in the next section.

3 CHALLENGES AND OUTLOOK

In the previous sections, we briefly alluded to some of the challenges
facing BCIs in biomechatronic systems. In the next few sections, we will
explicitly discuss some of these challenges as well as promising avenues
for future research.