Personal Space in Human-Robot Interaction at Work: Effect of Room Size and Working Memory Load

For HRI scientists, personal space represents an important variable especially in the field of mobile robotics [29]. As mentioned in the introduction, it is important that robots maintain a socially acceptable distance from humans to avoid negative consequences such as errors or discomfort [1]. Therefore, researchers tried to include the personal space of users when designing navigation algorithms and work environments with the goal that robots maintain a certain comfort distance toward humans [25, 33]. In this respect, several studies tried to investigate the influence of a variety of factors. A recent literature review and meta-analysis by Leichtmann and Nitsch [29] summarized the findings for each of the factors and added new insights by meta-analytic moderation analysis and theoretical considerations on theoretical transfer of psychological theories to HRI settings. However, the meta-analysis mostly revealed only mixed results, not allowing for clear conclusions. This was especially found for the effects of human-related factors such as gender, age, and personality. For example, while some studies reported larger distances by persons identifying as female compared to participants identifying as male (e.g., [26]), other studies reported opposite effects (e.g., [34]) or null-results (e.g., [24]), and thus the meta-analytical 95% confidence interval showed a wide range from negative to positive effects. Amongst several factors, only experience seemed to have a robust effect on comfort distance toward a robot (e.g., [19]). Robot-related factors such as robot anthropomorphism were theoretically argued to possibly influence the effects; however, to date there exist only a few studies on robot appearance [23] and a meta-analytical moderator analysis based on the current literature showed no significant effect [29]. Additionally, empirical investigation of other factors such as environmental factors is still missing.

There are multiple reasons for the variance in effects. For example, the analysis showed various methodological and statistical problems in original studies such as small sample size and thus low statistical power, as well as questionable research practices leading to higher false-positive rates and overestimation of effect sizes. Furthermore, variances could be explained by (hidden) moderators that are not known yet. Theoretical models could potentially explain such effects; however, many studies so far had not been based on solid theoretical models. In contrast, psychologists have proposed several theories and models throughout history in order to explain variances in human distancing behavior, which had been summarized in different reviews [1, 16, 20, 46]. In these theoretical frameworks variance in personal space may be explained by socialization processes and reflect socio-cultural norms, by social arrangements including effects of social status, or by a function of internal psychological states, just to name a few (an extensive overview is beyond the scope of this article and can be found elsewhere; see, for example, [1, 16] as a general overview, or [29] with an HRI focus).

As Leichtmann and Nitsch [29] argue, not all models or the estimated values of model parameters used to explain human-human distancing behavior may also be suitable to explain differences in distances for HRI scenarios, but may differ in effects of key factors or boundary conditions—thus, they may need to be adapted. Therefore, HRI research needs to test such models and theories and explore boundary conditions or the interaction of factors that might be unique to HRI situations. However, only a few studies in HRI have tested certain theories such as equilibrium models (e.g., [34]) or expectation violation theory (e.g., [4]), parameter ranges, and boundary conditions in empirical user studies. To sum up, the literature on personal space had been criticized for its lack of theory [16, 29]. However, theories are essential frameworks that connect different variables, as they help to identify the most impactful mechanisms and factors that lead to differences in human distancing behavior. Such theories can explain the effects of higher-order factors (e.g., room size) by explaining the variance in these effects through underlying mediating (e.g., arousal) and moderating (e.g., threat potential) factors. Future work on personal space in HRI thus needs to put more emphasis on theory and model development and testing. In the next section, we therefore describe theories especially useful in order to explain the effects of contextual factors on comfort distances.

To understand why one would expect differences in distance in the first place (ultimate explanation), Uzzell and Horne [51] describe three basic functions. As a protection function, larger distances serve as a buffer zone to avoid harm in a potentially threatening situation. As an arousal regulation function, distance to others is used to regulate the amount of incoming information and keep it on a level within a certain acceptable range to avoid overload. A third function is the communication function, for which distances are used to indicate information about the nature of the relationship within a social setting.

Based on these ultimate explanations, the following might be speculated: Especially in industrial work settings, a manufacturing robot can be perceived as a potential threat, causing a human to keep distance for protection. Simultaneously, it can also be understood as an additional source of information through its movement and interfaces. In mentally demanding work situations, a robot could thus lead to overstimulation of the human. Keeping larger distances might thus serve to reduce arousal.

In contrast to these ultimate explanations, proximate explanations describe how these functions lead to differences in distances. Widely used models to explain differences in comfort distances are conflict or intimacy equilibrium models [1, 2, 3]. According to these models, approach and avoidance forces exist in every situation and individuals have developed an “equilibrium point” of intimacy and a range surrounding this point within which changes are accepted. If situational changes result in levels higher than desired, “a reciprocal change occurs in one or more ...behaviors in order to restore the desired level” [2, p. 66].

This principle of equilibrium restoration is a realization of the more general principle of homeostasis, which can be defined as “a self-regulating process by which ...systems maintain stability while adjusting to changing external conditions” [7, p. 2]. Similarly, in Hockey’s [22] compensatory control model, “behavior is modified by reference to internal standards or set points (through negative feedback) so that currently active goals may be maintained, and purposive behavior promoted” [22, p. 75]. This regulatory activity may be experienced as mental effort or subjective strain as it attracts costs especially “under conditions of chronic perturbation from stress and environmental load” [22, p. 75]. Thus, to avoid strain and save costs (e.g., avoid an expenditure of mental resources or energetical costs due to sympathetic activation [22]), an internal control system strives to keep effort within a certain range defined by “internal standards.”

These principles can be used to explain the effect of contextual variables such as room size or working memory load on comfort distances in human-robot interaction situations. Contextual factors can be interpreted as factors potentially causing perturbations of inner variables. As suggested by Aiello [1], such inner variables can be a person’s arousal level and its perceived control within the situation. In situations in which arousal is already raised to a certain level or perceived control is lowered, an approaching robot can be an additional perturbation leading to a level of arousal or perceived control outside the acceptable range. A difference in distance toward that robot could be one behavioral adjustment to restore the desired level.

Past research on human-human distancing behavior has demonstrated the effect of room size on personal space [9, 10, 41, 48, 50, 56]. For example, people felt less comfortable and held larger distances toward confederates in smaller rooms compared to larger rooms [48, 56]. These effects can be explained by the protection function as a smaller room offers fewer possibilities to escape a potential threat (protection function).

Another contextual factor that can become especially important at work is the variance in cognitive task demands causing different levels of working memory load. Working memory is a “hypothetical cognitive system responsible for providing access to information required for ongoing cognitive processes” [55, p. 1] and is limited in its capacity. When already processing information, additional information (i.e., in situations with more demanding tasks) can lead to an overload, resulting in either performance impairment [12] or alternative strategies to avoid overload [57]. In human-robot interaction, an approaching robot could increase information processing, leading to overstimulation. Larger comfort distances would thus serve an arousal-regulation function. While there are not many studies on the effect of working memory load on personal space directly [18], the effect of information overload can be ascertained from studies on crowding [53] that leads to discomfort [13, 48] and lower perceived control [43, 44].

The laboratory experiment manipulated the room size surrounding a person and the working memory load in a 3 \(\times\) 2 within-subject design, while a mobile robot approached participants until they signaled the robot to stop. To achieve high mundane realism and thus high ecological validity and better transferability of these (artificial) laboratory results to practical application scenarios in manufacturing companies, the room was manipulated by the variance in storage corridors spanning a partly fenced-in area that is available to the people as an activity space. “Room” is defined here as a certain area, which is defined by corresponding boundaries in length, width, and height. The room size was manipulated on three levels by narrowing down the storage corridor. The working memory load was manipulated on two levels with and without a cognitive secondary task to simulate the variance in working memory load imposed by different cognitively demanding work tasks in manufacturing workplaces.

Based on the study design, sample size was calculated in a power analysis using G*Power for the two main effects of room size and working memory load on distance [14]. A medium effect size was assumed (\(f = .22\)) and the global \(\alpha\)-error probability was set to .05. To achieve a minimum statistical power of 80%, a sample size of \(N = 66\) would be needed.

As preregistered, participants were recruited through social media, advertisement posters in the city, and email lists of the local university (for detailed information of exclusion criteria see the preregistration document [28]). A total of \(N = 72\) participants took part in the experiment at the Augsburg Innovations Park in Germany. Thirty-nine participants were university students (54%); the other participants varied widely in job descriptions such as research assistant from different disciplines (~8%), school teacher, aircraft mechanic, bank assistant, police officer, home keeper, and others. Participants were rather unexperienced with robots, as 57 participants have never interacted with robots or only a few times, and almost no one (96%) had participated in a user study before. Thirty-six persons identified as female, thirty-six as male, and no person of any other gender. The mean age of the subjects was \(M = 30.42\) (\(SD = 12.22\)) years.

As a dependent variable, comfort distance was measured in centimeters from the participants’ tip of the toe to the front of the robot vehicle in each trial after the participants signaled the robot to stop (that is the just acceptable distance; for an exact description of the measurement see Appendix A). The mediating variable perceived as emotional arousal was measured using eight items of the calm-nervous subscale of the German original version of the Multidimensional Mood State Questionnaire by Steyer et al. [47] (\(\omega = .96\)). Perceived control was measured using four self-developed items (\(\omega = .96\), e.g., “How much did you feel you had an influence on what happened in the situation?”). For manipulation check, perceived narrowness of the room was measured using five self-developed items (\(\omega = .93\), e.g., “How oppressive was the situation for you?”), and safety perception was assessed with four items of the job safety subscale by Hayes et al. [21] (\(\omega = .91\)). All response scales were 5-point Likert-type scales. All self-developed items are listed in Table 1 in Appendix B. Additional variables were assessed with the purpose of exploratory analysis but are beyond the scope of this article. Only preregistered, confirmatory analyses are reported here.

The experiment consisted of three parts. In a first introductory part, participants were informed about the course of the experiment, the security concept, participation requirements, and data protection regulations. After giving informed consent, demographic information and information on previous experience with robots was collected.

The second part of the experiment consisted of interactions with a robot. In six different trials under different experimental conditions, participants were approached by a mobile manufacturing robot. The robot consisted of a mobile automated guided vehicle and a robotic arm attached to it (see Figure 1). In each trial, the robot started from 262 cm distance to the participant and approached the participant with a speed of 0.1 m/s. Participants were instructed to stand at a fixed point in the experimental setup during the approach. The experimental setup consisted of a corridor (3 m in length) of movable storage racks filled with cardboard boxes to simulate a warehouse situation (height range: 176–202 cm) (see Figure 1). For each approach, the participants were instructed to say “stop” when the robot reaches a distance at which point the participant is beginning to feel uncomfortable, and that is the point of the just acceptable distance (= comfort distance). The robot then stopped immediately and the distance between participant and robot was measured by the experimenter (stop-distance-procedure). If the participants did not signal the robot to stop, the robot automatically stopped at a minimum distance of 10 cm in front of them. After each run, the subject filled out questionnaires on perceived control, perceived narrowness, perceived safety, and emotional arousal, while the robot returned to its starting position for the next run. For filling out the questionnaires, participants were seated at a table nearby the experimental setup with one of the experimenters next to the table in case of questions (the experimenter did not watch participants answering the questionnaires to avoid social desirability biases). Meanwhile, another experimenter prepared the experimental setup for the next run (e.g., moving storage racks). In sum, each participant had a total of six such runs, which corresponded to the conditions resulting from the 2 \(\times\) 3 design (within-subject design). The order of the conditions was randomized using a random number generator to control for order effects. Care was taken to ensure that no sequence was run disproportionately often by chance (so that participants were about equally distributed across the possible order combinations). During all six approach trials, the robot was operated via a Wizard-of-Oz (WoZ) setup. That means the robot’s movements were remotely controlled by a human operator without awareness of the participants (they were told the robot moves autonomously and can be stopped by language control, as a cover story).

The trials differed according to the independent variables of room size and working memory load. Room size was manipulated using (1) a corridor with a 190 cm width and no obstacle in the participants’ back for the “large” room size condition (with potential escape routes to the side and backwards), (2) a corridor with a 110 cm width and no obstacle in the participants’ back for the “medium” room size condition (with potential escape backwards), and (3) a corridor with 110 cm width and an obstacle in the participants’ back for the “small” room size condition (without possibilities to escape). The manipulation of room size is depicted in Figure 2.

The two levels of the working memory load conditions were manipulated by running the approaches with or without a backwards counting task (counting backwards in three steps starting from a four-digit number). It was assumed that such an additional cognitively demanding task would result in different levels of working memory load.

After running all six trials, participants filled out additional questionnaires in the last part of the experiment. The study took about an hour. At the end of the experiment, the subjects were informed about the cover story (see WoZ) and were rewarded with €15.

While backwards counting is a widely used manipulation of working memory load in human factors research (e.g., [6, 52]), the manipulation of room size had been used for this study specifically. As a manipulation check, the non-parametric ANOVA-type statistic [8]¹ indicates that the perceived narrowness of the room significantly differed between room size conditions (\(F(1.82,\infty) = 95.47, p \lt .001\)), confirming that the manipulation of room size indeed affected the perception of those rooms (narrowness).

As preregistered, the effect of room size and working memory load on distance was tested using the non-parametric, rank-based ANOVA-type statistic [8]. Against our expectations (H1 and H2), neither room size (\(F(1.98,\infty) = 0.31, p = .73\)) nor working memory load (\(F(1,\infty) = 0.16, p = .69\)) had a significant effect on participants’ comfort distance. Comfort distance was thus about the same in the large (\(M = 50.56\), \(SD = 33.12\)), medium (\(M = 52.25\), \(SD = 34.89\)), and small (\(M = 52.05\), \(SD = 35.47\)) room size conditions, as well as in conditions with (\(M = 51.52\), \(SD = 34.49\)) or without (\(M = 51.72\), \(SD = 34.46\)) a secondary task. Boxplots can be obtained in Figure 3.

Although the direct effect of the independent variables on comfort distance was not statistically significant, we calculated the indirect path of the mediation. We tested the effect of the independent variables on arousal and perceived control (H3 and H4) as well as the effect of these on distance. For the mediation analysis, three regression equations were built according to MacKinnon et al. [31]. First, the effect of an independent variable (X; room size or working memory load) on the dependent variable (Y; comfort distance) is calculated (\(Y = i_1 + cX + e_1\)). Second, the effect of the independent variable on the mediation variable (M; perceived control or emotional arousal) is calculated (\(M = i_2 + aX + e_2\)). In a third equation, both the independent variable and the mediation variable are included as predictors (\(Y = i_3 + c^{\prime } X + bM + e_3\)). The mediated effect may then be calculated as \(\hat{a}\hat{b}\) or \(\hat{c} - \hat{c^{\prime }}\) . Note that \(i_1\), \(i_2\), and \(i_3\) represent the intercepts of the regression equations; a, b, c, and \(c^{\prime }\) represent the estimated parameters; and \(e_1\), \(e_2\), and \(e_3\), represent the error terms. The estimated parameters a, b, c, \(c^{\prime }\), and \(\hat{a}\hat{b}\) are reported as beta weights.

For the calculation of 95% confidence intervals (CIs) bootstrapping (700 simulations) was used. Since the study was a repeated-measurement design, it is assumed that individual differences affect the intercepts. Therefore, multi-level models were calculated. Room size was dummy coded with the large room condition as reference. All models are depicted in Figure 4.

While no significant direct effects of room size and working memory load on distance were found, models show significant indirect effects. Smaller compared to larger room size significantly led to higher emotional arousal (\(\beta _{medium} = .08, \beta _{small} = .35\)) and lower perceived control (\(\beta _{medium} = -.17, \beta _{small} = -.31\)). Higher levels of emotional arousal (\(\beta = .18\)) and lower levels of perceived control (\(\beta = -.19\)) in turn led to a significant increase in comfort distance (\(p \lt .001\)). Similarly, in situations with higher levels of working memory load, participants felt significantly more aroused (\(\beta = .44, p \lt .001\)) and felt significantly less in control (\(\beta = -.27, p \lt .001\)), leading in turn to an increase in comfort distance. The results confirm hypotheses 3 and 4. However, these indirect effects (\(\hat{a}\hat{b}\)) are small (ranging from .05 to .09).

As predicted, a regression analysis showed that participants kept shorter comfort distances toward the robot with increasing numbers of encounters (\(distance = 58.08\, cm \mathbin {-} 1.85\, cm * encounter\)). According to the model, comfort distance was reduced by \(b = 1.85\) cm (\(SE = .37, p \lt .001\)) in each of the six rounds (note that b is the unstandardized parameter here, not a beta weight, for better interpretation).

In this study, the effects of the contextual factors room size and working memory load on human comfort distance toward an approaching mobile manufacturing robot were tested. The variation in the size of a corridor and a mental secondary task did not affect the participants’ comfort distance (hypotheses 1, 2, and 5 rejected). However, indirect effects were significant, albeit small. Conditions with smaller room size and higher working memory load resulted in higher emotional arousal and lower perceived control (hypotheses 3 and 4 accepted). This increase was associated with larger comfort distances.

While the results clearly contradict our expectations in terms of the direct effects, the indirect effects confirm what was predicted based on equilibrium models (e.g., see [22]). Certain configurations of contextual factors during an approach by a robot can lead to a perturbation of inner standards. That means a smaller room with fewer opportunities to escape can be interpreted as a situation with higher threat potential (triggering a protection function), and higher workload can lead to overstimulation (triggering an arousal regulation function). This perturbation is realized as a certain level of emotional arousal and perceived control, outside an accepted range. However, according to equilibrium models [22], such a perturbation would lead to regulatory activity to keep inner states within a certain range (homeostasis). Our prediction that this regulatory activity would be expressed as a variation in comfort distance was not confirmed.

However, there are other possible regulatory activities. Hockey [22] describes four types of latent decrement caused by the motivation to restore equilibrium. Besides (1) strategic adjustments of comfort distance, another possibility is the allowance of (2) subsidiary task failures such as a larger amount of errors in the secondary task. Participants could have accepted higher (3) compensatory costs, that is, to mobilize additional resources, or (4) fatigue after-effects, that is, to shift to a “low effort mode” when demands change. However, the large number of possible alternative explanations is one of the main points of criticism [16]. This means that one can come up with many post hoc explanations derived from the model.

Besides these theoretical explanations, null results may also occur because of limitations of the experimental setup. Room size was manipulated using movable storage racks to form corridors. However, while the manipulation affected perceived narrowness as well as emotional arousal and perceived control, the manipulation might not have been effective enough to evoke the need of a behavioral response. That means the disturbance might not have been high enough for participants to take action.

As for every laboratory experiment, criticism on the artificiality of laboratory settings and thus limited validity also applies to this study [17]. The simulated situations differ from real-world situations. In laboratory experiments, people know that they are participating in an experiment and know about the high safety of such studies. For example, for safety reasons, participants had been told in advance that the storage racks were movable and that experimenters monitored the robot. The experimental situation itself thus influences the phenomenon [17], as an experiment itself is a special social situation.

The average comfort distance toward the robot in this study was 51.62 cm with a wide variation between participants and situations (\(SD = 34.44\) cm). In comparison to results of a recent meta-analysis by Leichtmann and Nitsch [29], this distance was smaller than the average distance of 81.61 cm including both humanoid and mechanic-looking robots [29]. However, it falls within the confidence interval of slightly smaller distances for mechanic-looking robots of 45.24 cm to 92.52 cm. In line with the findings of the review results on robot experience effects on distance [29], the results of this study show that comfort distance toward the approaching robot decreased across encounters (H6).

While we did not find the effects we would have expected based on the model, these null effects should not necessarily be seen as a limitation as it allows for further development of the theories and conclusions about parameter ranges in which the theory holds—or does not hold. The results of this study are valuable for further research for different reasons.

First, psychology had a replicability crisis recently; that is, a large amount of studies did not replicate the effects of the original studies, and thus the confidence in psychological research findings was questioned [35, 38, 54]. There are many reasons for that. For example, the use of questionable research practices as well as underpowered studies led to many false positives and overestimation of effect sizes [32, 45, 49]. This problem was even worsened by publication bias; that is, statistically significant results have higher probability of getting published, meaning that the outcome of studies rather than their quality led to publication [54]. This also led to higher false-positive rates in the literature and thus overestimation of effects. It is thus important to publish and consider null results as they will lead to a more accurate estimation of effects in future meta-analysis.

Second, theoretical work by Scheel et al. [42] argued that a major cause of unreliable results is the flexibility in data analysis because a priori hypotheses still had too much room for interpretation, not allowing for precise predictions. This flexibility, arbitrary defaults, and random heuristics led to weak tests that do not inform theory and do not lead to cumulative knowledge. A lot of model parameters, the strength of manipulations, the variance of measures, or the range of values that led to a certain effect of a parameter are often unknown. Researchers thus need to strengthen the link between the test and the tested theory. According to Scheel et al. [42] this could be done, for example, by parameter range estimation or the exploration of boundary conditions (that is, the regions of the parameter space in which the theory applies). The results of this study can inform the literature on personal space in this way. Although we did not find a direct effect in this study, indirect effects indicate that contextual conditions led to a stronger perturbation of inner states. Thus, differences in comfort distance might occur at some point. Smaller rooms led to higher arousal and lower perceived control, and higher arousal and lower perceived control in turn led to larger comfort distances. Effects had just been very small. However, it is conceivable that effects are significant at some point in the parameter range when the perturbation gets stronger. The manipulation might just have been too weak in this case. Based on this learning from this study, future studies could test other configurations of contextual factors that lead to stronger perturbations—that is, testing out the range of the arousal or perceived control parameter and determining at which point larger comfort distances occur. This could be achieved by more extreme levels of room size (e.g., even smaller rooms by additionally lowering the ceiling height) and working memory load (e.g., even more mentally demanding tasks) or other contextual factors not tested yet (e.g., crowding or noise). Additionally, future studies should control for alternative explanations derived from compensatory control models (see [22]). As mediating factors, participants’ arousal and perceived control seem promising. This study is therefore a start for testing internal state models in the context of HRI, and for exploring the effects of contextual factors. Future research could thus build on the results by exploring other configurations and parameter ranges. Furthermore, moderator effects are conceivable as HRI also varies highly depending on context. For example, while this study focused on HRI in a manufacturing setting with a mechanic-looking robot, future studies could test such effects in other, maybe more social contexts with humanoid robots.

Finally, it was pointed out by Leichtmann and Nitsch [29] that HRI research on personal space showed many methodological problems. This article thus also tried to tackle these problems by applying methods recommended by a recent reform movement [35, 54]. (1) The study was preregistered on OSF to restrict the degrees of freedom in data analysis and thus to prevent p-hacking. (2) Sample size was justified by a priori power analysis to ensure a minimum of 80% probability of actually finding an effect if this effect is true. (3) We adhere to the value of transparency and shared our analysis code for other researchers so that other scientists can follow our data analysis in more detail and thus better reproduce results and check robustness. In doing so we hope that future meta-analyses of personal space in HRI will achieve more reliable estimates and hope that other research projects will follow these reforms.

While the meta-analysis by Leichtmann and Nitsch [29] recommends an average comfort distance of 81.61 cm, the recommendation for this specific mobile manufacturing robot might be shorter (\(M = 51.62\) cm). However, comfort distance varies between persons and situations and thus—for safety reasons—a wider default distance is recommended. As we find shorter comfort distances toward the robot with more encounters, interaction designers might also set larger comfort distances at the beginning and shorten the comfort distance after a while when the workers have more experience with the new robot.

Although the study did not find a direct effect of contextual factors on comfort distance, the indirect effects of emotional arousal or perceived control on comfort distance can be used to derive recommendations. As such, situations that are expected to cause high emotional arousal or lower workers’ possibilities of control might also require a robot to keep a larger distance. For example, as configurations similar to those of the experiment (but without an experimenter observing the situation with the possibility to stop the robot remotely) could be more stressful in real-world situations, it might be conceivable that these situations require larger comfort distances. These effects need to be tested in field trials. To be on the safe side, a path-planning algorithm could keep larger distances toward human workers when entering small corridors or when entering certain “high workload areas.” Either the “high workload areas” could be predefined (for example, in assembly areas) or the robot needs to be enabled to identify the situation of the worker. This, however, is currently the subject of research. A different possibility is the control of the robot’s velocity when approaching a human. Within smaller corridors or high workload areas, it might be useful to reduce the robot’s velocity earlier. In addition, it would be an advantage if the robot visually displays its current safety zone. This would provide a transparent indication to humans of the distance to which the robot is approaching before it stops.

Past research on personal space in HRI has studied the effect of different factors on human comfort distance toward robots—however, contextual factors had not been considered. The experimental study in this article tried to address this gap. Although the study did not find a direct effect of room size and working memory load on comfort distance, the study showed indirect effects based on internal state models of human distancing behavior; that is, smaller rooms and tasks with higher working memory load led to higher emotional arousal and lower perceived control, which in turn led to larger comfort distances in general. Future research can build upon these results by exploring boundary conditions, that is, exploring whether more extreme or other contextual factors such as noise would lead to a certain level of emotional arousal, perceived control, and significant direct effects on comfort distance. In the study, moreover, methodological problems that studies of personal space in HRI revealed were avoided by a priori power analysis, preregistration, and open code and may lead to more robust results overall if widely applied in the future.