Eileen Roesler, Dietrich Manzey, Linda Onnasch
A meta-analysis on the effectiveness of
anthropomorphism in human-robot interaction
Open Access via institutional repository of Technische Universität Berlin
Document type
Journal article | Accepted version
(i. e. final author-created version that incorporates referee comments and is the version accepted for
publication; also known as: Author’s Accepted Manuscript (AAM), Final Draft, Postprint)
This version is available at
https://doi.org/10.14279/depositonce-12447
Citation details
Roesler, E., Manzey, D., Onnasch, L. (2021). A meta-analysis on the effectiveness of anthropomorphism in
human-robot interaction. Science Robotics, 6(58). https://doi.org/10.1126/scirobotics.abj5425.
Terms of use
This work is protected by copyright and/or related rights. You are free to use this work in any way permitted by
the copyright and related rights legislation that applies to your usage. For other uses, you must obtain
permission from the rights-holder(s).
ANTROPOMORPHISM IN HRI
1
The Effects of Anthropomorphism on Human-Robot Interaction: A Quantitative Meta-
Analysis
E. Roesler1*, D. Manzey1, & L. Onnasch2
1Technische Universität Berlin, Germany, 2Humboldt-Universitätzu Berlin, Germany
* Corresponding author (eileen.roesler@tu-berlin.de)
ANTROPOMORPHISM IN HRI
2
Abstract
The application of anthropomorphic design features is widely assumed to facilitate human-
robot interaction (HRI). However, a considerable number of study results point in the
opposite direction. There is currently no comprehensive common ground on the
circumstances under which anthropomorphism promotes interaction with robots. This meta-
analysis aims to close this gap. A total of 4,856 abstracts were scanned. After an extensive
evaluation, 78 studies involving around 6,000 participants and 187 effect sizes were included
in this meta-analysis. The majority of the studies addressed effects on perceptual aspects of
robots. In addition, effects on attitudinal, affective, and behavioral aspects were also
investigated. Overall, a medium positive effect size was found, indicating a beneficial effect
of anthropomorphic design features on human-related outcomes. However, closer scrutiny of
the lowest variable level revealed no positive effect for perceived safety, empathy, and task
performance. Moreover, the analysis suggests that positive effects of anthropomorphism
depend heavily on various moderators. For example, anthropomorphism was in contrast to
other fields of application, constantly facilitating social HRI. In conclusion, the results of this
analysis provide insights into how design features can be used to improve the quality of HRI.
Moreover, they reveal areas in which more research is needed before any clear conclusions
about the effects of anthropomorphic robot design can be drawn.
ANTROPOMORPHISM IN HRI
3
Introduction
Robots are making inroads into our working life and everyday world (1,2). Whereas
early robot generations were mainly limited to industrial robots that worked in safety cages,
kept apart from human workers, current robotic agents are increasingly interactive. In this
process, interaction is changing from a segregated coexistence to direct collaboration with
humans in the same space and time. The ability to collaborate, in turn, enables the
implementation of robots in more diverse domains (3). In addition to being deployed in
industrial settings, robots are also becoming more common in service and social fields of
application such as school teaching and elderly care. This general shift of robots entering the
world of humans is increasingly accompanied by the application of human-like features in
robot design (4–7). The postulated effectiveness of this anthropomorphic design approach is
mainly based on two assumptions. First, robots are used in an environment that is designed
and optimized for humans. For this reason, the application of human-like design is assumed
to support a naturalistic and functional embodiment (4). Structural and functional similarities
e.g., limbs and joints provide the capabilities, which can support a successful movement
through an environment and an interaction with artefacts built for humans (8,9). Second,
from a human-centered point of view, anthropomorphism promotes more intuitive interaction
for people because it enables the transfer of scripts that are well known from human-human
interaction (10,11).
Anthropomorphism in HRI is thereby a reciprocal phenomenon. On the one hand, it
describes the general tendency of people to attribute human characteristics including human-
like mental capacities to non-living objects (12,13). On the other hand, anthropomorphism
describes a human-like design of robots that in turn facilitates the attribution of human-like
characteristics to the robot (3). This design element is used to evoke expectations, which, if
met, represent a knowledge base for interaction and a better anticipation of robots’ actions,
ANTROPOMORPHISM IN HRI
4
even for first encounters with this often completely new technology (5,11,14). Figure 1
shows a number of examples of anthropomorphic robot designs in different domains of
human-robot interaction (HRI). The examples also illustrate that most straightforward
approaches of anthropomorphic robot design address the overall appearance of robots (e.g.,
face-like characteristics or body shapes). However, other approaches include more subtle
aspects such as anthropomorphic trajectories, language-based communication, or simply
different types of framing (e.g., giving robots human names or human-like descriptions).
Fig. 1. Examples of anthropomorphic implementations. Anthropomorphic design by means of depicting human-
like facial features or body features for the industrial (left: Sawyer; right: Nextage), service (left: Pillo Health ;
right: SnackBot), and social domain (left: BUDDY; right: Pepper) received from the Anthropomorphic Robot
(ABOT) Database (15)
But is this design approach generally beneficial for HRI? While current research in
social application domains broadly supports this assumption (4,5,12), a different picture
emerges in other domains. For example, studies focusing on industrial HRI suggest that
anthropomorphic design features may not necessarily be beneficial, and can undermine the
perceived reliability of robots (16) and raise concerns with regard to their safety (17). These
results are unexpected, because the transfer of human-human interaction scrips should make
interaction more familiar and trustworthy, independent of the application domain in question.
ANTROPOMORPHISM IN HRI
5
Interestingly, negative effects are not only observed in the industrial domain, but also in other
domains where humans have to perform a certain task in collaboration with a robot. In this
case, an anthropomorphic robot representation may again lead to counterproductive and
unintended effects, including a decrease in prosocial behavior (18), or overshooting effects
such as an inappropriately strong emotional attachment to the robot (19) .
Overall, these examples suggest that anthropomorphic design can lead to diverse and
unintended outcomes. However, our current knowledge about the context factors that make
anthropomorphic robot design beneficial have not yet been systematically identified, and a
comprehensive integration of the available research is lacking.
With this meta-analysis, we aim to close this research gap by (1) estimating the
overall effect of anthropomorphism on human-related outcomes, (2) separately estimating
effects of anthropomorphism on different facets of human-related outcomes, and (3) taking
into account possible moderators. The basic framework for this analysis, depicted in Figure 2,
includes and arranges the key variables considered in our meta-analysis.
Fig.2. Basic framework of the meta-analysis.
The anthropomorphism of the robot represents the relevant input variable. For this
reason, only studies that investigated the effects of at least two different degrees of
ANTROPOMORPHISM IN HRI
6
anthropomorphic robot design were considered in this meta-analysis to estimate the
effectiveness of increasing the anthropomorphism of robots. The primary aim of the analysis
was to examine the generally assumed positive effects of anthropomorphic robot design. We
therefore excluded studies that explicitly address what is commonly referred to as the
uncanny valley effect in HRI, which focuses on negative consequences of highly
anthropomorphic designs in terms of disturbance and eeriness (20).
The relevant dependent variables are summarized as human-related outcomes in terms
of subjective and objective interaction experiences (21–24). We identified four main
categories of outcomes based on an extensive analysis of the current body of research. The
first category is people’s perception of robots. Most of the relevant research in this area was
based on the Godspeed questionnaire series (25). Besides evaluating anthropomorphism and
animacy itself, this questionnaire series assesses how likeable, intelligent and safe a robot is
perceived to be by the human counterpart. The second category covers different attitudes
towards robots. Previous research has shown that attitudes such as trust, acceptance, and
empathy are important determinants of people’s actual behavior in HRI, and specifically their
willingness to work together with their robotic counterpart (26,27). Whereas trust (26,27)
and acceptance (28,29) are assumed to be mainly associated with effective and efficient
interaction, empathy seems to be especially relevant in social HRI settings (22,30). The
remaining two outcome categories include affective reactions (31–33), i.e., activation and
pleasure in terms of pleasure-arousal theory (34,35), and behavioral responses, including
task performance (36,37) and social behavior shown in interaction with a robot (18,22).
To investigate the circumstances under which anthropomorphism facilitates HRI, our
analysis further considers several moderating variables. Based on reviews (14,38)and a
recent taxonomy of HRI (39), we identified four central moderators that might explain
ANTROPOMORPHISM IN HRI
7
possible heterogeneity in individual study results. The first moderator relates to the
interaction environment, and sets the conditions and constraints for the configuration of
interaction, i.e., the field of application (39). The fields of application considered are
categorized as the social, service, and industrial domain. The social domain is defined as any
domain where robots are used in therapeutic, educational, or entertainment settings (39). The
service and industrial domain are defined based on the International Organization for
Standardization and (ISO 8373:2012) (40). In these fields of application, robots perform
useful functional tasks for humans such as transport, physical load reduction, and precision.
In addition, this moderator variable includes a fourth category (“none”), given that some HRI
studies focus on the pure perception of robots without any contextual information.
The next two moderator variables include different aspects of the robot itself. One is
the instrumentality of the anthropomorphic design feature. Studies suggest that it might make
a difference whether or not anthropomorphic features are related to the task in a meaningful
manner (e.g., randomly moving eyes vs. predictive eyes (41)). Whereas task-relevant
implementation may lead to increased task performance, this is probably less the case with
task-irrelevant implementation of anthropomorphic features. In addition, the impact of this
moderator might also be different for various outcome categories of HRI. In contrast to task-
irrelevant implementations, task-relevant anthropomorphic design might directly improve
actual performance, but it seems less obvious whether it also differently affects people’s
perception of or attitude toward robots. The third moderator addresses how anthropomorphic
features are implemented in the robot’s morphology (39), i.e., the appearance,
communication, movement and/or the context in which the robot is framed and introduced to
users.We assume that different implementations of anthropomorphism can be variously
effective with regard to different outcome categories. For example, whereas an
anthropomorphic appearance might not affect task performance (42), anthropomorphic
ANTROPOMORPHISM IN HRI
8
movements might do so by improving predictability of the robot’s actions, thereby enhancing
coordination in task fulfilment (43). In addition to the four implementation categories, a fifth
is added to cover cases where multiple anthropomorphic features are combined.
Finally, the last moderator in the framework comprises a more research-relevant
aspect, involving the question of how to expose humans to robots in HRI studies, i.e.,
whether humans interact directly with embodied robots (i.e., real machines) or must merely
imagine interaction based on depictions of robots (i.e., virtual two-dimensional agents). Both
approaches are used in HRI research, but there is no comprehensive ground yet regarding
how this might affect the results (44–46). To shed light on this issue, robot exposure is
included in this analysis by categorizing the robots used as either depicted or embodied (39).
In summary, although consequences of anthropomorphic features in HRI have been
investigated widely, we still lack knowledge about the generalizability of specific results
produced by individual studies. Based on the proposed framework, this study aims to
systematically review and quantify the effects of anthropomorphism on identified human-
related outcomes. Moreover, the analysis takes into account the role of moderators to enable a
differentiated understanding with regard to the circumstances under which
anthropomorphism can facilitate or hinder HRI. To achieve this goal, we applied quantitative
meta-analytic methods to the existing literature on anthropomorphism in HRI.
Results
Figure 3 illustrates the overall effect of anthropomorphism, as well as the effects for
the different outcome categories and specific variables.
ANTROPOMORPHISM IN HRI
9
Fig. 3. Forest plot of the overall effect size and all sublevels. Depiction of standardized mean differences
(Cohen’s d) shown by the positions of the squares, the 95% CIs by the whiskers, and the numbers of included
studies by the size of the squares.
Overall effect
The analysis revealed a positive overall effect of anthropomorphism on human-related
outcomes with a medium average effect size (d=0.501, 95% CI [0.394-0.608]). However, the
analysis also revealed a high level of heterogeneity (Q(186)= 1684.25, p<.001, I2=88.1%),
suggesting diverse effects on different outcome variables and/or an impact of moderator
variables.
Human-related outcomes
Perception. The analysis showed that people’s perception of robots is the most
frequently investigated construct (k=99) to evaluate the consequences of anthropomorphic
design in HRI. Overall, the reported effects of anthropomorphism on perception result in a
ANTROPOMORPHISM IN HRI
10
medium average effect size (d=0.570, 95% confidence interval (CI) [0.443-0.698]), again
with a high level of heterogeneity (Q(98)= 753.57, p<.001, I2=84.93%). The separate analyses
for the different subdimensions suggest that the overall positive effect of anthropomorphism
on people’s perception of the robot is mainly driven by the subcategories of likeability
(d=0.606, 95% CI [0.411-0.800]) and intelligence (d=0.647, 95% CI [0.467-0.827]). In
contrast, the data revealed no consistent effect for studies addressing the perceived safety of
robots (d=0.168, 95% CI [-0.131-0.466]).
Attitudes. A similar pattern of effect sizes emerged regarding attitudes towards
robots, although this aspect was based on a considerable smaller set of studies (k=25). The
analysis again revealed a positive overall effect (d=0.616, 95% CI [0.296-0.936]) with a
pronounced heterogeneity (Q(24)= 199.80, p<.001, I2=90.51%). The subset analyses showed
that the overall effect was mainly due to two subcomponents, i.e., a positive effect of
anthropomorphism on trust with a medium effect size (d=0.726, 95% CI [0.216-1.235]), and
a positive effect on acceptance with a large effect size (d=0.877, 95% CI [0.318-1.436]). In
contrast, no consistent positive effect was found for empathy towards robots (d=0.153, 95%
CI [-0.107-0.413]).
Affect. The effects of anthropomorphism on affect are least investigated, having been
addressed in only k=18 studies. The mean effect size of these studies is again positive
(d=0.386, 95% CI [0.181-0.591]). Compared to the effects on perception and attitudes, it is
somewhat smaller, but also more consistent with less remaining heterogeneity (Q(17)= 37.58,
p<.01, I2=55.67%). In this case, the overall effect is also representative for both
subcomponents, characterized as activation (d=0.441, 95% CI [0.202-0.682]) and pleasure
(d=0.351, 95% CI [0.023-0.678]).
ANTROPOMORPHISM IN HRI
11
Behavior. The effects of anthropomorphism on human behavior in HRI were
addressed in k=45 studies. Overall, anthropomorphism has a small positive effect on this
outcome category (d=0.318, 95% CI [0.046-0.590]). This positive effect can be mainly traced
back to beneficial effects on social behavior (d=0.378, 95% CI [0.140-0.616]). In contrast, no
consistent improvements emerged for task performance (d=0.259, 95% CI [-0.222-0.740]). In
line with the results for perception and attitudes, the analysis of this outcome category also
revealed a large degree of systematic heterogeneity between studies (Q(44)= 616.10, p<.001,
I2=91.68%), again suggesting the effects of moderator variables.
Moderators
The results presented show that the meta-analytic models used to analyze the effects
of the different studies almost always indicated a relatively high level of heterogeneity in the
data. This suggests that moderators most likely contributed to the differences between
studies. Figure 4 shows the results of the moderator analyses addressing the set of a priori
identified outcome categories. Each single graph in the figure illustrates the differences
between mean effect sizes dependent on the categories of a given moderator (columns) and
the different outcome categories (rows).
ANTROPOMORPHISM IN HRI
12
Fig. 4. Forest plots illustrating the effects of moderators. The plots show standardized mean differences
(Cohen’s d), the 95% CIs, and the number of effect sizes included, given separately for the overall effect and all
subcategories, dependent on the characteristics of the different moderators (columns). The moderator variables
are (i) field of application (SO, social; SE, service; IN, industrial; NO, none), (ii) task relevance (R, relevant; IR,
irrelevant), (iii) morphology (MT, multiple; MO, movement; CM, communication; AP, appearance; CX,
context), and (iv) robot exposure (DE, depicted; EM, embodied).
ANTROPOMORPHISM IN HRI
13
Field of application. On an overall level, the field of application explained only 0.9%
of heterogeneity (QM= 3.83, p=.28). Closer scrutiny reveals that a consistent positive effect
size across all different outcome categories was only found for the social domain, whereas no
comparable consistent effects of anthropomorphism emerged for studies of HRI in the service
domain. A somewhat mixed pattern of results emerged for the industrial domain. In this case,
anthropomorphism yielded small to medium effects for perceptual and affectional outcomes.
Finally, studies with no clearly defined field of application found consistent beneficial effects
of anthropomorphism for people’s perception of the robot only, while no comparable
consistent results were found for the other sets of outcome categories.
Task relevance. For the overall effect, task relevance did not account for any
heterogeneity (QM= 0.28, p=.597). Independent of whether or not anthropomorphic design
features were implemented in a task-relevant manner, they led to positive effect sizes for all
outcomes apart from behavioral ones. For this latter category, the task relevance of
anthropomorphic features seems to be a necessary condition for achieving positive effects.
Morphology. The overall positive effect of anthropomorphism was moderated by
how anthropomorphism was implemented, i.e., the dimension used to increase the
anthropomorphism of a robot (QM= 11.44, p<.05). Specifically, multiple implementations
(d=0.703, 95% CI [0.38-1.025], p<.01), implementations via movement characteristics (d=
0.645, 95% CI [0.41-0.879], p<.01), and implementation of human-like communication
(d=0.583, 95% CI [0.396-0.769], p<.05) significantly increased the positive effect compared
to using context framings only (d=0.054, 95% CI [-0.306-0.414]). Regarding appearance, at
least a non-significant trend for increased effectiveness compared to the context was found.
On the sublevel of outcome categories, communication and multiple implementations of
anthropomorphic features most consistently led to positive effects for three of the four
ANTROPOMORPHISM IN HRI
14
outcome categories. Anthropomorphic appearance and movement only resulted in a positive
effect size for perception, and the anthropomorphic context did not lead to any positive effect
on any of the outcome categories.
Robot exposure. The physical presence of the robot did not account for any
heterogeneity (QM= 0.099, p=.753) of the overall effect. Medium effect sizes with similar
values were present for both studies using depicted robots and studies using embodied robots.
On the sublevel of the outcome categories, a double-edged picture emerged. Whereas studies
using embodied robots report consistent beneficial effects across all outcomes, studies using
only depicted robots for their research merely found a positive effect with respect to
perception and attitudes.
Publication bias
The visual inspection of the data via a funnel plot showed a left-sided asymmetry,
which indicates that more effect sizes were included in our analysis that underestimate the
true effect compared to effect sizes that overestimate it. This asymmetry was supported by a
significant Egger’s regression test for funnel plot asymmetry (z=4.47, p<.001). More
precisely, the trim-and-fill method revealed that the estimated number of missing studies was
26 on the right side and none on the left. In comparison to the uncorrected overall effect of
anthropomorphism (d=0.501, 95% CI [0.394-0.608]), the trimmed-and-filled dataset resulted
in a slightly higher overall effect size (d= 0.655, 95% CI [0.542-0.7679]).
Discussion
The objective of this meta-analysis was to investigate the effects of anthropomorphic
design features on human-related outcomes, and to take into account relevant moderators.
The results reveal that adding anthropomorphic features to HRI leads to a considerable
overall positive effect, which is in line with previous research (21–24). Moreover, the results
ANTROPOMORPHISM IN HRI
15
show that this holds true for all different outcome categories considered in this analysis, with
moderate effects of anthropomorphism for perception and attitudes, and relatively smaller
effects for affect and behavior. The analysis further revealed that most studies thus far have
focused on the impact of anthropomorphism on perceptual aspects such as the perceived
intelligence or likeability of robots (25). Thus, the perceptual category represents the most
important source for the overall positive effect. This overrepresentation of perception
compared to other categories in HRI research does not seem to be justified by its greater
relevance. Instead, it seems to be primarily related to the ease of accessibility of this sort of
outcome variable. For instance, one of the most commonly used tools in HRI research is the
Godspeed questionnaire series (25) (and the according revised version (47)). This is a very
cost- and time-effective measure that addresses aspects of how people perceive robots (25,
48). In contrast, effects of anthropomorphism on affect or even behavioral outcomes require
more complex assessment approaches. However, attitudes are also less commonly
investigated. This is surprising for two reasons. First, the ease of accessibility of attitudes as a
subjective measure is comparable to that of perceptual evaluations (28). Second, the positive
effects of anthropomorphism on trust and acceptance are some of the most commonly
mentioned ones in the literature (12,27). Obviously, there is a gap between the theoretically
postulated importance of attitudes for a successful HRI and gaps in the research on this
specific topic that need to be closed by future studies. In addition, our results call for more
research on behavioral outcomes. Regardless of the domain in which humans and robots
collaborate, the primary goal of anthropomorphic design features will always be to improve
behavior (e.g., physical stimulation in therapeutic settings or smooth joint manipulation of
work pieces in industry). Of course, it is important to investigate subjective perceptions of
robots and attitudes towards them in HRI research (26,27), given that both presumably
ANTROPOMORPHISM IN HRI
16
determine people’s behavior and willingness to work together with a robot. However, actual
behavior will always be the key concern, and should not be neglected in research.
More detailed analysis on the specific variable level (per outcome category) further
suggests that anthropomorphic design features have no impact on the empathy towards
robots, the perceived safety of robots, or performance in joint tasks with robots. The non-
existent positive effect of empathy might again be related to the underrepresentation of
research on this rather specific aspect (k=7). In contrast, the missing effects on perceived
safety and task performance can certainly be considered a reliable finding because the
analysis was based on a relatively higher number of studies, specifically in non-social HRI
settings. The lack of evidence for improved task performance challenges the assumption that
equipping robots with anthropomorphic features might activate human-human interaction
schemes in HRI, which, then, intuitively supports task-related behavior (11). Combined with
the overall null effect on perceived safety, it suggests that anthropomorphic design features
might primarily be used to improve social aspects in HRI (5,12), but not task-related aspects.
The additional consideration of possible moderators generated further insights into the
specific circumstances that might determine the effectiveness of anthropomorphism. The first
moderator was the field of application. In line with an already sound body of research (4,5,
12), the results show that the social domain consistently benefits from the application of
anthropomorphism. This positive effect is not directly transferable to other domains, though.
Specifically, the service domain does not seem to benefit at all from anthropomorphic robot
design. A possible explanation could be that anthropomorphic features lead to an emotional
attachment (19), which might undermine a person’s willingness to use the robot as tool.
Whereas anecdotal evidence (49) for this assumption exists (e.g., delivery robots are used less
ANTROPOMORPHISM IN HRI
17
if they are anthropomorphized more), further research is needed to consolidate this
hypothesis.
The second moderator addressed whether or not it makes a difference if
anthropomorphic design features directly relate to the task at hand. Our data confirm the
expectation (41) that the task relevance of implemented anthropomorphic design features is
only a crucial factor for facilitating HRI with respect to behavioral outcomes. This finding
seems to be particularly important for actual work-related collaborative interactions. It
suggests that it is worthwhile to implement anthropomorphic features in a task-relevant
manner (e.g., social cues, predictive movements) whenever humans and robots collaborate on
certain tasks.
The third moderator considered in our analysis included effects of how specifically
anthropomorphic features were implemented, i.e., based on appearance, the communication
channel, movements, or just the type of framing. The data demonstrate that different
implementations of anthropomorphism can lead to a variety of effects. Not surprisingly,
approaches based on multiple as well as communicational anthropomorphic features turned
out to be most effective with regard to the different outcome categories. In contrast, the mere
use of different sorts of framing to induce an anthropomorphic context, e.g., giving a robot a
name and a personalized story (18,22), does not seem to be effective, having no reliable
overall effect on any of the outcome categories. There may be two reasons for this missing
positive effect of context anthropomorphism: the limited salience in comparison to more
visually detectable anthropomorphic features, and the possible masking of the robot’s
functional value by covering its task-related features as a tool (18). Other morphological
features were effective for some, but not all human-related outcomes. On the one hand, the
positive effect of appearance on perception is not surprising, because an anthropomorphic
ANTROPOMORPHISM IN HRI
18
appearance is described as the most salient characteristic (12,21). On the other hand, it might
be possible that anthropomorphic appearance had no effect on attitudes, affect and behavior
because of the non-functional character of appearance (39). In addition, appearance can
establish certain expectations regarding the robot’s functionalities that might get violated in
following interactions.
Finally, the last moderator variable addressed methodological issues of HRI studies
and investigated whether the efficiency of anthropomorphism depends on how the robots are
presented to participants, i.e., in a physically embodied manner that allows for lively
observation or even direct interaction, or merely by two-dimensional representations. Here,
our results reveal a gap between subjective and objective outcomes. Regardless of how
participants are exposed to a robot, positive effects of anthropomorphism emerged for
perception and attitudes, both of which are usually assessed via subjective questionnaires.
However, positive effects on affect and behavior, which concern actual physiological and
behavioral reactions (21–24), are usually only found in studies that involve presenting “real”
robots to the participants. Earlier research indicated both similarities (45)and differences (44,
46) between physically embodied robots and virtual two-dimensional representations. The
gap between subjective and objective reactions indicates a possible systematic explanation
for these mixed results and could be instructive for future research. If perceptions or attitudes
towards (anthropomorphic) robots are of the main interest, it seems sufficient and
ecologically valid to conduct studies using virtual agents or images of robots. However, if
affective or behavioral outcomes are central to an investigation, researchers should seek to
use studies involving physically present robots that enable real interaction so as to gain valid
insights.
ANTROPOMORPHISM IN HRI
19
Overall, the analysis suggests that it is counterproductive to draw general conclusions
on the impact of anthropomorphism on HRI when these are based solely on perceptual
evaluations. Apart from a handful of exceptions (i.e., in the service domain, implemented via
the context), anthropomorphism is always beneficial to people’s perception of robots.
However, this effect does not seem to be transferable to other more reciprocal interactional
outcomes such as the behavioral outcomes considered in our meta-analysis. Moreover, the
analysis illustrates another even more important issue regarding the transferability of effects
of anthropomorphism. Based on the shift of the robot’s role from a tool to a team partner
(39), it has often been assumed that the results gained in social HRI can be transferred to
other fields of application. However, the results suggest that the stable positive effect of
anthropomorphism in social HRI may not be directly transferable to other domains. For
example, essentially no positive effects of anthropomorphism were found in the service
domain, and only partial effects were determined in the industrial domain. This shows the
inadequacy of transferring insights from social HRI to more task-related settings.
Furthermore, the overall effectiveness of anthropomorphism on social behavior, but not on
task performance, challenges the usefulness of anthropomorphic features in those domains. In
sum, even though the analysis showed no evidence for a negative impact of anthropomorphic
design, anthropomorphism also does not generally improve the quality of HRI. Whereas
social HRI consistently benefits from anthropomorphic robot design, a mixed picture emerges
for other application domains. In addition, the way anthropomorphism is implemented seems
to determine its success. Most of all, our results suggest that interaction quality between
humans and robots can particularly be promoted by implementing anthropomorphic
communication features, by multiple implementations of anthropomorphism and by
implementing task-relevant anthropomorphism.
ANTROPOMORPHISM IN HRI
20
Meta-analyses must always be interpreted with caution, because they equally include
measures of various study designs involving different numbers of participants. However,
given the systematic procedure and the comparably high number of effect sizes included, we
assume that the global conclusions presented above are indeed reliable findings. Moreover,
the analysis of possible publication bias suggests that if a bias is present at all, it has biased
our analysis conservatively with regard to the impact of anthropomorphic robot design.
Nonetheless, one major limitation of the study concerns the non-consideration of different
degrees of anthropomorphism. Most of the empirical effects included in the analysis
contrasted only two different degrees of anthropomorphism, which could hardly be located
on an overall dimension. The main reason for this limitation is that the exact degree of
anthropomorphism of robots cannot be measured objectively. Thus, even though it was
possible to detect some major moderating factors of effects of anthropomorphism, we are
unable to make any conclusions about the degree of anthropomorphism required to induce
certain effects (25,48). This will be a matter of future research, and we hope that our meta-
analysis will be a good starting point for such research. The fact that we have the entire data
and material of this meta-analysis available online will enable other researchers to add more
data and to expand this data base over time. By taking this approach, our meta-analysis serves
not only as a state-of-the-art research synopsis, but moreover aims to iteratively create a
sound basis for investigating the consequences of anthropomorphism in future science and
practice.
Materials and methods
Before starting the systematic literature search, the meta-analysis was preregistered
and described in detail in the standardized procedure of preferred reporting items for
ANTROPOMORPHISM IN HRI
21
systematic review and meta-analysis protocols (50) via the open science framework (51). The
entire methodical procedure and all data generated during the process, from the literature
search to the actual analysis of the data, are available online to enable other researchers to
replicate and further extend the analysis in the future (51).
Based on the objective of the study, the terms used for the literature search included
combinations of <human-robot interaction or social robot> and <anthropomorphism or
anthropomorphic or humanlike> and <experiment or subject or participant or user study)>.
The literature search was conducted between April and June 2020. The comprehensive
procedure, encompassing also the list of inclusion criteria, is illustrated in Figure 5.
Fig. 5. Search flow diagram. Depiction of the entire process of data collection, including the sources searched,
the inclusion criteria, and the selected articles.
ANTROPOMORPHISM IN HRI
22
The first step involved scanning entries of the most common electronic databases of
scientific literature, as well as the first 500 Google Scholar hits. The 4,856 resulting abstracts
were analyzed, and all studies that did not violate the inclusion criteria were selected,
resulting in a total of 325 articles, without duplicates and non-accessible full texts, available
for further inspection. Two independent reviewers then reviewed these articles in depth with
regard to the fulfillment of the inclusion criteria. This inspection yielded a total of 78 articles
with 89 independent samples, including data of 5,973 participants. Most of the participants
identified themselves as female (60%) and were university students (64%) with an overall
mean age of 31.7 years.
All relevant data from these studies were summarized in a template to compute an
effect size for each dependent variable examined. Based on this summary data, standardized
mean differences between experimental groups exposed to robots varying in
anthropomorphism were calculated. Most studies reported a comparison of means. However,
the data sets were often incomplete, e.g., with no mention of means or standard deviations.
Cohen’s d was therefore chosen as a standard measure to describe the effect sizes. Note that
Cohen’s d represents an entire family of effect sizes, which makes it widely applicable for
different study designs (e.g., Cohen’s dav for within study designs). In addition, it can be
calculated from a wide range of statistical values received from different inferential statistical
methods (e.g., ANOVAs or t-tests) (52). By using this measure as a standardized measure of
effect sizes, we were able to compute a total of 294 effect sizes from the available data base.
Different effect sizes derived from the same samples and similar outcome variables within a
single study were averaged via the arithmetic mean. This was done so as not to overestimate
those studies in comparison to others.
Overall, this resulted in a total of 187 effect sizes. The final set of effect sizes was
then analyzed deductively by starting with the estimation of the overall mean effect of
anthropomorphism on human-related outcomes via a random-effects model. The calculated
mean effect size indicates the magnitude of the overall effect in terms of a standardized mean
difference. If the 95% confidence interval does not include “zero”, it can further be concluded
ANTROPOMORPHISM IN HRI
23
that this mean difference indeed represents a statistically significant effect that can be
expected to be replicated in further studies. To illustrate the effect size relative to its
confidence interval, a forest plot was created. The square reflects the effect size; the size of
the square shows the effect size weight with respect to the number of effect sizes included
and confidence intervals are shown by the length of the whiskers (see Fig. 3 for illustration).
In addition, the use of the random-effects model in this analysis also enabled us to assess the
degree of heterogeneity of effect sizes. In contrast to random sampling errors as a cause of
between-study differences, the heterogeneity estimates the true variation due to systematic
differences in study design, sample, and measurements used (53,54). To estimate the level of
heterogeneity, we used Qtests, which indicate whether or not a significant level of
heterogeneity is present, and 𝐼2, which represents the proportion of variance in the model that
can be explained by unaccounted factors (54).
The second and third steps involved conducting a subset meta-analysis for each of the
different superordinate outcome categories (i.e., perception, attitude, affect, behavior) and the
respective subdimensions. Again, the analysis, based on random-effect models, allowed for
assessing the mean effect sizes for different human-related outcomes and respective 95%
confidence intervals, which were again illustrated via forest plots. In addition, we estimated
the heterogeneity between the effects in different studies caused by hitherto unknown
moderators.
Finally, a variety of moderator analyses were conducted, based on the set of possible
moderators that had been identified a priori, i.e., the field of application, task relevance,
morphology, and robot exposure. For the overall model, mixed-effect models were used for
this purpose in order to include the moderators for diverting the directions or strength of the
relationship between a predictor and an outcome (53,55). Moreover, we estimated the
presence of heterogeneity via QMand the amount of heterogeneity via I2(in percent)
ANTROPOMORPHISM IN HRI
24
accounted for by the different moderators. For the superordinate outcome categories, we
abstained from using mixed-effect models, and limited our analysis to merely calculating the
mean effect sizes and 95% confidence intervals in order to identify whether an effect was
present at all. This somewhat constraint procedure was chosen because substantial
heterogeneity in the data set can considerably reduce the statistical power of tests in mixed-
effects models, which in turn would have increased the risk of failing to detect effects even if
they were actually present (56).
In an additional analysis, the current data set was used to examine the degree of
publication bias in the field of HRI. This was done because it has been suggested that
unpublished results might systematically differ from published ones, especially because non-
significant results may be submitted and published less frequently (57). Two different tools
were used to detect such possible asymmetry between effects reported by published versus
unpublished data, including a funnel plot to visually explore such bias and an Egger’s
regression test as an inferential statistical indicator. In the event of asymmetry, the two-sided
trim-and-fill method was used to correct the data set for publication bias. This method is used
to remove (trim) studies leading to asymmetry and replace the omitted studies (fill). It models
the data as if effect sizes and standard errors were symmetrically distributed as they should be
had all samples been unbiased estimators of the same mean value. As a result, the method
generates an estimate of the number of missing studies and an adjusted effect size of a meta-
analysis including the filled studies.
ANTROPOMORPHISM IN HRI
25
References
1. International Federation of Robotics, in World Robotics 2020: Industrial Robots
(2020;
https://ifr.org/img/worldrobotics/Executive_Summary_WR_2020_Industrial_Robots_1
.pdf), pp. 13–16.
2. International Federation of Robotics, in World Robotics 2020: Service Robots (2020;
https://ifr.org/img/worldrobotics/Executive_Summary_WR_2020_Service_Robots.pdf
), pp. 11–12.
3. L. Onnasch, E. Roesler, A Taxonomy to Structure and Analyze Human–Robot
Interaction. Int. J. Soc. Robot. 13, 833–849 (2021).
4. T. Fong, I. Nourbakhsh, K. Dautenhahn, A survey of socially interactive robots. Rob.
Auton. Syst. 42, 143–166 (2003).
5. C. Breazeal, Toward sociable robots. Rob. Auton. Syst. 42, 167–175 (2003).
6. V. Villani, F. Pini, F. Leali, C. Secchi, Survey on human–robot collaboration in
industrial settings: Safety, intuitive interfaces and applications. Mechatronics.55, 248–
266 (2018).
7. E. Matheson, R. Minto, E. G. G. Zampieri, M. Faccio, G. Rosati, Human–robot
collaboration in manufacturing applications: A review. Robotics.8, 100 (2019).
8. C. G. Atkeson, B. P. W. Babu, N. Banerjee, D. Berenson, C. P. Bove, X. Cui, M.
DeDonato, R. Du, S. Feng, P. Franklin, M. Gennert, J. P. Graff, P. He, A. Jaeger, J.
Kim, K. Knoedler, L. Li, C. Liu, X. Long, T. Padir, F. Polido, G. G. Tighe, X.
Xinjilefu, in IEEE-RAS International Conference on Humanoid Robots (IEEE
Computer Society, 2015), vols. 2015-December, pp. 623–630.
9. J. Luo, Y. Zhang, K. Hauser, H. A. Park, M. Paldhe, C. S. G. Lee, M. Grey, M.
Stilman, J. H. Oh, J. Lee, I. Kim, P. Oh, in Proceedings - IEEE International
ANTROPOMORPHISM IN HRI
26
Conference on Robotics and Automation (Institute of Electrical and Electronics
Engineers Inc., 2014), pp. 2792–2798.
10. G. Hoffman, C. Breazeal, in AIAA 1st Intelligent Systems Technical Conference
(American Institute of Aeronautics and Astronautics, Reston, Virigina, 2004;
http://arc.aiaa.org/doi/10.2514/6.2004-6434), p. 6434.
11. A. Clodic, E. Pacherie, R. Alami, R. Chatila, in Sociality and Normativity for Robots:
Studies in the Philosophy of Sociality, R. Hakli, J. Seibt, Eds. (Springer, Cham, 2017;
https://doi.org/10.1007/978-3-319-53133-5_8), pp. 159–177.
12. B. R. Duffy, Anthropomorphism and the social robot. Rob. Auton. Syst. 42, 177–190
(2003).
13. A. Waytz, J. Cacioppo, N. Epley, Who sees human? The stability and importance of
individual differences in anthropomorphism. Perspect. Psychol. Sci. 5, 219–232
(2010).
14. J. Złotowski, D. Proudfoot, K. Yogeeswaran, C. Bartneck, Anthropomorphism:
Opportunities and challenges in human-robot Interaction. Int. J. Soc. Robot. 7, 347–
360 (2015).
15. E. Phillips, X. Zhao, D. Ullman, B. F. Malle, in ACM/IEEE International Conference
on Human-Robot Interaction (IEEE Computer Society, 2018), pp. 105–113.
16. E. Roesler, J. I. Maier, L. Onnasch, The effect of anthropomorphism and failure
comprehensibility on human-robot trust. Proc. Hum. Factors Ergon. Soc. Annu. Meet.
64, 107–111 (2020).
17. S. Stadler, A. Weiss, N. Mirnig, M. Tscheligi, in 2013 8th ACM/IEEE International
Conference on Human-Robot Interaction (HRI) (2013), pp. 231–232.
18. L. Onnasch, E. Roesler, Anthropomorphizing robots: The effect of framing in human-
robot collaboration. Proc. Hum. Factors Ergon. Soc. Annu. Meet. 63, 1311–1315
ANTROPOMORPHISM IN HRI
27
(2019).
19. K. Darling, in Robot Ethics 2.0: From Autonomous Cars to Artificial Intelligence
(2017), pp. 173–188.
20. M. Mori, K. F. MacDorman, N. Kageki, The Uncanny Valley [From the Field]. IEEE
Robot. Autom. Mag. 19, 98–100 (2012).
21. K. S. Haring, D. Silvera-Tawil, K. Watanabe, M. Velonaki, in Proceedings of the 2016
International Conference on Social Robotics: Lecture Notes in Computer Science, A.
Agah, J.-J. Cabibihan, A. M. Howard, M. A. Salichs, H. He, Eds. (Springer, Cham,
2016), vol. 9979, pp. 392–401.
22. K. Darling, P. Nandy, C. Breazeal, in Proceedings of the 24th IEEE International
Symposium on Robot and Human Interactive Communication (RO-MAN) (IEEE,
2015), pp. 770–775.
23. G. Hoffman, O. Zuckerman, G. Hirschberger, M. Luria, T. Shani-Sherman, in
Proceedings of the 10th ACM/IEEE International Conference on Human-Robot
Interaction (HRI ’15) (2015), pp. 3–10.
24. T. Zhang, D. B. Kaber, B. Zhu, M. Swangnetr, P. Mosaly, L. Hodge, Service robot
feature design effects on user perceptions and emotional responses. Intell. Serv. Robot.
3, 73–88 (2010).
25. C. Bartneck, D. Kulić, E. Croft, S. Zoghbi, Measurement instruments for the
anthropomorphism, animacy, likeability, perceived intelligence, and perceived safety
of robots. Int. J. Soc. Robot. 1, 71–81 (2009).
26. T. Sanders, A. Kaplan, R. Koch, M. Schwartz, P. A. Hancock, The relationship
between trust and use choice in human-robot interaction. Hum. Factors.61, 614–626
(2019).
27. P. A. Hancock, D. R. Billings, K. E. Schaefer, J. Y. C. Chen, E. J. De Visser, R.
ANTROPOMORPHISM IN HRI
28
Parasuraman, A meta-analysis of factors affecting trust in human-robot interaction.
Hum. Factors.53, 517–527 (2011).
28. G. Charalambous, S. Fletcher, P. Webb, The development of a scale to evaluate trust in
industrial human-robot collaboration. Int. J. Soc. Robot. 8, 193–209 (2016).
29. C. Bröhl, J. Nelles, C. Brandl, A. Mertens, C. M. Schlick, in Proceedings of the HCI
International 2016 – Posters’ Extended Abstracts. Communications in Computer and
Information Science, C. Stephanidis, Ed. (Springer, 2016), vol. 617, pp. 97–103.
30. T. Nishida, Toward mutual dependency between empathy and technology. AI Soc. 28,
277–287 (2013).
31. X. Dou, C.-F. Wu, K.-C. Lin, S. Gan, T.-M. Tseng, Effects of different types of social
robot voices on affective evaluations in different application fields. Int. J. Soc. Robot.
(2020), doi:10.1007/s12369-020-00654-9.
32. M. R. Fraune, S. Sherrin, S. Sabanovi, E. R. Smith, in Proceedings of the 10th
ACM/IEEE International Conference on Human-Robot Interaction (HRI ’15) (IEEE,
New York, NY, USA, 2015), pp. 109–116.
33. S. L. Müller, S. Stiehm, S. Jeschke, A. Richert, in Proceedings of the 2017
International Conference on Social Robotics: Lecture Notes in Computer Science
(Springer Verlag, 2017; https://doi.org/10.1007/978-3-319-70022-9_59), vol. 10652,
pp. 597–606.
34. R. Reisenzein, Pleasure-Arousal Theory and the intensity of emotions. J. Pers. Soc.
Psychol. 67, 525–539 (1994).
35. J. A. Russell, A. Weiss, G. A. Mendelsohn, Affect grid: A single-item scale of pleasure
and arousal. J. Pers. Soc. Psychol. 57, 493–502 (1989).
36. S. Kuz, C. M. Schlick, in Proceedings of the 19th Triennial Congress of the IEA
(Melbourne, 2015; https://www.researchgate.net/publication/282665386).
ANTROPOMORPHISM IN HRI
29
37. M. Salem, F. Eyssel, K. Rohlfing, S. Kopp, F. Joublin, To err is human(-like): Effects
of robot gesture on perceived anthropomorphism and likability. Int. J. Soc. Robot. 5,
313–323 (2013).
38. J. Fink, in Proceedings of the 2012 International Conference on Social Robotics:
Lecture Notes in Computer Science (Springer, 2012), vol. 7621, pp. 199–208.
39. L. Onnasch, E. Roesler, A taxonomy to structure and analyze human–robot interaction.
Int. J. Soc. Robot. (2020), doi:10.1007/s12369-020-00666-5.
40. International Organization for Standardization, ISO 8737:2012 Robots and robotic
devices (2012), (available at https://www.iso.org/obp/ui/#iso:std:iso:8373:ed-2:v1:en).
41. M. Khoramshahi, A. Shukla, S. Raffard, B. G. Bardy, A. Billard, Role of gaze cues in
interpersonal motor coordination: Towards higher affiliation in human-robot
interaction. PLoS One.11, e0156874 (2016).
42. E. Roesler, J. I. Maier, L. Onnasch, in Proceedings of the Human Factors and
Ergonomics Society Annual Meeting (2020).
43. M. P. Mayer, S. Kuz, C. M. Schlick, in Lecture Notes in Computer Science (including
subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
(Springer, Berlin, Heidelberg, 2013; http://link.springer.com/10.1007/978-3-642-
39182-8_11), vol. 8026 LNCS, pp. 93–100.
44. J. Li, The benefit of being physically present: A survey of experimental works
comparing copresent robots, telepresent robots and virtual agents. Int. J. Hum.
Comput. Stud. 77, 23–37 (2015).
45. J. Wainer, D. J. Feil-Seifer, D. A. Shell, M. J. Matarić, in Proceedings of the 16th
IEEE International Symposium on Robot and Human Interactive Communication (RO-
MAN 2007) (IEEE, 2007), pp. 872–877.
46. K. E. Schafer, T. Sanders, T. T. Kessler, M. Dunfee, T. Wild, P. A. Hancock, in
ANTROPOMORPHISM IN HRI
30
Proceedings of the 2015 IEEE International Multi-Disciplinary Conference on
Cognitive Methods in Situation Awareness and Decision (CogSIMA 2015) (IEEE,
2015), pp. 113–117.
47. C.-C. Ho, K. F. MacDorman, Revisiting the uncanny valley theory: Developing and
validating an alternative to the Godspeed indices. Comput. Human Behav. 26, 1508–
1518 (2010).
48. A. Weiss, C. Bartneck, in Proceedings of the 24th IEEE International Symposium on
Robot and Human Interactive Communication (RO-MAN 2015) (IEEE, 2015), pp.
381–388.
49. P. Madden, L. Feingold, The Robots Are Here: At George Mason University, They
Deliver Food To Students. NPR (2019), (available at
https://www.npr.org/2019/04/07/710825996/the-robots-are-here-at-george-mason-
university-they-deliver-food-to-students?t=1620380296102).
50. D. Moher, L. Shamseer, M. Clarke, D. Ghersi, A. Liberati, M. Petticrew, P. Shekelle,
L. A. Stewart, P.-P. Group, Preferred reporting items for systematic review and meta-
analysis protocols (PRISMA-P) 2015 statement. Syst. Rev. 4, 1 (2015).
51. E. Roesler, L. Onnasch, D. Manzey, Same same, but different - A meta-analysis
regaring the consequences of anthropomorphism in human robot interaction:
PRISMA-P Protocol. available at osf.io/egtk6 (2020).
52. D. Lakens, Calculating and reporting effect sizes to facilitate cumulative science: A
practical primer for t-tests and ANOVAs. Front. Psychol. 4, 863 (2013).
53. S. Jain, S. K. Sharma, K. Jain, Meta-analysis of fixed, random and mixed effects
models. Int. J. Math. Eng. Manag. Sci. 4, 199–218 (2019).
54. H. Cooper, L. V. Hedges, J. C. Valentine, Eds., The handbook of research synthesis
and meta-analysis (Russell Sage Foundation, New York, NY, USA, ed. 3, 2019).
ANTROPOMORPHISM IN HRI
31
55. W. Viechtbauer, Conducting meta-analyses in R with the Metafor package. J. Stat.
Softw. 36, 1–48 (2010).
56. L. V. Hedges, T. D. Pigott, The power of statistical tests for moderators in meta-
analysis. Psychol. Methods.9, 426–445 (2004).
57. K. Dickersin, in Publication Bias in Meta-Analysis: Prevention, Assessment and
Adjustments, H. R. Rothstein, A. J. Sutton, M. Borenstein, Eds. (John Wiley & Sons,
Chichester, UK, 2006), pp. 9–33.
