IEEE Robotics & Automation Magazine - December 2022 - 144

The first strategy has already been mentioned,
namely, that safety, reliability,
and efficiency requirements are facilitated
by considering isolated problems.
A second strategy is to have safety
layers and fallback solutions around the
ML algorithms. As Tenorth says, " Our
vertically integrated system gives us
some freedom in building fallback solutions.
For example, we have a dedicated
safety controller that allows us to navigate
safely next to humans without having
the whole software stack safety
certified. " Implementing fallback solutions
is essential, but they have diminishing
returns. According to Abbeel, " If
you need to rely on those fallback
options too often, then it becomes too
costly. So you need to have high reliability
such that your fallback is triggered
only very sporadically. And that way
you create real value. "
These diminishing returns lead to a
long tail of errors and failures that need
to be addressed, which is costly and time
consuming. As Tenorth shares: " I was
very impressed by how much is possible
if everyone really has a common goal.
For instance, I was really surprised how
fast we had some pick-and-place demos
running after a few months. But what
you really underestimate is how long the
time is to get it reliable. " " As Moritz, I
was initially surprised by how much we
could get done with a relatively small
amount of money. But then as we grew
and engaged with more customers, I was
also surprised by how much extra engineering
we had to do to keep them all
happy, " says Konidaris.
The final solution is to consider an
entirely new approach entirely for a
robotic system or one of its components.
Should the system be evolved to
become more reliable through engineering,
or rather, should time and
money be invested in trying out revolutionary
ML methods?
As an example, the success of simultaneous
localization and mapping
Postures and Movement Patterns
of the Human Hand
A blueprint of hand behavior
Comprehensive and practical
(SLAM) hinged on the pervasive use
of probabilistic representations, revolutionary
at the time. On the other
hand, " When SLAM was considered
'solved' in academia, it was probably
5-10 years before you could even
think about deploying it in industry
because it needed to be much more
reliable and work off the shelf, "
explains Konidaris. Whether to revolutionize
or evolve to create value is
one of the most challenging decisions
to make, in ML and beyond.
Conclusion
From the interviews presented in this
article, we learned that some of the key
obstacles for deploying ML methods in
robotics are the following: the unpredictability
of the engineering effort
required to address the long tail of
potential errors; the uncertainty about
whether, after all this effort has been
made, the end result actually creates
value for customers, or still relies heavily
on suboptimal fallback solutions; and
the fact that these two questions require
both deep technical expertise and a
" hard-nosed industry " perspective.
Academic robotics researchers considering
a spin-off using ML methods
ideally master not only the technical
aspects of their field but also have the
perseverance for dealing with long
error tails and the right intuition about
IEEE Robotics & Automation Magazine - December 2022

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - December 2022
