IEEE Robotics & Automation Magazine - December 2022 - 141

INDUSTRY ACTIVITIES
Innovation Paths for
Machine Learning in Robotics
By Freek Stulp , Michael Spranger, Kim Listmann, and Stéphane Doncieux
Based on interviews with Moritz Tenorth, George Konidaris, Pieter Abbeel
A
dvances in artificial intelligence
(AI), especially in machine
learning (ML), are changing
the business models of many
companies, and creating entirely new
ones. Recent research estimates that AI
could boost profitability rates by 38%
worldwide, leading to an economic
boost of €12 trillion across a variety of
industries by 2035. This immense
number is an accumulation of many
smaller numbers, related to the
successful deployment of ML at
individual companies, including smalland
medium-sized enterprises (SMEs)
and start-ups.
On the other hand, it appears that
40% of Europe's self-proclaimed AI
start-ups do not use AI technology in a
way that is essential to their business. In
a recent IEEE International Conference
on Intelligent Robots and Systems
(IROS) workshop and the ongoing
European project " VeriDream, (VERtical
Innovation in the Domain of Robotics
Enabled by AI Methods) " we
address several questions related to the
successful deployment of AI methods in
robotics companies, especially concerning
ML methods that have resulted
from academic research projects. These
questions include the following:
●
●
What are the consequences of building
on academic open source software,
and what is the impact of the
software license chosen?
●
How can the long tail of failures that
impact safety and reliability, particularly
important in robotics, be effectively
addressed?
In this " Industry Activities " column,
we summarize the main insights and
personal experiences of three leaders in
the field of robotics who have earned a
name for themselves both as academic
researchers and as successful innovators,
based on a set of interviews at the
aforementioned IROS workshop. We
believe these insights to be valuable to
companies who are contemplating the
deployment of ML, as well as researchers
in academia whose aim it is to one
day found a start-up. This article categorizes
the use cases and challenges
related to ML methods, describes recommendations
for spinning off a
start-up on ML and robotics, and
highlights that reliability is often the
main bottleneck for successful innovation.
(See " Terminology. " )
Which Application Use Cases are
Appropriate for ML Methods?
In the domain of AI, the concept that
speaks to the imagination the most is
" artificial general intelligence, " i.e., a system
that is able to achieve human-level
intelligence or beyond. We believe such
systems will remain science fiction for
quite a while to come. As Moritz
Tenorth mentions: " I'm skeptical that
you can really have one system that
learns it all. And the other question is,
even if you could do this, should you do
this? " (See " Moritz Tenorth. " )
In practice, the first choice for
achieving reliable intelligent behavior is
still engineering as " powerful engineering
enables reuse, easy configuration,
and easy debugging during operation, "
according to Tenorth. On the other
hand, problems for which no engineering
solution is known are candidates
for ML methods. George Konidaris
describes the thought process as follows:
" This is a hard problem and I
don't know how to solve it. Therefore, I
will use learning because learning is
more efficient than trying to engineer
How can start-ups manage the transition
from novelty-driven research
to identifying and satisfying customer
needs?
Digital Object Identifier 10.1109/MRA.2022.3213205
Date of current version: 2 December 2022
Terminology
We consider AI to be a property of a system, i.e., whether a system is able to
display intelligent behavior or not. There is no common agreement on the exact
definition or scope of the term, but it usually includes the ability to perceive,
plan, reason, learn, and interact with humans. Common approaches for achieving
intelligent behavior in systems include automated reasoning, optimization,
evolutionary algorithms, statistical learning methods, artificial neural networks, and
deep learning. The latter three belong to an important subclass of AI methods called
Machine Learning, where tasks and representations are learned from data. The
main focus of this article is on ML.
DECEMBER 2022 * IEEE ROBOTICS & AUTOMATION MAGAZINE *
141
http://orcid.org/0000-0001-9555-9517
IEEE Robotics & Automation Magazine - December 2022

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