IEEE Robotics & Automation Magazine - March 2022 - 100

●
In-situ testing: where R is assessed
against the actual working of the system
in a real-world environment [3].
Again, there are many variations and
options here ranging from the " certain "
(in the case of formal proof) to the stochastic.
The TC aims to encompass
work across all these areas as well as to
link and support aspects such as transparency
[11] and modularity [6], [9].
This TC is
concerned with
the development
of tools and
techniques to
verify autonomous
systems even in
such unconstrained
and unstructured
environments.
What Is " Autonomy " ?
Essentially, autonomy is the ability (and
often requirement) of a system to make
its own decisions and take its own
actions. This TC takes an inclusive view
on system autonomy, covering full
autonomy, where decision making and
action is fully
within the system's
software
(and so assessment
of why
decisions are
made becomes
crucial); adaptive
systems,
where decision
making and ac -
tion are driven
by (often continuous)
interactions
with the
environment;
and automated
systems, where
decision making
and action are prescripted, and so
on. Which form of decision making is
utilized will also have a strong impact
on the effectiveness of any of the
ve rification techniques to which it can
be applied.
Why Is This Important?
As the range of systems that are
expected to act on their own expands,
the need for verification of these
autonomous systems becomes more
important. When there is a " human in
the loop, " i.e., a human providing oversight
and control of a system, the key
decisions about the system can be delegated
to that human, leaving the system
analysis to address issues such as
reliability. However, once there is a
need for the system to make key decisions,
much more evidence and confidence
in these types of systems will be
required. Developing the ability to
establish and provide this evidence is
then essential not only for engineers
but also for all stakeholders such as regulators,
the public, and governments,
and so for this TC.
If the abilities of systems and the
environments in which they are to
work remain constrained, then realistic
boundaries for system behaviors can be
provided. However, once autonomous
systems are deployed in hard to predict
or unknown environments and we
expect them to make key, and sometimes
(safety, mission, security) critical
decisions, then a much stronger analysis
is required. In addition, what
requirements might be assessed depend
crucially on what is known of the system
and its environment. Traditionally,
it has been assumed that we can assess
(before deployment) all potential
issues/concerns and mitigate against
these, which might be the case in
highly controlled, closed environments.
However, with autonomous systems
increasingly being used in open,
uncontrolled environments and with
internal, software behavior able to
change in various ways, our ability to
predict " everything that might go
wrong " is severely limited. Furthermore,
stochastic models of complex,
unknown environments can never be
complete and may have hard to predict
errors. Therefore, this TC is concerned
https://www.ieee-ras.org/verificationof-autonomous-systems https://www.ieee-ras.org/verificationof-autonomous-systems
IEEE Robotics & Automation Magazine - March 2022

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