IEEE Robotics & Automation Magazine - June 2023 - 110

push the field forward. The rest of
this article briefly summarizes these
three themes and concludes with a
discussion of promising directions.
HUMANS AND AI WORKING
TOGETHER
Although many large-scale manufacturing
processes are automated to
some degree, there are always humans
involved at multiple levels. The
humans may be physically interacting
with the system (e.g., loading parts into
a machine) or interacting through a
computer (e.g., defining the production
schedule). Different subject matter
experts (SMEs) have different types of
expertise; some may have a deep technical
knowledge of a specific process,
such as maintenance of a specific
machine, while others may have a
broader understanding of a system.
Humans are expected to be adaptable
and able to react to both small disturbances
(slightly out-of-order product
flow) and large changes (new product
arriving).
Most AI deployed in manufacturing
plants today is relatively fixed;
i.e., it does not have the ability to learn
beyond what it has been trained to do,
nor can it ask questions when it does
not understand what to do. The panel
felt that there was a significant opportunity
to develop AI that can work in
a more integrated fashion with SMEs,
at different levels, perhaps answering
questions posed by the SME and
even posing questions when something
becomes uncertain. While some AI
tools that take input from a human have
been developed for data science applications
(e.g., Google Vision AI), these
capabilities need to be extended to the
manufacturing environment. In this
way, the knowledge held by the SMEs
can be leveraged with the knowledge in
the AI, and no human knowledge will
be left behind. This integration of SME
and AI can also lead to better trust of
the AI by the SME, so that the AI is
seen as a tool to help get the job done
well-empowering the SME rather
than threatening to eliminate the job.
Many currently deployed AI systems
in manufacturing detect anomalies that
11 IEEE ROBOTICS & AUTOMATION MAGAZINE JUNE 2023
occur-things that are out of the ordinary-which
can lead to preventing
poor quality or machine breakdown.
However, in many cases, the anomaly
detection just alerts an operator, who is
called over to investigate the situation
and make a decision about the appropriate
response (such as ignore, shut
down, or repair). Most of these solutions
are developed in a specific domain,
using SMEs knowledgeable about the
process. Generalizing these solutions
to new domains is difficult and generally
requires new SMEs to be engaged.
Could an AI be developed that would
facilitate this type of generalization?
What would it need to learn, and where
would it get the information?
In many manufacturing plants today,
although robots and humans may work
together to accomplish a job, they do
not work in close proximity, nor do they
request knowledge from
each other. They a re
separated because of
safety constraints, with
the robots sequestered
in cages. Newer collaborative
robots can
operate safely around
humans, although they
are generally restricted
to move more slowly
than their caged counterparts.
Despite closer
proximity, collaborative
robots do not request
additional information
from a human when new
or unfamiliar challenges
arise. As percept ion
technologies improve, there are opportunities
for AI to help robots and humans
work together on physical tasks, such as
assembly. This could help address the
challenge of personalized production.
As we move into the era of humancentric,
sustainable manufacturing
[3], questions of cooperation between
humans and AI become increasingly
more important. One question that
was raised was whether it was essential
to differentiate between human
intelligence and AI. From a manufacturing
system perspective, would it be
possible to start looking at any intelligent
system (humans, robots, computers,
etc.) as an intelligent agent with
specific capabilities? By blurring the
line between human intelligence and
AI, we can start to develop an evolutionary
process to support learning
resulting from a continuum of human
and AI knowledge.
"
IMPROVED COORDINATION
AMONG
UNIVERSITIES, INDUSTRY,
AND GOVERNMENT
AGENCIES
CAN FACILITATE
GREATER OPPORTUNITIES
TO PUSH THE
FIELD FORWARD.
„
INFRASTRUCTURE FOR AI IN
MANUFACTURING
Manufacturing systems are incredibly
complex. From the nonlinear physics of
low-level processes (such as welding
and 3D printing) to the material handling
within a plant, scheduling, and
the supply chain, everything needs to
come together for a high-quality part to
be produced and delivered to a consumer.
There are hundreds of thousands
of variables, of different types
and measured at different sample rates,
that interact in diverse
ways to describe what is
happening. Many manufacturing
plants operate
multiple shifts per
day or even 24/7, leading
to huge volumes of
data. There are thousands
or even millions
of different things that
can happen. AI technology
needs data to learn,
and although there are a
lot of data to be found
in manufacturing systems,
they
are not
always AI ready. Or,
even though there are a
lot of data, these data
may not be sufficiently diverse and
capture enough samples of important
events to develop effective learning and
intelligence applications. In short, most
data management systems were not
developed or put in place with AI in
mind (e.g., maintenance management
databases), and so they are not AI ready
in terms of many data quality dimensions
(e.g., accuracy, freshness, granularity,
volume/archive length, variety,
and context richness).
Data are collected through sensors,
and the panelists spent some time
discussing the limitations of current
IEEE Robotics & Automation Magazine - June 2023

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2023
