IEEE Robotics & Automation Magazine - June 2020 - 20

FROM THE GUEST EDITORS

Deep Learning and
Machine Learning in Robotics
By Fabio Bonsignorio, David Hsu, Matthew Johnson-Roberson, and Jens Kober

D

eep learning has gone through
massive growth in recent
years. In many fields-com-
puter vision, speech recogni-
tion, machine translation, game playing,
and others-deep learning has brought
unprecedented progress and become
the method of choice. Will the same
happen in robotics and automation?
In a sense, it is already happening.
Today, deep learning is often the most
common keyword for work presented
at major robotics conferences. At the
same time, robots, as physical systems,
pose unique challenges for deep learn-
ing in terms of sample efficiency and
safety in real-world robot applications.
With robots, data are abundant, but
labels are sparse and expensive to
acquire. Reinforcement learning in
principle does not require data labeling
but does require a significant number of
iterations on real robots. Transferring
the capabilities learned in simulation to
real robots and collecting sufficient data
for practical robot applications both
present major challenges. Further, mis-
takes by robot learning systems are
often much more costly than those by
their counterparts in the virtual world.
These mistakes may cause irreversible
damage to robot hardware or, even
worse, loss of human lives. Safety is thus
paramount for robot learning systems.
A related issue is interpretability, i.e.,
the ability for humans to understand
the learning process. We need shared
supervision and control approaches
Digital Object Identifier 10.1109/MRA.2020.2984470
Date of current version: 10 June 2020

20

*

IEEE ROBOTICS & AUTOMATION MAGAZINE

*

JUNE 2020

where humans have the option to inte-
grate with machine learning systems
and overcome the systems' cognitive
limitations. This requires that humans
understand the decision processes of
the machine in sufficient detail.
This special issue of IEEE Robotics
and Automation Magazine focuses on
approaches that have been validated on
real-world robots, scenarios, and auto-
mation problems. It features an excep-
tionally large number of scientific
articles, 11 in total, in keeping with the
recent massive growth in the field of
robot learning. The articles were careful-
ly selected by the guest editors in two
rounds of reviews and provide a good
account of the depth and breadth of the
research in this area. The methods range
from algorithms based on probability
theory to deep learning. The whole spec-
trum of robotics-from manipulators to
legged robots to drones to autonomous
vehicles-is covered, with applications
ranging from rehabilitation to cargo
transportation. Additionally, this issue
contains two other loosely connected
regular articles.
"Movement Primitive Learning and
Generalization" by You Zhou et al. con-
siders the representation of movements
capable of generalizing according to
variations in the environment and task,
rather than simply reproducing a dem-
onstration. The focus of this article is
the ability to represent behaviors that
have multiple modes (i.e., different
strategies) in a joint representation. The
final evaluation is a throwing task
where the robot decides whether to
bounce the ball off a wall or not

depending on the target's location.
"Gaussians on Riemannian Manifolds"
by Sylvain Calinon provides an over-
view of the use of Riemannian geome-
try in robotics. The article then
illustrates the use of Gaussians on Rie-
mannian manifolds for movement gen-
eration in more detail. This is important
as a possible way to reduce the compu-
tational burden of learning strategies by
considering the manifold structure of
the data coming from robot motions.
Similarly, the following article proposes
a structured approach to deal with tem-
poral features, another characteristic
aspect of robotic systems. "Interactive
Learning of Temporal Features for
Control" by Rodrigo Pérez-Dattari et
al. explores how robots can be taught
new skills interactively by human
teachers. The authors focus on tasks
where the learning agent needs to be
equipped with memory to succeed.
"Multifingered Grasp Planning via
Inference in Deep Neural Networks"
by Qingkai Lu et al. employs deep
learning to predict grasp success,
using a voxel representation of the
object and the grasp pose as inputs.
This model, jointly with a prior over-
grasp pose, is then used to infer the
grasp with the highest chance of suc-
cess on unseen, novel objects. "Opti-
mal Deep Learning for Robot Touch"
by Nathan Lepora and John Lloyd shows
how deep learning methods can be
employed to estimate object poses
based on input from optical tactile
sensors. The authors also demonstrate
how the method can then be employed
for 3D object exploration.
IEEE Robotics & Automation Magazine - June 2020

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - June 2020
