IEEE Robotics & Automation Magazine - September 2019 - 74

Figure 4(a) shows that, although NRI papers in traditional
robotics areas primarily cite other research in traditional
robotics areas (namely, computer science, engineering, and
mathematics), there are substantial numbers of citations to
less traditional fields, most noticeably medicine as well as
neuroscience, social science, suborganismal biology, and
physics. Figure 4(b) shows that NRI papers published in lesstraditional robotics areas cite a more even mix of fields. Engineering and computer science are still prominent, but they
are much less prominent than in Figure 4(a). For example,
biomedical fields are more prominent than in Figure 4(a).
This suggests that some NRI awards are making substantial
use of previous research well outside the traditional robotics
domain. Figure 5 shows dot plots of the percentage of citations from NRI awards to selected fields. For example, a dot
at 50% over biochemistry, genetics, and molecular biology
(BIOC) indicates that, for one NRI award, 50% of that

Table 1. Scopus four-letter research area codes
and definitions [11].

Code

Research Area

AGRI

Agricultural and biological sciences (including
ecology and organismal biology)

ARTS

Arts and humanities (including linguistics)

BUSI

Business, management, and accounting

BIOC

Biochemistry, genetics, and molecular biology

CENG

Chemical engineering

CHEM

Chemistry

COMP

Computer science

DECI

Decision sciences (including information
systems and statistics)

DENT

Dentistry

EART

Earth and planetary sciences

ECON

Economics, econometrics, and finance

ENER

Energy

ENGI

Engineering

ENVI

Environmental science (including ecology)

HEAL

Health professions

MATE

Materials science

MATH

Mathematics

MEDI

Medicine

MULT

Multidisciplinary

NEUR

Neuroscience

NURS

Nursing

PHAR

Pharmacology, toxicology, and pharmaceutics

PHYS

Physics and astronomy

PSYC

Psychology

SOCI

Social sciences

VETE

Veterinary

IEEE ROBOTICS & AUTOMATION MAGAZINE

award's citations are to papers classified as BIOC. Although
many awards have a small number of citations to nontraditional robotics fields, a number of awards cite these fields
for 20% or more of their references. In particular, for
26 awards (29% of those for which we have publications),
more than 25% of the citations are to at least one nontraditional robotics field.
Expanding the Scope of Robotics Research
The previous section showed that NRI researchers are
drawing on prior work outside traditional robotics. What
about dissemination? Given the interdisciplinary aims of
the NRI, is it also encouraging researchers to publish outside traditional robotics journals? To address this question,
we focus on nine research areas: suborganismal biology,
decision sciences, Earth sciences, materials science, medicine, neuroscience, physics, psychology, and social sciences.
For each of these research areas, a moderate number of
NRI-funded authors (more than five and fewer than 50)
have published at least one NRI-funded paper in the area.
We consider the publication record of each of these
authors, looking at all papers they have published since
2000, including those that were not funded by the NRI and
those published outside the target areas. The sizes of these
data sets are given in Table 2.
For each research area, we fit a multilevel Bayesian logistic
regression model for the probability that a given paper will be
published in the area, conditional on publication year and
whether the paper was NRI-funded. We partially pool the
intercept term and NRI effect across authors (within a given
area). By partially pool, we mean that we estimate intercept
and NRI effects for each individual author but model these
effects as draws from a common, unobserved distribution.
Intuitively, partial pooling means that the author-level effects
should be similar (they are connected by the unobserved distribution) but not constrained too tightly (they are not identical, for example). This allows us to make informed estimates
even when a given author has only a small number of publications in the data set [13].
More precisely, our models are all specified as a generalized linear regression
p( y i = 1) = logit -1( b year yeari + b j + b nri, j nri i ),

SEPTEMBER 2019

(1)

where y i = 1 indicates that publication i is in the research
area, logit is the logistic function, b year is a random effect for
the publication year, j is the author of publication i, b j is an
intercept term (constant or baseline effect) for author j, nri i
indicates whether publication i is NRI-funded, and b nri, j is an
author-specific NRI effect. In these models, the intercept term
b j can be interpreted as the probability (more precisely, the
log odds) that author j publishes in the given area outside the
NRI program, and b nri, j is the change in this probability when
author j publishes under the NRI program. Therefore, positive values of b nri, j indicate that the NRI increases publishing
in the target research areas.

IEEE Robotics & Automation Magazine - September 2019

Table of Contents for the Digital Edition of IEEE Robotics & Automation Magazine - September 2019