The embodiment of robots, specifically the face in the case of humanoids, is a sensitive issue, since studies on social robots involve complex socio-demographic factors such as age, gender, culture and others [1
Appearance has an important effect on anthropomorphism, as humans, for instance, commonly have different expectations from mechanical-looking robots and more human-looking robots [2
], and attribute different tasks among different kinds of appearances [3
]. Acceptance of humanoids can also be influenced by cultural factors between western and Asian countries [4
]. Cultural backgrounds has been proven to influence the perception and intention to use a robot as an instrument [5
], and preferences about their role in society [6
], and the degree of anthropomorphism [7
] and mind attribution [9
From the appearance and behaviour, the first impression that we have of a robot matters in the same way as with humans, as depending on time, motivation, and effort in social judgments, mental shortcuts are used when forming an impression of others. People seem to rely on simple rules of thumb and engage in automatic information processing to come to an estimate about a person or a non-human entity [10
]. To do so, humans take into account key features that indicate social category membership of a person or entity, relying on visual cues that indicate age, gender, or ethnic background of a person. Similarly, from physical and functional cues humans can infer a robot “personality” which may impact human–robot interaction [11
Subtle visual cues that indicate the gender of a robot may activate peoples’ gender stereotypical knowledge structures, which lead to biased expectations regarding the robot's abilities [14
]. For instance, the experiment in [11
] has documented that indeed, people perceive a long-haired robot as more female than a short-haired counterpart, and accordingly, judgments regarding these “feminine” vs. “masculine” robot prototypes turn out to be more gender-stereotypical, with the robots being differentially rated in their suitability for gender-stereotypical tasks.
Previous research starts from the gender-stereotypic response toward computers: in [15
], synthetic voices were investigated and a dominant female voice was found to be perceived more negatively than a male one in a computer, which was also taken more seriously in the case of praise. Another experiment on a gendered voice was performed in [16
An investigation on children’s preferences [17
] concluded that, according to the design implications of their research, robots should have a female gender and be brightly coloured for positive behaviours. This is because female gender was associated with positive robot traits such as happiness and friendliness, unlike male gender, which was not distinguishable for positive and negative robot traits.
Similar to human–human interaction, many factors influence the persuasiveness of a robot, including appearance, style and content of communication, and non-verbal behaviour, and gender plays a fundamental role.
Siegel et al. [18
] showed that men were more likely to be persuaded to donate money to a female robot, as they felt more trust and engagement. In general, participants tended to rate the robot of the opposite sex as more credible, trustworthy, and engaging. In a recent study about persuasion [19
], robot gender was manipulated by changing its voice and name: the results found an influence on preferences in making donations, as the female robot received more donations from the participants. In a study on the robot guard RobotMan [20
], its two tasks of security and guidance were associated with gender. The expression of the eyes and the tone of the voice influenced not only the perceived gender-related traits, but also physical attributions like body size as well as likeability.
These results suggest that we should consider manipulating the gender when designing robots for persuasive applications, which can be desireable for specific tasks.
1.1. Female Robots
In science fiction and in the entertainment industry, gendered robots have been portrayed quite often, mirroring the role of females in society. Much discussion has been raised over the risk of de-humanising women as a process of dehumanising female robots, and on the other hand, new terms such as gynoid and fembot have been created.
Meanwhile in robotics research, gendered robots are not too common.
In real life, some jobs are more typically associated with female gender, such as the receptionist, which is a role in which robots can be employed. In some countries like Brazil, more than 80% of receptionists are women [21
]. In the survey described in [22
], 38% of the participants expressed the explicit preference for a female robotic receptionist, compared to 6% who explicitly indicated a male robotic receptionist. A “roboceptionist” was first employed in Carnegie Mellon University, and it featured a conversational agent called Valerie displayed on a screen [23
]. More recent works include the android Otonaroid and Saya [24
]. In [22
], a female agent was used as a receptionist in a hi-tech centre. The research in [25
] investigated the effect of voice pitch on the evaluation of a social robot receptionist, showing that a female high-pitched voice was perceived significantly more attractive in terms of behaviour and personality compared to a calm, low-pitched voice.
The guidance robot IOmi [26
], although in a stylised design, features a female appearance. Its torso has a yaw twisting motion supported by a cylindrical bearing to distribute the load throughout the skeleton. Such a cylinder is contained in a lower body that resembles a long skirt and that also contains the large-wheeled platform. This is a typical example of matching form with fuction, as the skirt is bigger at the bottom. Similarly, IOmi’s head has one degree of freedom in the hair bun. As it is common for women to work having their hair tied, the top knot can symbolise that the robot is on duty and ready to provide services. At the same time, it can move, acting as a communication mean, effectively matching form with function.
The robot Flobi [27
] is also worth mentioning, as it is a robotic head that was projected in a modular way, in order to be easily characterised as male or female, through the use of different sets of hair and mouth.
The present research will shed more light on the role of visual gender cues in robots, as these are particularly crucial for product design.
Thus, we aimed to investigate the perception of gender in robot design, focusing specifically on the proportion between chest, waist, and hips to indicate robot gender. By manipulating body proportions, it will be possible to influence user perception of the robot gender, as documented in previous research, and in particular for humanoids, enhance the feeling of intimacy. The novelty in the present paper consists in the manipulation of body proportions, which although extensively studied in anthropometrics, has never been done before on pictures of robot bodies.
We explore the chest–hip ratio and waist–hip ratio as subtle visual cues that are utilised to form gender-based impressions of novel robot prototypes. In anthropometrics, shoulder-to-hip ratio (SHR), chest-to-hip ratio (CHR), and waist-to-hip ratio (WHR) are typical indicators of human body types. We reviewed a number of studies that inspired the stimuli we used which were produced by a professional designer: in our study, a figure of a generic robot was used and adapted to different combinations of CHR and WHR.
], shoulder-to-hip ratio was measured for human models: an average SHR of 1.39 was found for male models and a SHR of 1.23 was obtained for females. In comparison, average males had a SHR of 1.21, while average females had a SHR of 1.08. In their experiment [28
], the authors used stimuli that were characterised as having a SHR of 1.2 for males and a SHR of 1.05 for females.
Chest-to-hip ratio represents a similar metric which does not take into account shoulder width. It is used in product design: for example, in the female doll “Bratz”, the CHR was 0.82. This CHR score is similar to that of an adolescent female body [29
Furthermore, the difference between upper and lower body is typically used in symbolism for public toilets: a triangle can be used to represent females because of its resemblance with a skirt. These kinds of symbols are widespread and recorded by institutes such as the American Institute of Graphic Arts (AIGA). Reversed triangle and circle can also be used as symbols for men and women, respectively. They recall broader shoulders of men and rounded bellies of pregnant women [30
]. Squares are also typically associated with males and circles with females [31
Waist-to-hip ratio is also an important measurement related to gender: higher testosterone levels in men stimulate fat deposits in the abdominal region while inhibiting fat deposits on the hips and thighs: healthy adult men generally have higher WHRs than women, with values ranging from between 0.85 and 0.95 [32
]. For women, the ideal value is typically 0.70, while for men the ideal is approximately 0.90. In [28
], the measurements of WHR resulted in a range of 0.70–0.90 for women and 0.90–1.10 for men.
Across cultures, a low waist-to-hip ratio represents a predictor of female body attractiveness [29
]. Furthermore, the existence of a cross-cultural difference in the preferences regarding body fat and thinness is widely known [33
]. In particular, plump females are typically linked with femininity and fertility among many non-western societies. At the same time, in some Asian countries, Confucianism may have an impact on food consumption, thus influencing the thin ideal of female body [34
]. However, [33
] also indicates that the largest differences in body size ideals are much less present due to modernisation and westernisation.
Since the last century, ideals in the west have been historically oscillating between the desirability of a round, plump body, and sexual attractiveness without maternity [33
]. In the nineteenth century, the use of waist training and corsets allowed some women to achieve 0.50 WHR. The Bratz doll is hyper-feminised in such way, featuring a WHR of 0.52 [29
]. Designers exploit product proportions and arrangement to evoke associations: a pronounced waist is typically used to indicate female gender [35
The robot IOmi [26
] features a pronounced feminine CHR and WHR in order to increase the perception of a female robot. Although it is not possible to calculate accurately since the exact position of the hip of IOmi is not defined, its CHR is around 0.6 and its WHR is around 0.5.
] features a CHR of approximately 1.30 and WHR of approximately 0.45. The large chest, necessary for the touch screen, contrasts with the small waist, and this may produce a mixed response in terms of visual and social perception.
1.3. The Present Experiment
In order to further study this topic, we conducted a study on the perception of stimuli consisting in robot body shapes, in which we manipulated CHR and WHR.
In the present experiment, our hypothesis is that the two body measurement ratios will impact the perception on gender of robots, namely: a higher CHR suggesting a male gender and a lower CHR a female gender; and a higher value of WHR (close to 1) indicating male gender and a lower WHR, a female gender.
As reported in the Introduction, background culture may play an important role in this kind of studies. For this reason, we performed a cross-cultural comparison between Peru and Japan, also including participants from other countries who were collected in the two countries, for a broader east–west comparison. Peru was the country from where this research originated subsequent to the development of the female robot IOmi [26
]. Japan was chosen as a country for the main comparison because while being one of the world leading countries in terms of development of humanoids, it is very distant in terms of culture, and the role of women and gender stereotypes are distinct [37
]. Due to these differences, we expect to find differences in the attribution of gender-stereotypical traits to the robots.
A total of 161 participants took part in the online survey on volunteer basis, after giving consent. Data were collected in Peru and Japan, through social networks and within students of the Pontifical Catholic University of Peru and of Waseda University.
This number does not include all the respondents who filled in the survey incompletely, and duplicates. All those cases were excluded from the count.
Among the 161 (72 males, 89 females), the mean age was 25.9 (standard deviation (SD) = 8.49), ranging from 17 to 61. They were divided into 4 groups according to their nationality. One outlier was left out of this classification as her nationality was US/Taiwan and did not fit in any group; two more were left out because they stated in the comments that they answered randomly to the questions.
Japanese: male = 22; female = 34; total = 56; age mean = 25.17; age SD = 7.86
Other Asians: male = 12; female = 15; total = 27; age mean = 21.93; age SD = 3.74. Specifically: Chinese = 5; Taiwanese = 2; Indonesians = 3; Myanma = 1; South Koreans = 9; Mongolian = 1; Singaporeans = 3; Vietnamese = 2; Thai = 1.
Peruvians: male = 27; female = 27; total = 54; age mean = 27.95; age SD = 9.59
Other Westerners: male = 11; female = 10; total = 21; age mean = 28.33; age SD = 10.01. Specifically: Spanish = 3; Argentine = 1; Uruguayan = 1; Mexican = 3; Americans = 9; Italian = 1; German = 1; Finnish = 1; French = 1.
The analysis in the following sections is mainly based on Japan v Peru, with additional data on Asians v westerners.
The dimensions of agency and communion of Bem items were verified to be reliable. Internal consistencies (Cronbach’s alpha) were for agency: A1:0.79; A2:0.70; B1:0.80; B2:0.80; C1:0.75; C2:0.79. For communion: A1:0.87; A2:0.90; B1:0.89; B2:0.88; C1:0.91; C2:0.89.
3.3. Gender in Language Use
and Figure 3
reveal the percentage scores regarding the use of personal pronouns “he” or “she” in the text passage for each of the robot body type. Participants referred to the robots most of the time using a male pronoun: only for the case of Peruvians and in general westerners, A2 and C2 were referred as “she”. Only a small number of participants appeared undecided.
A generalized estimating equations (GEE) model was built to ascertain the effects of CHR (Type A/Type B/Type C), WHR (1/2) and cultural group (East/West) on the robot’s sex identification made by participants (Male/Female). The GEE model displayed a QIC statistic (goodness of fit) value of 980.24. In such model, it was found that CHR (Wald’s X2 = 25.81, p < 0.001) and WHR (Wald’s X2 = 78.86, p < 0.001) significantly predicted the sex that participants attributed to the robots, and that cultural group did not exert an influence on participants’ responses (Wald’s X2 = 3.60, p = 0.058). Similarly, the same GEE model was built considering only Peruvian and Japanese participants, and the findings remained constant, QIC = 647.20, CHR Wald’s X2 = 17.06 (p < 0.001), WHR Wald’s X2 = 65.93 (p < 0.001), Cultural Group Wald’s X2 = 2.08 (p = 0.149).
3.4. Stereotypical Attributes
Hereafter, we report the result on the investigation of the attribution of gender-stereotypical traits to the robot body types.
Regarding masculine traits, a 3 × 2 × 2 mixed factorial analysis of variance (ANOVA) revealed the main effects of the within subjects factors CHR, F(1.869) = 29.634, p < 0.001, np2 = 0.16; and WHR, F(1) = 4.995, p = 0.027, np2 = 0.03. No effect was found for the between subjects factor cultural group (East/West), F(1) = 0.048, p = 0.827, np2 < 0.01. All interactions were non-significant and considered irrelevant for further analysis, all F ≤ 0.1.995, p ≥ 0.14, ηp2 ≤ 0.01. Regarding the CHR found effect, three Bonferroni corrected pairwise comparisons indicated that CHR type B (M = 1.67, SD = 0.95) scored higher in masculine traits than CHR type A (M = 1.34, SD = 0.83), p < 0.001, d = 0.37; and CHR type C (M = 1.30, SD = 0.87), p < 0.001, d = 0.43. Analogously, a 3 × 2 × 2 mixed factorial ANOVA was computed to ascertain the effects of CHR, WHR and cultural group considering only Peruvian and Japanese participants. The within subjects findings remained constant and no effect was found for cultural group (Peru/Japan), F(1) = 0.054, p = 0.816, np2 = 0.001.
Regarding feminine traits, a 3 × 2 × 2 mixed factorial ANOVA revealed the main effects of the within subjects factors CHR, F(2) = 8.441, p
< 0.001, np2
= 0.05; and WHR, F(1) = 5.792, p
= 0.017, np2
= 0.04. No effect was found for the between subjects factor cultural group east/west, F(1) = 0.711, p
= 0.400, np2 < 0.01. All interactions were non-significant and considered irrelevant for further analysis, all F ≤ 0.2.662, p
≥ 0.07, ηp2
≤ 0.01. Regarding the CHR found effect, three Bonferroni corrected pairwise comparisons revealed an absence of significant differences among CHR types, all p
≥ 0.113, d
≤ 0.07. Figure 4
and Figure 5
illustrate the breakdown of culture group-wise results. Similarly, a 3 × 2 × 2 mixed factorial ANOVA was computed to ascertain the effects of CHR, WHR and cultural group considering only Peruvian and Japanese participants. The within subjects findings remained constant for CHR, but not for WHR, F(1) = 3.631, p
= 0.059, np2
= 0.033. In addition, no effect was found for cultural group (Peru/Japan), F(1) = 0.479, p
= 0.491, np2
In order to further explore the statistical relationship regarding the attribution of traits, we conducted Spearman correlation analyses. The degree of attribution of female communion traits was negatively correlated both with familiarity with product design (r(155) = −0.24; p = 0.003), and with familiarity with robots (r(156) = −0.19; p = 0.019). Nothing significant was found between familiarity with products design and agency (r(155) = −0.05; p = 0.5), and familiarity with robots and agency (r(156) = −0.12; p = 0.12). Age of participant was also not correlated with either communion (r(154) = −0.12; p = 0.12) or agency (r(154) = −0.11; p = 0.16).
In this manuscript we present the results of an investigation on the perception of gender in robot design, focusing specifically on the proportion between chest, waist and hips to indicate robot gender. Our goal was test whether the manipulation of chest-to-hip ratio and waist-to-hip ratio in robots would orientate the perception of robot gender and elicit gender-stereotypical trait attributions (agency and communion). We performed this experiment with a total number of participants of 158, from different countries.
Our hypotheses regarding body proportions were confirmed by the statistical analysis, and both WHR and CHR appear to influence the participants’ choices of the pronoun. The two factors also play a role in influencing the more subtle choice of stereotypically male or female traits.
This pattern happens regardless of the cultural group east/west, contrary to our expectations. The more specific comparison Japan/Peru follow approximately the same pattern.
Perceived communion trait of robots was also found to be negatively correlated with familiarity with product design and with robots. The latter result was expected, as familiarity with robots and design exposes, respectively, how the robots and their pictures are made. A degree of habituation to robots may make them seen more as machines and reduce the degree of anthropomorphism. Similarly, a familiarity with product design may make the participant aware that the gender is characterised following some design rules, and this might specifically reduce the subtle attribution of traits.
One last thing to remark is the generally low scores of communion and agency, always below half of the scale (which is made of 5 points). This result suggests the need for further design cues to characterise the gender of robots.
4.2. Findings for Robot Design
Let us analyse the results with the concrete purpose of finding design advices for making gender-characterised robots.
From the statistical finding we can determine that, specifically, a CHR visibly greater than 1 suggests a male gender, whereas visibly less than 1 suggests a female gender. Body types A and C are considered more feminine, in an equivalent way differing from B.
WHR can also indicate female when visibly less than 1, whereas values close to 1 indicate male. Body type 2 is considered more feminine than 1.
Therefore, A2 and C2 are the most feminine, and B1 the most masculine.
Additionally, it seems that WHR is stronger than CHR to influence in participants choices.
The three other body types A1, B2 and C1 indeed convey mixed visual signals, which may end up in contradictory perception. This is confirmed by the mismatch visible from a qualitative analysis: A1 and C1 are seen mostly male, judging from the choice of pronoun, while having a high degree of attribution of feminine traits; conversely, B2 is attributed masculine traits, while a relevant share of participants (32 to 37%) choose a female pronoun for it.
These kinds of mismatches should be avoided in robot design, especially the case of humanoids and androids, in which misalignment of visual cues can be critical in the characterization of gender, unless specifically aimed for.
While analyzing the results and the participants’ comments, we realised of the presence of some confounds, which were unavoidable because of language.
In the Spanish language, the introduction sentence “This robot is ... (robot name) ...” reads “Este robot es ...”. The demonstrative pronoun este is male even though commonly applied to any case when gender is not specified. The absence of a neutral form may have slightly biased the answers towards male gender. The same happens in other language such as French, and may have indirectly biased French speaking participants, even though filling the English questionnaire.
A few participants expressed a priori that the word ‘robot’ would be inherently male, and were also biased in that direction.
Additional minor confounds were found in colour, shape and naming of the stimuli. One subject opted for a male pronoun because according to him, “grey colour is male”. Another participant was distracted by the shape of the head, despite it being made as neutral as possible.
A few other participants stated that the consonants Z, X, V, N, C, B and K were either suggesting a male or female name, and were influenced in their choice according to their own perceptions.
4.4. Limitations and Future Work
Besides the confounds, the main limitations of this study were that it was performed by using drawn figures rather than photos of a robot. While on the one hand this choice was made in order to use the most possible generic stimuli, on the other the external validity was limited, as the realistic photo of a robot can give a different impression.
Nevertheless, given the findings of the present research, we can highlight some additional developments. The present study showed the effectiveness of body size manipulation, however limited regarding the extent of subtle attribution of stereotypical traits. This fact prompts the need to stress gender through additional design cues: as stated in [12
], it is important for robot designers to consider a combination of physical characteristics rather than focusing specifically on certain features in isolation. We can find new ideas from the detected confounds of our study. Naming, shape and colour could be manipulated and influence gender perception in a subtle way. Further research should focus on these aspects without resorting to trival means of characterisation.