Not All Robots are Evaluated Equally: The Impact of Morphological Features on Robots’ Assessment through Capability Attributions

As mentioned above, we previously developed a theoretical framework to explain differences in the evaluation of physically embodied robots in comparison to virtual instantiations ([13], Figure 1). Our core hypothesis is that evaluation and interaction outcomes (i.e., evaluation of an artificial entity, interaction evaluation, and behavioral responses) can be explained by perceived capabilities of the entity, which are influenced by various moderating variables related to the robot (e.g., physical embodiment, morphology), the person (e.g., prior experience, expectations), or the environment (e.g., the scenario in which the interaction occurs). Furthermore, we assume that perceived capabilities are associated with concrete physical features of robots and other artificial entities. The presence, absence, and combination of various body features such as eyes, arms, or wheels causes humans to attribute certain abilities to the artificial entity and anticipate according behaviors. For example, an entity with eyes is expected to be able to perceive the environment, an entity with wheels or legs is expected to be able to move. Hence, a favorable evaluation of a robot in a transportation task can be explained through the perceived capability for tactile interaction that is derived from the presence of manipulators that allow for tactile interaction. The relevant capabilities to evaluate artificial entities which we identified earlier [13], are perceived (nonverbal) expressiveness, shared perception, mobility, tactile interaction, and corporeality (i.e., the perception of an entity as physically existent in the real world, Figure 1). These somehow abstract capabilities are naturally related to visible body features such as arms, legs, or eyes. One can imagine that perceived capabilities vary according to the presence and combinations of body features in different morphologies (e.g., human-like versus animal-like). Morphology is thus hypothesized to moderate how capabilities of an artificial entity are perceived, regardless of the physical embodiment of the entity.

By morphology we refer to the form and appearance of an artificial entity, which often resembles other entities such as animals (zoomorphic morphology) or humans (anthropomorphic morphology). Related research that considered robots with various morphologies showed that morphology indeed has an impact. Phillips et al. [27] specified the physical human-likeness of robots based on morphological features. To this end, they have built a large database of anthropomorphic robots that are evaluated for the presence of various appearance features such as eyes, arms, legs, or wheels, as well as their measured physical human-likeness. Their analysis of 200 anthropomorphic robots stored in the database resulted in four major dimensions that determine physical human-likeness: (1) Surface Look, (2) Body-Manipulators, (3) Facial Features, and (4) Mechanical Locomotion. All dimensions can be expected not only to alter ratings in human-likeness of a robot, but also to affect assumed capabilities of a robot, and related evaluations. For instance, the presence of body-manipulators should increase the perceived ability to move and touch objects, or the presence of facial features should increase the perceived expressiveness of a robot. In line with this, Phillips et al. [27] remarked that physical features are related to human’s expectations of a robot’s capabilities, e.g., body manipulators evoke expectations about capabilities such as the transportation of objects, and facial features trigger expectation about a robot’s communicative functions. DiSalvo et al. [7] already demonstrated that differences in robotic heads determine how human-like people perceive these heads. Their investigation of 48 different robots revealed that the dimensions of the head and the number of visible facial features (e.g., mouth, nose, and eyes) mainly determine the perception of robots’ heads. Regarding the evaluation of robots with varying morphologies, a survey by Rosenthal-von der Pütten and Krämer [29] revealed that robots with similar design characteristics are equally evaluated with regard to likeability, threat, submissiveness, familiarity, human-likeness, and mechanicalness. Toy-like and zoomorphic robots (cluster 2 in their paper) were, for instance, evaluated as the most submissive, whereas android robots (cluster 4) were rated as most likable and human-like. Other researchers found similar effects of different morphologies on perceived intelligence [2], trust towards robots [30], perceived presence and on similarity to on oneself [1]. Attitudes toward robots also varied by robot morphology. For example, Thellman and Ziemke [34] showed that participants’ self-reported attitudes toward the social influence of robots varied significantly when participants were exposed to images of a robot with a semi-anthropomorphic or functional morphology. Regarding the core dimensions of social perception, recent research has shown that social robots’ ratings of warmth and competence are also predicted by certain body characteristics (e.g., eye-to-head ratio, visual acuity, and degrees of freedom) that can be associated with specific morphologies [3]. Furthermore, other investigations demonstrated that quite similar looking robots such as NAO and Pepper also evoke different reactions and evaluations [21, 35]. These differences might be related to differences in morphological details such as the possession of legs compared to wheels, the amount of fingers (NAO has three, Pepper has five fingers), or the size of the robots, and associated capabilities of the robots. However, the authors [35] did not further investigate what exactly caused the observed differences and whether perceived capabilities are an explaining variable.

In summary, previous work suggests that morphological differences are a crucial determinant for the perception and evaluation of artificial entities. Therefore, investigating the impact of varying morphologies is of great interest to understand how capabilities are determined by morphology. As outlined above, morphology of an artificial entity sets the potential for mutual perturbation with the environment [cf. 6]. Features incorporated in a morphology thus trigger associated skills, e.g., presence of manipulators triggers the ability to manipulate objects; presence of eyes triggers the ability to perceive the environment. We thus believe that the investigation into capabilities associated with different robot morphologies is important to gather a deeper understanding of what determines the reflexive, initial appraisal, i.e., the perception and evaluation, of artificial entities. In line with our framework [13], we hypothesize that different robot morphologies lead to different perceived capabilities and likewise evaluations (cf. Figure 1). To test this hypothesis, we utilized a set of standardized pictures of robots from the Anthropomorphic Robot Database² (ABOT: [27]), and presented them to subjects in two large online survey studies. Both studies address the research question of how differences in morphological features of physically embodied robots affect the perceived capabilities of a robot as an explanatory variable for the accompanying assessment. As a result, the analyses revealed that robots with similar morphological features can be summarized in clusters that differ in their assessment regarding body-related capabilities and overall evaluations. Moreover, results from our second study validated that varying evaluations of robots with different morphologies in terms of social attributes and acceptance can be explained through capability attributions (see Figure 12 for a simplified summary of all findings). Our findings expand previous work on robot perception by adding an explanatory variable to the puzzle of how initial evaluations of artificial counterparts originate, namely: perceived capabilities. This helps researchers to better understand why visible morphological features of social robots (or other artificial entities) trigger varying evaluations. In addition, the results are informative to engineers and designers who want to build robots with morphologies that match human expectations of a social robot’s capabilities.

Most (n = 39) of the used images were taken from the ABOT database [27], which hosts a total of 200 standardized images of anthropomorphic robots along with physical human-likeness score for each robot, and feature scores for the dimensions: body-manipulators, surface-look, facial features, and mechanical locomotion (Figure 2). To generate a diverse stimulus set, we followed three criteria: (1) for each dimension of anthropomorphic features we choose at least two robots that represented high, medium, or low scores, respectively; (2) following the four categories of robot morphologies as outlined by Fong et al. [11] (anthropomorphic, zoomorphic, caricatured, and functional) we selected at least five robots per category, plus a set of androids; (3) we also made sure that robots used in earlier studies that compared robots and virtual agents were included [see meta-analysis by 18]. Because of that, we added images of seven more robots that were not listed in the ABOT database (i.e., zoomorphic robots: Aibo, Karotz, Keepon, Miro, Pleo, and a KUKA gripper and a Roomba to represent functional robots). To ensure that the images are comparable to those from the ABOT database, we chose images that depicted the robot in full size in front of a white background.

For each robot we assessed several perceptual dimensions using three different instruments. The Embodiment and Corporeality of Artificial Agents Scale [EmCorp: 13] was used to measure participants’ bodily related perceptions of the robots’ embodiment and corporeality. The scale consists of 20 items which are rated on a fully labeled 6-point Likert scale (“strongly disagree” to “strongly agree”) and form the four factors (Shared)Perception and Interpretation (7 items, e.g., “The robot is able to perceive what I perceive”, Cronbach’s \(\alpha\) = .938), Tactile Interaction and Mobility (6 items, e.g., “The robot is able to carry objects”, Cronbach’s \(\alpha\) = .846), (Nonverbal) Expressiveness (4 items, e.g., “The robot is unrestricted in its facial expressions”, Cronbach’s \(\alpha\) = .862) and Corporeality (3 items, e.g., “The robot is existent in the real world”, Cronbach’s \(\alpha\) = .661). Besides, we used the Robotic Social Attributes Scale [RoSAS: 4] as an evaluative output variable to test whether different morphologies evoke different evaluations in terms of warmth (6 items, Cronbach’s \(\alpha\) = .935), competence (6 items, Cronbach’s \(\alpha\) = .921), and discomfort (6 Items, Cronbach’s \(\alpha\) = .902). The items were rated on a 9-point Likert scale. To see how physical human-likeness of different robot morphologies is linked to perceived capabilities, we further measured perceived (physical) human-likeness of the robots with a single item slider from 0 to 100, as introduced by Phillips et al. [27]. To control for individual differences, we further asked for participants’ attitudes towards robots using the Negative Attitudes towards Robots Scale [NARS: 24, Cronbach’s \(\alpha = .88\)]. The scale consists of 14 items which are rated on a 5-point Likert scale (“Strongly Disagree” to “Strongly Agree”). In addition, we considered participants’ general tendency to anthropomorphize (i.e., apply human characteristics) objects as a determinant to perceive robots as more human-like, and thus to perceive robot capabilities differently. We used the Anthropomorphism Questionnaire by Neave et al. [23], which consists of 20 items rated on a 7-point Likert scale (“Not at all” to “Very much so”, Cronbach’s \(\alpha = .95\)).

On the starting page of the online-survey, participants were informed that they were being asked for their opinions on various robots. Participants first completed the NARS and subsequently viewed pictures of three robots which were each evaluated on the EmCorp-Scale, the RoSAS, and the physical human-likeness slider. Each robot was separately presented at the top of the page with a height of 500px. The scales were presented below the picture of the robot (one page per scale to ensure that the picture was visible while rating the items). The different robots were presented in random order. Their presentation time was not limited. Participants regarded them for as long as they wanted. Afterwards, participants completed the Anthropomorphism Questionnaire. Finally, we asked participants to briefly describe the three robots they have seen previously and give demographic information (age, gender, prior contact to robots). Participants’ descriptions of the robots were screened as a data quality check to eliminate possible inattentive users or bots. Every participant viewed and evaluated three robots. To guarantee randomized but also equal distribution of evaluators to robots, for each participant three robots were randomly drawn from the pool of robots until the urn of 46 robots was “empty”. Then the urn was filled up again to be drawn from.

A total of 673 participants were recruited via Amazon Mechanical Turk. Those participants who provided nonsense descriptions for the three robots at the end of the survey were excluded (e.g., 1-2-3; nice-good-well; karatas-mamibot-grimlock). Finally, data of \(n=510\) were analyzed (age 18–76 years, \(M= 37.41\), \(SD=12.17\)), 245 female, 260 male, 1 queer, 1 trans-gender, 2 nonbinary and 1 who did not indicate sex at all). Their prior contact with robots was mixed: \(54\%\) never had contact with robots, \(22\%\) had contact with chat-bots like Siri and Alexa, \(14\%\) mentioned that they encountered a robot at least once, while \(5\%\) have contact with robots at work. Only \(2\%\) indicated to own a robot, and the remainder did not comment on the question. Participants attitudes towards robots were overall moderate (NARS overall: \(M=2.85; SD=0.77\) on a 5-point scale). Their overall tendency to anthropomorphize things was low (\(M=2.62, SD=1.31\) on a 7-point scale).

We hypothesize that different robot morphologies cause distinct capability attributions (EmCorp) and thus evaluations (RoSAS) that are not a result of differences in the physical embodiment. We further assume that the capability dimensions are linked to specific body features, and that thus specific features are more important for one capability than another, e.g., the presence of wheels is important for the perception of mobility, whereas it is less important for shared perception. To identify robot morphologies that evoke similar capability perceptions, we clustered the robots according to their assigned capabilities. For that purpose, we ran an agglomerative hierarchical cluster analysis with squared Euclidian distance measures using Ward’s minimum variance method [38] on 46 cases (i.e., 46 robots with different morphological features) on the basis of the calculated mean values per EmCorp-subscale. Data were standardized by converting them to z-scores. The six-cluster solution is most reasonable, because it marks the last considerable change in agglomerated coefficients (Table 2) and the dendrogram in Figure 3 also supports this solution. The following sections describe each cluster separately.

For the validation of the clusters we ran a MANOVA with the six clusters as fixed factor and the four EmCorp-subscales as dependent variables. As depicted in Table 3, the average perceived capabilities vary significantly between the clusters. To disentangle differences between single clusters (C1–C6), we calculated posthoc comparisons (LSD) for all capabilities. The results are reported per capability in the following subsections. For a better comprehensibility, Figure 10 gives a visual summary of the order of mean cluster ratings per EmCorp capability.

We explored the evaluations in terms of warmth, competence, and discomfort as well as the assigned physical human-likeness between the clusters (Table 3).

In this first study, we aimed to investigate how differences in morphological features of physically embodied robots affect the perceived capabilities of a robot and the accompanying evaluation. The results show that the capabilities humans infer from the mere appearance of robots on photographs depend on morphological features in the robots’ design. Furthermore, analysis reveal that different morphological features also lead to differences in the evaluation of robots in terms of perceptual measures such as human-likeness, warmth, competence, and discomfort. In addition, we demonstrated that the attribution of different capabilities predict differences in these perceptual measures. Based on these findings, the next step was to further test the assumptions of the EmCorp Framework, namely the mediating effect of attributed capabilities on the relation between robot morphologies and differences in perceptual output variables (see Study 2).

In Study 2, we reduced our stimulus set down to 20 robots. The selection of robots was based on the six clusters identified in Study 1, with the aim to map the variety of each cluster in terms of differences in morphology. This resulted in six groups of stimuli consisting of 3 to 4 robots per group (see robots labeled ’\(\#s2\)’ in Figures 4–9). Same as in Study 1, most robot-images (16 images) were taken from the ABOT database [27]. For the four robots (Keepon, MiRo, Roomba, and Kuka) that were not represented in the ABOT database, we used the same images as described in Study 1.

Similar to Study 1, we assessed several perceptual dimensions using subjective measures. To measure participants’ bodily related perceptions of the robots’ embodiment and corporeality we used the EmCorp Scale [13] (subscales: Corporeality, Cronbach’s \(\alpha\) = .635; Nonverbal Expressiveness, Cronbach’s \(\alpha\) = .857; Tactile Interaction and Mobility, Cronbach’s \(\alpha\) = .859; Shared Perception and Interpretation, Cronbach’s \(\alpha\) = .920). We used the RoSAS [4] as an evaluative output variable for the different robot morpholgies (subscales: warmth, Cronbach’s \(\alpha\) = .926; competence, Cronbach’s \(\alpha\) = .912; discomfort, Cronbach’s \(\alpha\) = .899). We assessed perceived (physical) human-likeness of the robots with a single item slider from 0 to 100, as introduced by Phillips et al. [27]. To control for individual differences, we further asked for participants’ attitudes towards robots using the NARS [24] (Cronbach’s \(\alpha\) = .857) and their general tendency to anthropomorphize using the Anthropomorphism Questionnaire by Neave et al. [23] (Cronbach’s \(\alpha\) = .961). Furthermore, we added 4 items assessing participants’ general acceptance of the evaluated robot [13, 14] as a second evaluative output variable. The four items asked for, e.g., whether the robot arouses curiosity about the topic of robots, or whether participants can imagine to perform certain tasks with the help of the robot, on a 5-point Likert Scale (“fully disagree” to “fully agree”, Cronbach’s \(\alpha\) = .817).

The design of the second survey followed that of Study 1, except that participants only evaluated one robot and had to fill in the four general acceptance items after the human-likeness rating. We also changed the control questions to three individual questions for each robot regarding their appearance to automatically delete participants who did not consciously complete the survey.

In total, 593 participants completed the survey via MTurk and answered the control questions correctly. After screening the remaining datasets, we excluded data by 7 more participants since they showed suspicious answering patterns (i.e., equal ratings for all questions). Thus, the final dataset consists of n = 586 participants between the age of 18 to 74 (M = 36.40, SD = 11.43), 285 female, 300 male, 1 agender. With regard to prior contact with robots, 48% did report not having contact to robots so far, 15% mentioned that they encountered a robot at least once, some had contact to chat-bots (27%), and the minority had contact with robots at work (1%) or actually own robots (4%). Their negative attitudes towards robots were moderate (NARS overall: \(M=2.98, SD=0.93\); on a 5-point scale), and their tendency to anthropomorphize things was low to moderate (IDAQ overall: \(M=3.13, SD=1.63\) on a 7-point scale). The sample was hence similar to that of Study 1.

Robot clusters were built based on the results of Study 1 (see robots labeled “\(\#s2\)” in Figures 4–9). We therefore added a new variable that holds the number of the assigned cluster. Based on the mapping of the robots to the clusters we calculated mean ratings for all dependent measures (see Measurements) per cluster for further analyses (Table 3, Study 2).

To validate the clusters we ran a MANOVA with robot cluster as fixed factor and the four EmCorp-subscales as dependent variables. Again, as depicted in Table 3 (Study 2), the average perceived capabilities differed significantly between the clusters.

As in Study 1, we analyzed the evaluations in terms of warmth, competence, discomfort, and the assigned physical human-likeness between the clusters. Additionally, we considered the overall acceptance of the robots in the clusters to compare the results with previous findings [13] (Table 3).

The major aim of Study 2 was the test of the mediating role of perceived capabilities on evaluative outcomes as proposed in [13]. Therefore, we ran multiple mediation analyses for the RoSAS-subscales and the acceptance measure as the dependent variables.⁴

For each analysis, we opposed two clusters, respectively, as independent variable that turned out to differ significantly in the posthoc comparisons (e.g., C2 vs. C1). As mediators, we included all four EmCorp-subscales. The detailed results for each mediation analyses are available as supplementary material. For a better understanding of our findings, a short summary of the observed patterns can be found in Table 5. Overall, the results reveal that different acceptance ratings of robot clusters can primarily be explained through perceived Tactile Interaction and Mobility, i.e., robots that are evaluated as more capable to touch and move are more accepted in general. For some comparisons (C2 vs. C1, C5 vs. C1), (Shared) Perception and Interpretation further explained higher acceptance of robots. A similar pattern was observable for competence ratings: All pairwise differences were mediated by Tactile Interaction and Mobility, and additionally, differences between C5 and C1, as well as C4 and C5, were further mediated by (Shared) Perception and Interpretation. The perceived warmth and discomfort individuals assigned to the robots were explained by (Nonverbal) Expressiveness in the majority of the cases. Differences in warmth were also often explained by (Shared) Perception and Interpretation. This was also true for discomfort when C5 and C6 were opposed. None of the differences was mediated through Corporeality. (A simplified visualization of which capability mediates the effect on which outcome variable can be found in Figure 12).

Our second study supported the findings from Study 1. As predicted in the EmCorp framework, robots in clusters with similar morphological features differ significantly in body-related capabilities as measured by the EmCorp-Scale. Additionally, mediation analyses revealed that robots’ perceived body-related capabilities further explain different assessments of robots in terms of the robot’s acceptance and robotic social attributes (i.e., warmth, competence, and discomfort). Our findings show that the sub-scales Tactile Interaction and Mobility and (Shared) Perception and Interpretation explain ratings of acceptance and competence, whereas (Nonverbal) Expressiveness and (Shared) Perception and Interpretation explain warmth and discomfort ratings. Although variance in Corporeality ratings was visible for the different robots (Table 3), Corporeality was not the distinguishing factor to explain different assessments of acceptance, warmth, competence or discomfort. This is perfectly plausible, as all stimuli were physically embodied robots (e.g., as compared with virtual characters; cf. [13]) and corporeality might hence be a neglecting factor for these comparisons.

Cluster 1 - Caricatured robots with eyes. Their abstract, comic-like appearance might explain why robots in this cluster were rated the least capable on all dimensions. The comic-like morphologies might trigger toy-related mental models and thus lower peoples’ expectations for a robot’s capabilities in general. The absence of manipulators and legs or wheels explains why they were perceived as incapable to touch, transport objects, and move. Although the majority of the robots possesses eye-like features, they score low on (Shared) Perception and Interpretation as well as on Nonverbal Expressiveness. This might be due to the aforementioned associations with toys that are predominantly static and incapable of perception. However, the majority of the robots in this cluster are in fact quite expressive, e.g., Keepon can move and make sounds. These dynamic features of the robots were of course not discernible from the pictures, especially when participants had not encountered the robots in real-live or video before, which is a limitation of the static approach. Corporeality was lower than in the other clusters, but still moderate (3.18 on a 6-point scale). It seems plausible that abstract appearances that resemble comic characters are rated as less “real”.

Cluster 2 - Full body, humanoid, and android robots. Robots with a full humanoid body, including legs to walk, arms and hands to grasp, a torso and a head with facial features, were rated the most sophisticated regarding all capabilities. Although only the upper body of the android robots were depicted in three cases, the robots were still assumed to be highly capable. The same could be observed for anthropomorphic, expressive robot heads (C4). It seems that humans assume high capabilities only based on the sophisticated design of visible parts. Robots with a humanoid or android shape were also perceived as most (physically) human-like. Similarity to the human shape thus appears to be an indicator of high capabilities: if it looks like a human, it must be capable to do what humans do. Eliciting such high capability expectations through design might backfire, if the robots’ actual capabilities do not match their perceived ones. The robots were further rated as the “realest” robots (Corporeality), which suggests that humans’ definition of what makes an entity corporeal is not necessarily the differentiation between is real (robot) or is not (virtual character). Instead it seems related to a degree of realism in the kind of an entity, e.g., human-like entities are more real than functional objects, which are, however, still more real than cartoon-like entities, at least with regard to robot embodiment.

Cluster 3 - Anthropomorphic but functional robots with grippers. While robots in this cluster were perceived as capable for tactile interaction, mobility, and corporeality as were the full body, humanoid and android robots (C2), they were evaluated less capable of shared perception and expressiveness. The comparison of the robots heads shows that robots in C2 possess more surface features and more sophisticated faces (in the case of the androids), whereas the majority of the robots in C3 only have eyes, and some eyes and a mouth. Hence, possessing more body features that allow for perception and resemble human faces increase the perceived capability for shared perception and interpretation. This is an important information, since not all facial features of robots are designed for perception. Nevertheless, they do trigger expectations. On the contrary, sensors that allow for visual or sound perception might be invisible to humans but still incorporated in a robot. If it is important that the user has knowledge about the robots capabilities it might be worth to considering features that are related to humans’ perceptions of the indicated capabilities, e.g., microphones in ear-like features, or speakers in mouth-like features of a robot.

Cluster 4 - Anthropomorphic, expressive robot heads. These robots are rated moderately high on the dimensions that can be related to facial features, i.e., shared perception and nonverbal expressiveness. Despite the lack of a torso, arms, legs, or wheels, their mobility and corporeality was still rated moderately, but low in contrast to the other clusters, except for caricatured robots (C1). This supports the assumption that not only the presence of a feature (e.g., eyes) determines a capability, but beyond that the human-likeness of the feature’s design seems to produce a Halo effect and might affect judgments for other capabilities. Therefore, expressive human-like robot heads (C4) were rated as more capable than comic-like robots with eye-like features (C1) to perceive their environment. Moderate, and not high, ratings in (Nonverbal) Expressiveness could be a result of the robot’s restriction in gestures, since the items for expressiveness cover gestures as well. However, it is surprising that the same robots were rated as capable to move and manipulate objects although they do not possess a torso or legs. It seems as if the sophisticated surface looks of these robots trigger mental completion processes: human-like heads belong to full bodies, although they are not present on the pictures. This would be the same for a portrait picture of a human, where it is clear that the human has a body, although it is not visible. It is vital to mention that such misconceptions would not arise in actual encounters where the absence of a torso would be inevitably visible to the viewer.

Cluster 5 - Mobile robots with facial features but no manipulators. While the robots in C5 look quite similar as those in C3, they are rated lower regarding movement, touch, and expressiveness. This is surprising, since most of them are capable to move. However, their bodies do not resemble a human shape. They are animal-like or abstractly formed, move on a pedestal, or it is unclear from the picture whether the robot has wheels or not (e.g., Karotz). Together with the abstractness of some facial features, this reason could also explain the low ratings on (Nonverbal) Expressiveness. It has to be noted that the perception of these robots might change in a real interaction where the features that seem static on the picture are actually quite expressive (e.g., mouth and eyebrows of iCat). This positive discrepancy between the expected low capabilities and the actual higher extent might result in highly positive evaluations of these robots in actual encounters, presumed that an initial impression has been formed based on the static appearance before.

Cluster 6 - Functional robots with grippers or wheels. The function of these robots can be more easily derived from their appearance, e.g., grasp and manipulate objects. The robots are either wheeled (Roomba, Clocky) or possess manipulators (Kuka, Franka Emica) resulting in overall moderately high ratings on Tactile Interaction and Mobility. They were perceived as restricted in nonverbal expressiveness since they neither possess facial features, nor a body that could show gestures. Also, their ability to perceive the world was considered low, perhaps due to the lack of mammalian-like features indicative of perceptual abilities. Of course, many of these robots have sensors that allow for perception, e.g., collision detection. However, due to the absence of visible features, these capabilities are not as salient for these morphologies compared to the others in our study. Higher ratings in corporeality compared to more cartoon or toy-like robots as in C1 can be explained as above, i.e., industrial robots or vacuum cleaner robots can be regarded as tools that are more “real” than fictional cartoon characters. Conclusively, humans have rather low expectations for functionally-looking robots in terms of social capabilities. However, if these robots should become more interactive and integrated in social contexts, it is a socio-technical challenge to equip these robots with cues that allow human users to form adequate mental models about their capabilities which might go beyond pick-and-place.

As outlined in the EmCorp-framework [13] and supported by the present studies, the assessment of robots can (to some extent) be explained through the expected capabilities humans attribute to them based on their morphological appearance. How capable a robot is perceived to move, touch, see, understand, and express itself determines how warm and competent a robot is rated, how much discomfort humans assume in its presence, and finally, how likely people are to accept it.

First, making judgments about robots’ assessment on the basis of static pictures is a major limitation of the present work. Albeit, the short exposure to a picture was the most controllable possibility to expose participants to a large variety of robots. Nonetheless, we have to admit that the exposure to a picture of a robot on a screen does not equal standing in front of a robot, even if it is not moving at all. Also, with regard to the robotic heads (C4) that were rated as capable to move and touch, it is vital to mention that such misconceptions would not arise in actual encounters where the absence of a torso would be inevitably visible to the viewer. Hence, direct comparisons of capability ratings based on pictures and live exposure to the same robot will be highly informative in the future. In addition, comparisons of just observing a co-present robot versus directly interacting with it should be taken into account. As summarized above, humans seem to assume high capabilities based on the design visible in the pictures. Whether these (high) expectations endure live encounters remains an open question. Furthermore, initial expectations of a robot’s capabilities can easily change based on interaction experiences. For example, if a robot with visible features that imply vision (e.g., eyes) does not respond to motion in the environment, this should change perceptual expectations. Eliciting high capability expectations through design can hence backfire, if the robots’ actual capabilities do not match their perceived ones (e.g., [26]). On the one hand, this suggests that the evocation of certain capability expectations through a robot’s morphology should be taken seriously. On the other hand, it suggests that actual interaction experience can overwrite initial impressions in a good (“better as expected”) as well as a bad sense. Second, conducting online surveys with MTurk comes with limitations. We took several steps to ensure data quality (check questions at several points in the survey, in-depth analyses of answers). However, potential inattentiveness of respondents remains a problem. Furthermore, the rating of the robots was based on static pictures that did not reveal the size, sound, or movements of the robots. Because of that the actual co-presence of the robots, which is one key feature of physically embodied robots [18, 37], was not given. Although we believe that no actual interaction with the robots is necessary to answer the question whether morphology (which can be regarded as a stable and static feature) impacts capabilities inferred from robot appearance, it remains unclear whether the findings still apply to the same extent in dynamic environments (video or actual interaction). Third, some findings from the presented studies suggest to rethink the combination of the theoretically separated capabilities Tactile Interaction and Mobility into one sub-scale. This becomes especially visible with regard to the high ratings in Tactile Interaction and Mobility assigned to robots in cluster 5 that do not possess body manipulators to touch or carry objects. However, these robots have wheels or four legs that make them mobile. Last, the clusters in their current form subsume very different categories of robots with similar morphological features. For instance, Cluster 2 consists of full body, humanoid and android robots such as NAO and Geminoid (cf. Figure 5), which might evoke quite different reactions. Hence, further subdivisons, e.g., into humanoid and android robots, could allow for more fine grained comparisons within the presented clusters.

Our findings expand previous work on robot perception by adding perceived capabilities as an explanatory variable to untangle the assessment of artificial entities with varying morphology. Our results reveal that initial exposure to visual cues of robots incorporated in their morphology trigger certain expectation about their body-related capabilities, i.e., the capabilities to move in space and to touch objects, to express oneself, to share perceptions, and to be corporeal. These capability related expectations further explain why robots with different morphologies receive varying assessments in terms of acceptance but also socially relevant evaluations. This knowledge informs researchers, on the one hand, to better understand why visible morphological features of artificial entities or social robots trigger varying evaluations. On the other hand, the results are relevant to engineers and designers that aim at building robots with morphologies that match human expectations of robot’s capabilities. Regarding previous work, our results suggest that conflicting findings could be to some extent caused by different capability attributions that were caused by different morphologies of the used robots (e.g., humanoid, full body robots [8, 12, 15, 16] or zoomorphic, toy-like robots [14, 17, 19]). As HRI researchers, we are aware that viewing pictures of robots is different from experiencing the co-presence of a social robot, its size, its movement, and the accompanying motor sound. An important open question for future research is thus: Which role do morphological differences play in the assessment of robots during actual HRI? Do visible static features significantly alter how humans appraise a robot and how they react towards it? Or do static features become less salient and thus less important during live HRI when the attention is directed towards the task and the performance of a robot? Are initial impressions, as we tried to infer from the ratings of pictures, actually consequential for humans’ decision to approach or avoid a robot in real life? Can these initial expectations predict how people will behave in front of a robot? An important step to answer these questions will be the systematical variation of morphological features in live interactions. This can be realized through comparative studies that utilize different robots, e.g., robot from different clusters as presented here. Or, through variations of the visible features of one robot, e.g., by covering parts of a robot (cf. [5], or dismounting grippers (if possible). Virtual and augmented reality applications further seem to be a fruitful test bed to study the impact of morphological differences in live interactions. In addition, research on the role of robot identity and its relationship to possessing a single or multiple bodies suggest that people are able to recognize the same robot identity within a new body if certain cues such as the eyes or the voice are kept equal [20]. More research in this realm is necessary to understand whether morphological cues associated with capabilities are affected by dynamically changing robot identities. For example, it seems plausible to assume that the same robot identity in another body knows (cognitive capability) the same information, whereas it is not plausible to assume that it will be able to transport objects if the new body is not equipped with manipulators (physical capability). How this discrepancies might affect the overall assessment of a robot should be considered in future work. Furthermore, it remains open whether perceived capabilities, such as those related to embodiment (EmCorp-subscales), are stable perceptions, or whether perceived capabilities can change over time. As subsumed in the framework (Figure 1), it can be expected that contextual factors and enabled behaviors in live interactions will render certain capabilities more salient. For instance, one can expect that performance shortcomings such as dropping a cup might result in lowered ratings of the robot’s capability for tactile interaction, although it has been initially assumed to be high due to the presence of grippers. The same can be expected for a robot that has eyes that do not include vision sensors which allow for reaction to visual stimuli. Moreover, contextual factors can render certain capabilities more important than others. In a task that includes the manipulation of physical objects, like the towers of Hanoi task [cf. 14, 37], shared perception and reaching out to manipulate objects is more relevant than nonverbal expressiveness. Thus, an industrial robot such as the Kuka Gripper might be perceived as more capable for the task than a robot with a highly realistic face but no manipulators (e.g., Flobi). Future studies can expand this line of research by including capabilities beyond body-related ones, e.g., cognitive or communicative capabilities, which might also be linked to visible features of social robots.