IEEE Robotics & Automation Magazine - December 2022 - 142

it. This is what we see in computer
vision. " (See " George Konidaris. " )
Tenorth confirms: " We do use ML in
several cases, mainly for perception,
and typically for isolated problems. "
This approach-what is known need
not be learned, and what is not known
must be learned-is summarized nicely
by Tenorth: " We often know how the
robot should behave because it's determined
by law, by processes, or by some
other specification. Then you have the
Moritz Tenorth
Moritz Tenorth is the CTO at Magazino GmbH (Munich,
Germany). He received his Ph.D. from the Technische
Universität München in 2011 on the topic of " Knowledge
Processing for Autonomous Robots. "
" I was very impressed by how much is possible if everyone
really has a common goal. "
Magazino GmbH develops and builds intelligent, mobile
robots for intralogistics. Their products include the TORU robot for autonomous
shoe box picking, and SOTO, for industrial production supply. AI methods are used
in the overall architecture called Advanced Cooperative Robot Operating System.
ML is used for individual problems, especially perception.
George Konidaris
George Konidaris is director of the Intelligent Robot Lab and an
associate professor at Brown University (Providence, Rhode Island,
USA). He is the chief roboticist at Realtime Robotics. His Ph.D.
thesis was " Autonomous Robot Skill Acquisition " at the University of
Massachusetts Amherst in 2011.
" Real robot hardware keeps you honest. "
Realtime Robotics offers products for the intuitive
programming of automated, collision-free motion plans and real-time control
to execute these plans as well as spatial perception capabilities for dynamic
unstructured environments. AI methods are used for motion planning and
perception.
Pieter Abbeel
Pieter Abbeel is director of the Berkeley Robot Learning Lab
and a professor at the University of California, Berkeley. He is
a cofounder of Gradescope and covariant.ai. He received his
Ph.D. from Stanford University in 2008 with the thesis
" Apprenticeship Learning and Reinforcement Learning With
Application to Robotic Control. "
" From day one, make sure that you work hard on
understanding your customers. It is critical to know what
really matters to them. "
Robots at covariant.ai learn general abilities such as robust 3D perception,
physical affordances of objects, few-shot learning, and real-time motion planning.
This allows them to adapt to new tasks just like people do: by breaking down
complex tasks into simple steps and applying general skills to complete them.
142 * IEEE ROBOTICS & AUTOMATION MAGAZINE * DECEMBER 2022
choice of either programming it yourself
or collecting data and reverse engineering
the learning process so that the
outcome of the learning process is what
you had in mind beforehand. And that
is the learning problem upside down. "
The fact that robots act and make
decisions makes them such a challenging
application domain. As Konidaris
explains, " Robotics, because of its
sequential decision-making aspect, is
fundamentally harder than supervised
learning. " So indeed, we see that much
of the ML successes in robotics have
been achieved in problems that can be
formulated as supervised learning problems,
such as computer vision.
If ML is the method of choice to solve
an application task, workflows for data
acquisition are essential. As Pieter Abbeel
puts it, " A big part of our job is collecting
a lot of data. " (See " Pieter Abbeel. " ) Simulation
and simulation-to-real approaches
are important tools to facilitate data collection.
Care should be taken to focus on
the real problem at hand, rather than the
simulated problem. The way Konidaris
sees it, " What you find when you build
the simulation is that you end up cobuilding
the simulation, the benchmark, and
the learning algorithms so that everything
sort of works out together. Real
robot hardware keeps you honest. "
To summarize these insights and
comments, innovation with ML methods
in robotics will currently most likely
be successful for isolated problems that
can be formulated as supervised learning,
and in which data acquisition is not
too expensive, or can be automated with
simulators that mirror the robot hardware
sufficiently well.
Innovation: From Research Ideas
to Products and Services
Innovation is the process of turning ideas
into products or services. The transition
from academic researcher to innovator
thus requires a shift of perspective. As
Tenorth explains, " Before I looked at it
from a research point of view. What is
novel? What hasn't been researched
beforehand? Also, what can you write a
paper about? In a company, it's about
what works. Maybe a simple heuristic is
good enough in some cases and saves
you a lot of complexity. " All the interviewees
corroborate the following:
" From day one, make sure that you work
hard on understanding your customers.
It is critical to know what really matters
to them, " explains Abbeel. From Konidaris'
perspective, " If you're not talking to
customers, then your symbols are all
ungrounded when you're thinking about
application. " Understanding customer
needs requires a substantial effort that
does not involve technological progress
http://www.covariant.ai http://www.covariant.ai
IEEE Robotics & Automation Magazine - December 2022

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