Differences in Human Response When Interacting in Real and Virtual (VR) Human–Robot Scenarios

Abstract

The utilization of robots has become an integral aspect of industrial operations. In this particular context, the study of the interaction of humans and robots aims to integrate their relevant capabilities with the intention of attaining maximum efficiency. Moreover, in the private sector, interaction with robots is already common in many places. Acceptance, trust, and perceived emotions vary widely depending on specific contexts. This highlights the necessity for adequate training to mitigate fears and enhance trust and acceptance. Currently, no such training is available. Virtual realities have frequently proven to be helpful platforms for the implementation of training. This study aims to evaluate the suitability of virtual realities for training in this specific application area. For this purpose, simple object handovers were performed in three different scenarios (reality, virtual reality, and hybrid reality). Subjective evaluations of the participants were extended by psychophysiological (ECG and EDA) and performance measures. In most cases, the results show no significant differences between the scenarios, indicating that personal perception during interaction is transferable to a virtual reality. This demonstrates the general suitability of virtual realities in this context.

1. Introduction

Nowadays, robotic support is an integral part of industrial production processes [1]. Especially in the area of mass production, robots have become standard and are increasingly spreading to medium-sized companies. According to [2], it is very likely that robots of various kinds will soon find application in more than just their traditional form in industry and also, to an increasing extent, in private households and other work environments. This is primarily due to the rapid advancement of technology and the new possibilities it offers.

As production environments become increasingly flexible, mobile, and versatile, assistant and transport robots are becoming essential [3]. In such applications, robots are no longer completely separated from humans but work together with them, at least to some extent. A number of cooperative or collaborative robots have been available on the market for a while and can be used in a wide variety of areas [4]. A common motivation for introducing cooperation between humans and robots is to reduce monotony and avoid physically difficult tasks that affect ergonomics.

A basic distinction can be made between different forms of human–robot interaction (HRI) [5,6]. A classification is given in the literature by assigning them to a hierarchical structure, which varies in detail depending on the author. Onnasch et al. divide interaction with robots into three classes: coexistence, cooperation, and collaboration [5]. While coexistence does not involve a common work goal between humans and robots, this is certainly the case with cooperation. Although there is no direct dependency between the tasks of humans and robots, they are working on a common, overarching task. Collaboration differs from cooperation in terms of the joint processing of a subtask. In this case, particular attention must be paid to the coordination of both interaction partners. Bender et al. not only include the joint task in the classification but also the respective workspace of humans and robots [6].

Robots that can work collaboratively are called cobots [7]. Cobots are, above all, characterized by the fact that they achieve a significantly higher level of safety despite direct interaction with humans. This is enabled, for example, by using multiple compact robots instead of one large one. In addition, special sensor technology allows for more sensitive perception of the environment and, as a result, better collision detection. The standard [8] specifies safety aspects that are particularly important for collaborative robots. Weber et al. point out that employee qualifications are essential to ensure a sufficient level of safety [9].

In addition to safety, ergonomics also plays a central role in human–robot interaction [10]. The correct use of robot technologies can, under certain circumstances, not only increase cognitive ergonomics but also reduce the physical strain on humans. Strengthening trust in the robot and reducing negative emotions is also important for successful interaction [11,12].

When working directly with robots in a human–robot team, trust in the robot is crucial [13]. Especially in high-risk scenarios, trust directly affects how well people accept working with robots and how they act in different situations. For example, people who do not trust the robot are more likely to intervene in the robot’s automated process. In contrast, too much trust can also have a negative effect on cooperation and situation awareness because people may ignore the robot for a longer period of time. On the one hand, robot performance, size, type, and behavior appear to have a major influence on trust. On the other hand, environmental factors such as task- and team-related aspects also have an impact. Lewis et al. also divide the relevant factors into system characteristics (reliability, errors, predictability, transparency, and degree of automation) and environmental characteristics (risk due to task or context) and also introduce the personal characteristics of user(s) as a factor group [14]. The latter primarily includes personal tendency to trust in general and self-confidence of individuals.

According to [15], the acceptance of robots differs mainly in terms of the task to be performed and also in terms of the domain in which the application takes place. For example, [16] showed that for the healthcare sector, robots that only assist with physical tasks such as transportation are relatively widely accepted. In contrast, robots that work closely with patients, for instance, assisting them with eating, are significantly less accepted. Another factor influencing the acceptance of robots is personal experience [17]. If no personal experience has been gained with the specific robot to be worked with, negative attitudes toward it are more common [18].

In addition to acceptance and trust, people’s emotional reactions can also influence HRI. For instance, a study by [19] provides initial evidence regarding the influence of the robot’s movement speed on the emotional response of the user. Various distances between humans and robots were also investigated. The relevance of these factors is confirmed by the recommendations in the publication by [20], which suggests both a specific maximum speed of the robot and a minimum distance from it.

To sum that up, to make the constantly increasing fields of human–robot interactions pleasant, effective, and productive for humans, and to maximize the quality of the interaction, we are required to develop ways to reduce fears and increase acceptance and trust. Regarding the reduction of fear in human–robot interaction, various research works show the effectiveness of interventions and training [21,22]. Additionally, as the study by [23] shows, targeted training can also influence people’s trust in a system. By familiarizing them with a collision warning system in a car, trust in the system was strengthened in the long term. In order to make training as effective and efficient as possible, technologies that were not used in classic training courses are becoming increasingly popular. One of these technologies is virtual reality (VR) [24]. The use of VR in training offers a number of advantages. First, participants are placed directly in the context of the situation to be trained. This enables them to participate far more actively in the scenario than in other forms of training. It also encourages experimenting with steps that would not have been tried in reality due to a perceived greater risk. The targeted use of gamification can also increase motivation. Another important factor is the possibility of collaboration over long distances, which enables shared experiences in groups. The ability to adapt virtual scenarios to individual users, for example, through different training speeds or learning methods, is another advantage. Various forms of training were already being implemented with the help of virtual realities. Examples include training and professional development for various career groups, including doctors [25]. VR training has been used explicitly to reduce anxiety, as a study by [26] shows. Additionally, training using VR is already available in robotics. However, these are mainly used as a training platform to teach users how to operate the robot correctly [27].

As far as the authors know, no successful training programs to reduce anxiety and other negative emotions while increasing acceptance and trust in human–robot interaction exist. Furthermore, there seems to be a research gap regarding whether VR can be suitable for this kind of training. In order to weigh up the usefulness of VR training, it should first be determined whether virtual interactions with robots have a similar effect on human reactions as real interactions do. In addition to subjective measures, objective measurement methods should ideally also be used to obtain a holistic picture of the processes taking place within the human body. Examples are the cardiovascular and the electrodermal activity, which are non-invasive and mostly cannot be influenced intentionally [28]. In the past, measurements of cardiovascular and electrodermal activity were found to be suitable for assessing the relevance of various factors for human–robot interaction [19,20] as well as for a VR teaching tool [29].

Previous paragraphs have demonstrated the need for training to reduce anxiety and other negative emotions and to increase acceptance of human–robot interaction. The aim of this study is to examine whether emotional reactions, acceptance, and trust in human–robot interaction differ between virtual scenarios and real interactions. This is intended to investigate the suitability of training in virtual realities for this specific use case. Of particular interest is whether emotional and mental activation changes during a virtual interaction with a robot and to what extent it does so. This leads to the following two research questions:

Research question 1. 

Does human–robot interaction in a virtual environment lead to human reactions different to those observed in real-life interactions?

Research question 2. 

Are interactions with a robot in a virtual reality, which is as realistic as possible, subjectively perceived as different from interactions with a real robot?

Learning effects observed during repeated human–robot interactions are also of interest for assessing the suitability of VR training. Special attention should be paid to the differences from possible learning effects in real interaction. When virtual scenarios show fewer objective human reactions or a more positive subjective evaluation with an increasing number of repetitions, this may indicate that training in virtual reality is a good option. That leads to research question three:

Research question 3. 

Is there a difference in the learning effect of human–robot interaction between a virtual environment and a real interaction after multiple repetitions?

When designing VR training, one of the first questions that arises is to what extent the level of detail in the virtual representation of a scene plays a role in the implementation. To answer this question, the fourth research question was formulated.

Research question 4. 

Does the level of detail of the digital twin (e.g., modalities used) have an influence on objective human reactions and the subjective evaluation?

2. Materials and Methods

The study described here is part of a series of various experiments conducted as part of a larger research program [12,15,16,19,30,31]. In addition, this exploratory study is divided into a qualitative and a quantitative experimental section. This paper refers exclusively to the quantitative section of the study. The results of the qualitative section can be found in a publication by [30], but the procedure is also described here for completeness.

2.1. Sample

The sample size consisted of 30 people. Participation in the study was explicitly open to anyone over the age of 18. No special skills or other requirements were necessary. The participants did not receive any compensation. A total of 16 men and 14 women participated in the study. The mean age of the participants was 24.83 years (SD = 3.82, Md = 24.50). The minimum age was 21, and the maximum age was 40. Fourteen people stated that they had worked with robots in the past, while 16 reported that they had not. However, none of the participants worked with robots on a regular basis (at least once a month). The goups did not differ significantly regarding age, $t\left(28\right)=-1.09,p=0.285$, and gender, ${\chi }^2(1,N=30)=2.08,p=0.149$. In addition, participants were asked to rate their general attitude toward human–robot collaboration on a five-point scale. Five people rated this as “partly true,” 14 people rated it as “I think it’s pretty good,” and 11 people rated it as “I think it’s very good.” The response options “I think it’s very bad” and “I think it’s pretty bad” were not picked by anyone. All participants provided their informed consent at the beginning of the study. The study was approved by the ethics committee of Furtwangen University (22-030).

2.2. Study Design

A mixed methods design was used for the study (see Figure 1). Three independent variables were selected to answer the research questions. The first independent variable (IV1) was the type of interaction with three levels: The first type of interaction, “Real interaction”, involved direct collaboration with a real robot. The interaction type “Virtual interaction” involved collaboration with a virtual robot (digital twin) in a virtual reality environment. The third type of interaction, “Mixed interaction”, combined elements from the first two interaction types. In other words, the robot was represented in a virtual reality environment, but was also present and moved in the real world. The continuous synchronization of all the robot’s characteristics with its digital twin ensured that the positions of the two robots were always aligned. The participants were divided randomly but evenly into the three groups. The second independent variable (IV2) was gender. Since there were no participants in the diverse category, IV2 was defined as a dichotomous variable. In addition, each person performed the interaction task a total of six times (T1 to T6). Therefore, the third independent variable (IV3) was a six-level measurement repetition factor.

The dependent variables were derived from the measured subjective and psychophysiological measures, which were combined in this study. The psychophysiological measures were the cardiovascular activity with the dependent variables heart rate (HR) and heart rate variability (HRV RMSSD), as well as electrodermal activity with the dependent variables skin conductance level (SCL), sum amplitude, number of non-specific skin condictance responses (NS.SCR), and mean sum amplitude (sum amplitude/NS.SCR).

Different questionnaires were used to record subjective measures several times during the study: The Self-Assessment Manikin (SAM) [32] with the dependent variables valence, arousal, and dominance, in addition, modifications of the acceptance scale from van der Laan et al. [33], with the dependent variables usefulness and satisfaction and the trust score according to Wagner-Hartl et al. [12], with the dependent variable trust in the robot. The acceptance scale [33] was not queried using five-point items as in the original publication, but using four-point items. The trust scale, according to [12], was again checked for internal consistency using a reliability analysis (Cronbach’s alpha). The four items were answered a cumulative total of 240 times within this study. This results in $\alpha =0.814$. That indicates high or good internal consistency [34]. The Cronbach’s alpha value is similar to the Cronbach’s alpha calculated in the study by [12].

The subjective assessment was completed with a questionnaire about the current emotional state, with the dependent variable emotional status. The questionnaire consists of seven five-point semantic differentials: “relaxed—stressed”, “excited—calm”, “nervous—calm”, “competent—overwhelmed”, “confident—helpless”, “fearless—anxious” and “comfortable—uncomfortable”. Items two and three are inverted. The items were checked for underlying factors using a principal component analysis. A significant Bartlett test ${\chi }^2\left(21\right)=915.77,\mathit{p}<0.001$ and a Kaiser-Meyer-Olkin criterion of 0.876 indicate that the data is suitable for principal component analysis [35]. There is only one component with an eigenvalue greater than 1.0 (4.33). This component accounts for 61.8% of the variance. The resulting component was then checked for internal consistency using Cronbach’s alpha. The cumulative response rate for the items was 240. This results in $\alpha =0.894$, indicating that the internal consistency can be described as good or high [34]. The results of the items can therefore be summarized by calculating the arithmetic mean to form the emotional state scale. Values from 1.0 (positive emotional status) to 5.0 (negative emotional status) can be achieved. Furthermore, the enjoyment of the task was recorded as another dependent variable using a five-point semantic differential scale “very little—very much”. Finally, the performance measures “Correct execution”, “Leaving the area” and “Error-free return” were recorded as dependent variables, which are described in more detail in the following chapters and are based on the standard “Ergonomics of human-system interaction—Part 11” [36].

2.3. Interaction Task

This chapter describes the interaction task with the robot mentioned before, which is derived from a similar task in the work of [19]. This task was the central component of the study. The task was chosen to represent an introductory training scenario that would also be feasible for novices who had no prior experience with robots. It was also important to choose a standard task in order to minimize the influence of possible effects that can arise from complex sequences of actions. One part of the task was performed by the robot and the other part by the participant. Initially, the participant was asked to stand on a marked spot on the floor with his/her body facing the robot. The task began when the robot picked up a wooden cube with an edge length of 4 cm from a storage position and offered it to the participant. The participant picked up the cube and returned to the marked position on the floor. A visual inspection of the cube was then carried out, in which the participant had to check if a black “X” appeared on any side of the cube (see also Figure 2a). The result of this visual inspection was reported to the study supervisor. The performance measure “Correct execution” was derived from the reported result. The cube was then returned to the robot, which was still in the handover position. If the cube fell during this process, the performance measure “Error-free return” was answered with “No”. After the test subject had returned to the position marked on the floor, the robot placed the cube back in the storage position. If the participant only left the marked position to hand over the cube during the entire task, the performance measure “Leaving the area” was answered with “No”. The transfer position was located at a height of 135 cm. The distance to the robot was one meter and was based on the average of five distances used in the work of [19] in the context of a similar task. The movement speed of the robot, at 70 percent of the maximum possible speed, also follows the average speed used in the study described before [19].

2.4. Materials and Measures

Various hardware and software components were required to conduct the study, which are described in more detail in this section. First, two wooden cubes with an edge length of 4 cm were used for the interaction task. One of these cubes had a black cross in one corner on one of its surfaces (Figure 2a). To improve the adhesion of the cube in the robot’s gripper, the cube was covered with a strip of adhesive tape on all sides. The second cube was identical to the first, except for the missing cross. To enable the robot to grip the cube correctly, a suitable gripper was designed and printed using an FDM 3D printer.

The robot used was a HORST600 from fruitcore robotics [37]. The virtual scenario was experienced using the HTC Vive Pro head-mounted display (HMD) with the corresponding controllers. This HMD provides precise user positioning through the integration of the Lighthouse system. Precise positioning, perspective and accurate spacial dimensions were ensured with the help of permanently installed base stations in the room and trackers on all relevant objects (e.g., table and robot). To ensure that the virtual space matches reality as closely as possible, calibration was performed using HTC Vive Tracker 3.0 [38] before each run. The freedom of movement of the participants was restricted by the wired HMD and normally requires special attention from the study supervisor. However, the VIVE wireless adapter allows the HMD to be converted to a wireless version, thereby avoiding the problems described above. The study was recorded via camera.

The virtual scenario was developed using the Unity engine, a real-time development platform that combines graphical scene design with targeted programming of specific functions. Unity also managed the entire experiment autonomously and automatically recorded key events. Digital models of the cubes and the room were created using the software Blender 2.93 LTS and integrated into the virtual scenery. The real HORST600 [37] was controlled using the corresponding operating software. The virtual model of the robot (see Figure 2b) was exported from the original control software, which already provides an exact digital twin corresponding to the real model. During the preparation phase, all movements of the physical robot were systematically recorded via the Modbus/TCP protocol. This process enabled the subsequent digital recreation of these movements as virtual movements via the digital twin. This ensures that the movement paths of the real robot correspond exactly to those of the virtual robot, both in terms of time and space, in order to create comparable test conditions. Additionally, an online questionnaire was used to develop a comprehensive study guide, incorporating all relevant standardized questionnaires and supplementary participant questions. The aim was to ensure that the study was conducted as standardized as possible.

Regarding the psychophysiological reactions of the participants, two measurement methods were used in the study to record these parameters: first, electrocardiography (to measure cardiovascular activity, ECG), and second, electrodermal activity (EDA). Cardiovascular activity was measured using the EcgMove4 (1024 data points per second) from movisens [39], while the EdaMove4 (32 data points per second) [40] from movisens was used for electrodermal activity. The subjective assessments are already described in Section 2.2. In the condition “Real interaction”, all questionnaires were printed and laid out. In the other two conditions, the questionnaires were displayed on the wall behind the table.

2.5. Study Procedure

The procedure was divided into four parts. All sub-areas were independently evaluated during a preliminary test to verify their respective functionalities. The overall duration of the study ranged between 60 and 90 min. Participation was voluntary for all individuals and no compensation was provided. The research setup in the conditions “real interaction”, and “virtual interaction” can be viewed in Figure 3.

2.5.1. Preparation

Upon arrival, the participants were welcomed. This was followed by a brief explanation of the conditions of participation and the important points of the informed consent, which were discussed with the study supervisor. The participants were then required to carefully read the informed consent form, indicate their consent or refusal, and create their participant code within the online questionnaire. The sociodemographic data and the questions about previous experience with robots were also assessed using this questionnaire. The initial phase was completed by attaching the measurement devices for electrodermal [40] and cardiovascular activity [39] and subsequently checking the electrodes.

2.5.2. Familiarization and Baseline

Before recording the baseline, there was a short familiarization phase in the virtual environment to bring all participants to the same level. The participants first had the opportunity to familiarize themselves with the virtual environment and determine the correspondence between reality and the virtual space by touching the table, chair, and robot. They were also able to try out the collision of the controller with the described objects. The second part of the familiarization scenario involved training participants to handle the cube. For this purpose, a sample cube was created. Participants were instructed to try to pick up the cube with the controller and view it from all sides. Once the participant indicated that they felt comfortable handling the cube, the familiarization phase was completed. The baseline measurement was subsequently conducted while the participants were seated, consisting of a five-minute resting assessment. Participants were instructed to fix their gaze on the fixation cross, move as little as possible, and relax.

2.5.3. Part 1: Quantitative Assessment

The quantitative survey phase consisted of performing the described interaction task six times. Except for the type of cube used (with or without a cross), the six tasks were identical. The order in which cubes with or without a cross were presented was permuted. Each possible sequence of the total of 20 variants was performed by at least one person. Participants were assigned to the independent variable “Type of Interaction” using a permuted randomization procedure to ensure balanced gender distribution. Each participant completed all six interaction tasks exclusively within their assigned interaction type. The three types, “real interaction”, “virtual interaction”, and “mixed interaction”, are presented in detail as follows. The study supervisor had full control over the robot at all times and could shut it down immediately if necessary. In the event of a direct collision with the participant, the robot automatically performs an emergency shutdown. The real interaction was the regular form of interaction. No virtual scenario was required for this. The participant performed the task with the real robot and thus had the opportunity to perceive the robot directly using all modalities. Natural grabbing and viewing the cube was also possible in this type of interaction and did not require the use of a controller. The virtual interaction was characterized by the exclusive movement of the digital reflection of the robot. In this condition, the participant receives neither direct haptic nor auditory feedback. In the mixed interaction, a combination of the two types of interaction “real” and “virtual” takes place. Although the participant was in the virtual scenario through a head-mounted display and could therefore only see the robot in virtual form and interact with the virtual cube there, the real robot performs the exact same movements as the virtual robot at the same time and in the same position. This provided auditory feedback and also enabled the controller to collide with the gripper or the robot body, as it is located where it is seen.

After each individual interaction task was performed, the participants provided a subjective assessment. This included the Self-Assessment-Manikin questionnaire [32], the acceptance scale variant according to [33], the trust score [12], the questionnaire on the emotional state, and the question regarding the enjoyment of the situation. In all types of interaction, the corresponding questions were provided for reading. The questions were answered verbally and recorded by the study supervisor. This was done to ensure consistency between the different types of interaction. After completing the subjective assessment, participants were instructed to be seated in the chair for a 90 s resting measurement. This took place after each individual subjective assessment.

2.5.4. Part 2: Qualitative Assessment and Conclusion

The qualitative part of the survey began with the removal of the measuring devices. The participants were then allowed to try out the two types of interaction they had not seen before by performing each interaction task once. The order in which the two remaining types of interaction were performed was permuted. The subjective assessment from Part 1 was provided after each of the two remaining types of interaction. The subsequent interview with the participants was conducted as a semi-structured interview. As already mentioned, the qualitative part of the study is not included in this publication and will therefore not be described in detail. More details about it are presented in [30]. At the end of the study, the participants had the opportunity to ask open questions and were given a brief goodbye.

2.6. Data Analysis

Regarding the quality control of the recorded psychophysiological data, they were reviewed and checked for any artifacts after each participant. Any artifacts found were recorded with the corresponding time stamp so that they could be checked for relevance later during further processing. The psychophysiological data was processed using Data Analyzer from movisens and internal tools. All mean values generated by the program were baseline-corrected prior to the following analysis. Video recordings were used to check for incomprehensible data and special events that occurred during the test. The questionnaires and scales on subjective evaluation were prepared and evaluated according to the published methods.

All statistical tests were performed with a significance level of $\alpha =0.05$. Since this paper follows an exploratory approach [41], results that are only tendentially significant (significance level $\alpha =0.10$) are also taken into account. The results were calculated using (mixed) ANOVAs, MANOVAs, and Cochran’s Q. The psychophysiological measures were collected exclusively in Part 1 (see Section 2.5.3) and analyzed with regard to group differences between the types of interaction and learning effects. The subjective surveys were conducted in both Part 1 (analysis of group differences between the types of interaction and learning effects) and Part 2 (see Section 2.5.4; analysis of sequence effects and differences between the types of interaction as a within-subjects factor). The raw data was examined for movement artifacts and excluded if necessary. Additionally, some sensor problems occurred, resulting in incorrect data recording. These data sets were also excluded from the analysis. For example, in Section 3.1.2, two people were excluded from the virtual interaction condition, and one person was excluded from the real interaction condition.

3. Results

3.1. Psychophysiological Measures

The results of the measured psychophysiological parameters are shown as follows: Section 3.1.1 deals with the analyses relating to research questions one and four, while the analyses in Section 3.1.2 and Section 3.1.3 are used to address research question three.

3.1.1. Differences Between Real, Virtual, and Mixed Interactions

To examine differences between groups of the independent variable type of interaction, two MANOVAs were calculated. One for the dependent variables of the electrodermal activity (DVs: skin conductance, number of spontaneous fluctuations, and mean sum amplitude) and one for the cardiovascular activity (DVs: heart rate and heart rate variability). It is worth noting that the sum amplitude was not included in the analysis of electrodermal activity due to multicollinearity. It was decided to use the mean sum amplitude (sum amplitude/NS.SCR) for the analyses. None of the effects reach the level of significance (see

Table 1. Differences between the types of interaction in the electrodermal and cardiovascular activity (baseline-corrected)—results of the MANOVAs.

Dependent Variable (Group)	F	df	dferror	p	$\mathit{\eta }$2part.	$\Lambda$
Electrodermal activity
type of interaction	0.72	4	52	0.585	0.052	0.898
Cardiovascular activity
type of interaction	0.13	6	46	0.991	0.017	0.966

).

3.1.2. Learning Effect: Electrodermal Activity

In order to examine learning effects, mixed ANOVAs were analyzed. The results show that regarding the skin conductance level (SCL), a significant effect of the interaction task can be seen, $F_{GG}(1.98,47.47)=12.40,p<0.001,{\eta }_{part}^2=0.341$. Post hoc tests (Bonferroni) show a significantly lower skin conductance level in interaction task T1 compared to tasks T4 $(p=0.025)$, T5 $(p=0.009)$, and T6 $(p=0.005)$. Interaction task T2 also shows a significantly lower skin conductance level compared to task T6 $(p=0.014)$. The same applies for task T3 compared to tasks T4 $(p=0.004)$, T5 $(p<0.001)$, and T6 $(p<0.001)$. Figure 4 shows the mean values of the skin conductance level in the various interaction tasks. All other effects do not reach the level of significance (see

Table 2. Learning effect: electrodermal activity (baseline-corrected)—results of the mixed ANOVAs.

Dependent Variable	F	df	dferror	p	$\mathit{\eta }$2part.
Skin conductance level (SCL)
interaction task 1	12.40	1.98	47.47	<0.001	0.341
interaction task*type of interaction 1	1.38	3.96	47.47	0.256	0.103
type of interaction	1.82	2	24	0.184	0.132
Mean sum amplitude
interaction task	0.75	5	120	0.590	0.030
interaction task*type of interaction	0.69	10	120	0.736	0.054
type of interaction	0.42	2	24	0.662	0.034
Sum amplitude
interaction task	1.03	5	120	0.401	0.041
interaction task*Type of interaction	0.80	10	120	0.629	0.063
type of interaction	0.30	2	24	0.740	0.025
NS.SCR frequency
interaction task	1.11	5	120	0.358	0.044
interaction task*type of interaction	0.89	10	120	0.542	0.069
type of interaction	0.15	2	24	0.861	0.012

¹ Mauchly’s test of sphericity indicates that the assumption of sphericity is violated—sphericity correction: Greenhouse–Geisser.

).

3.1.3. Learning Effect: Cardiovascular Activity

To answer the research question regarding the differences in cardiovascular activity between the different interaction tasks and types of interaction, a mixed ANOVA was analyzed for the dependent variables heart rate and heart rate variability. Following the results, a significant interaction task x type of interaction was shown regarding heart rate differences, $F_{GG}(6.71,87.24)=2.22,p=0.042,{\eta }_{part}^2=0.146$. Furthermore, a significant effect of the interaction task, $F_{GG}(3.36,87.24)=7.71,p<0.001,{\eta }_{part}^2=0.229$, and a significant effect of the type of interaction, $F(2,26)=13.80,p<0.001,{\eta }_{part}^2=0.515$ were revealed. Concerning the interaction task x type of interaction (see also Figure 5), post hoc tests (Bonferroni) show that for interaction task T2 the heart rate in the condition “real interaction” was significantly lower than in conditions “virtual interaction” $(p<0.001)$ or “mixed interaction” $(p=0.001)$. The same applies for the interaction tasks T3 (real–virtual: $p=0.004$, real–mixed: $p=0.002$), T4 (real–virtual: $p<0.001$, real–mixed: $p=0.003$), T5 (real–virtual: $p=0.015$, real–mixed: $p=0.007$), and T6 (real–virtual: $p<0.001$, real–mixed: $p<0.001$). Additionally, for the condition “real interaction”, the post hoc test reveals a significantly higher heart rate in interaction task T5, compared to T6 $(p=0.008)$. In the condition “virtual interaction”, there was a significant lower heart rate during interaction task T1, compared to T3 $(p=0.048)$, T4 $(p=0.028)$, T5 $(p=0.020)$, T6 $(p=0.038)$ and a tendency toward significance regarding T2 $(p=0.082)$. In the condition “mixed interaction”, there was also a significantly lower heart rate during interaction task T1, compared to T3 $(p=0.037)$, T5 $(p=0.014)$, and a tendency toward significance regarding T6 $(p=0.090)$. All other effects do not reach the level of significance (see

Table 3. Learning effect: cardiovascular activity (baseline-corrected)—results of the mixed ANOVAs.

Dependent Variable	F	df	dferror	p	$\mathit{\eta }$2part.
Heart rate (HR)
interaction task 1	7.71	3.36	87.24	<0.001	0.229
interaction task*type of interaction 1	2.22	6.71	87.24	0.042	0.146
type of interaction	13.80	2	26	<0.001	0.515
Heart rate variability (HRV)
interaction task 1	1.23	3.38	95.72	0.304	0.045
interaction task*type of interaction 1	0.71	7.36	95.72	0.670	0.052
type of interaction	1.35	2	26	0.277	0.094

¹ Mauchly’s test of sphericity indicates that the assumption of sphericity is violated—sphericity correction: Greenhouse–Geisser.

).

3.2. Subjective Assessments

Next, the subjective evaluations were analyzed. In the respective subsections, analyses are presented and described to answer the research questions two, three, and four. For this purpose, mixed ANOVAs were used.

3.2.1. Valence, Arousal and Dominance

To analyze the differences in the dimensions, valence, arousal, and dominance, of the Self-Assessment-Manikin [32] while interacting within the real, the virtual, and the mixed environments, mixed ANOVAs were conducted. The results show a tendency toward significance for the effect type of interaction regarding the dominance, $F(2,48)=2.95,p=0.062,{\eta }_{part}^2=0.109$. Post hoc tests (Bonferroni) show a tendency toward higher dominance in the real interaction, compared to the mixed interaction ($p=0.066$, see Figure 6). All other effects do not reach the level of significance (see

Table 4. Valence, arousal, and dominance of the Self-Assessment-Manikin [32]: Differences between types of interaction, sequence effect, and learning effect—results of the mixed ANOVAs.

Dependent Variable	F	df	dferror	p	$\mathit{\eta }$2part.
Valence
Sequence effect & type of interaction
type of interaction	1.00	2	48	0.374	0.040
type of interaction*sequence	1.29	10	48	0.262	0.212
sequence	0.20	5	24	0.959	0.040
Learning effect
interaction task 1	2.24	2.82	76.25	0.095	0.076
interaction task*type of interaction 1	0.99	5.65	76.25	0.435	0.068
type of interaction	0.67	2	27	0.521	0.047
Arousal
Sequence effect & type of interaction
type of interaction 1	0.51	1.61	38.57	0.563	0.021
type of interaction*sequence 1	1.58	8.04	38.57	0.163	0.247
sequence	1.12	5	24	0.377	0.189
Learning effect
interaction task 1	0.23	3.00	80.99	0.874	0.009
interaction task*type of interaction 1	0.58	6.00	80.99	0.745	0.041
type of interaction	0.24	2	27	0.786	0.018
Dominance
Sequence effect & type of interaction
type of interaction	2.95	2	48	0.062	0.109
type of interaction*sequence	1.27	10	48	0.276	0.209
sequence	1.31	5	24	0.292	0.215
Learning effect
interaction task 1	1.06	2.59	69.81	0.366	0.038
interaction task*type of interaction 1	1.35	5.17	69.81	0.252	0.091
type of interaction	1.07	2	27	0.356	0.074

¹ Mauchly’s test of sphericity indicates that the assumption of sphericity is violated—sphericity correction: Greenhouse–Geisser.

).

3.2.2. Acceptance

To analyze differences in the dimensions, usefulness and satisfaction, of the acceptance scale based on [33] for different interaction tasks within the types of interaction, mixed ANOVAs were used. The investigation of sequence effects and differences between the types of interaction reveals a significant effect type of interaction for the dimension usefulness, $F_{GG}(1.54,36.86)=3.85,p=0.041,{\eta }_{part}^2=0.138$. Post hoc tests (Bonferroni) show a significantly lower usefulness of the real interaction, compared to the virtual interaction ($p=0.006$, see Figure 7). For the dimension satisfaction, a significant effect of the type of interaction is present when examining sequence effects and differences between the types of interaction, $F(2,48)=3.48,p=0.039,{\eta }_{part}^2=0.127$. Post hoc analyses (Bonferroni) show a tendency toward satisfaction for the virtual condition, compared to the real condition ($p=0.068$, see Figure 8). All other effects did not reach the level of significance (see

Table 5. Acceptance scale, adapted from [33]: Differences between types of interaction, sequence effect, and learning effect—results of the mixed ANOVAs.

Dependent Variable	F	df	dferror	p	$\mathit{\eta }$2part.
Usefulness
Sequence effect & type of interaction
type of interaction 1	3.85	1.54	36.86	0.041	0.138
type of interaction*sequence 1	1.14	7.68	36.86	0.359	0.192
sequence	0.80	5	24	0.561	0.143
Learning effect
interaction task 1	2.04	3.00	70.09	0.124	0.070
interaction task*type of interaction 1	1.46	5.12	70.09	0.212	0.098
type of interaction	2.32	2	27	0.118	0.146
Satisfaction
Sequence effect & type of interaction
type of interaction	3.48	2	48	0.039	0.127
type of interaction*sequence	1.32	10	48	0.249	0.215
sequence	2.04	5	24	0.109	0.299
Learning effect
interaction task 1	0.62	3.61	97.35	0.630	0.023
interaction task*type of interaction 1	0.70	7.12	97.35	0.673	0.050
type of interaction	1.73	2	27	0.196	0.114

¹ Mauchly’s test of sphericity indicates that the assumption of sphericity is violated—sphericity correction: Greenhouse–Geisser.

).

3.2.3. Trust

Mixed ANOVAs were used to analyze differences in the trust score [12] for the interaction tasks and the types of interaction. In this case, no significant differences are present (see

Table 6. Trust score [12]: Differences regarding the types of interaction, sequence effect, and learning effect—results of the mixed ANOVAs.

Dependent Variable	F	df	dferror	p	$\mathit{\eta }$2part.
Trust
Sequence effect & type of interaction
type of interaction	1.23	2	48	0.303	0.049
type of interaction*sequence	1.62	10	48	0.128	0.253
sequence	1.07	5	24	0.401	0.182
Learning effect
interaction task 1	0.85	2.73	73.64	0.463	0.030
interaction task*type of interaction 1	1.38	5.45	73.64	0.237	0.093
type of interaction	0.22	2	27	0.806	0.016

¹ Mauchly’s test of sphericity indicates that the assumption of sphericity is violated—sphericity correction: Greenhouse–Geisser.

).

3.2.4. Emotional State

To analyze differences in the emotional state of the participants regarding the interaction tasks and the types of interaction, mixed ANOVAs were used. Following the results, the effect of sequence tends to be significant, $F(5,23)=2.28,p=0.081$, ${\eta }_{part}^2=0.331$. When determining learning effects, the effect of the interaction task is significant, $F_{GG}(2.33,60.56)=3.65,p=0.026,{\eta }_{part}^2=0.123$. Post hoc tests (Bonferroni) show significantly more positive emotional states in the interaction task T5 than in the interaction task T1 ($p=0.032$, see Figure 9) and a tendency toward more positive emotional states than in interaction tasks T2 $(p=0.080)$, T3 $(p=0.098)$, and T4 $(p=0.080)$. All other effects do not reach the level of significance (see

Table 7. Emotional state: Differences between types of interaction, sequence effect, and learning effect—results of the mixed ANOVAs.

Dependent Variable	F	df	dferror	p	$\mathit{\eta }$2part.
Emotional state
Sequence effect & type of interaction
type of interaction	1.75	2	46	0.186	0.071
type of interaction*sequence	0.37	10	46	0.953	0.075
sequence	2.28	5	23	0.081	0.331
Learning effect
interaction task 1	3.65	2.33	60.56	0.026	0.123
interaction task*type of interaction 1	0.27	4.66	60.56	0.920	0.020
type of interaction	0.26	2	26	0.773	0.020

¹ Mauchly’s test of sphericity indicates that the assumption of sphericity is violated—sphericity correction: Greenhouse–Geisser.

).

3.2.5. Enjoyment of the Situation

The results of the analysis (mixed ANOVAs) of differences in the enjoyment of the situation show a significant effect of the interaction task, $F_{GG}(2.84,76.65)=6.20$, $p<0.001,{\eta }_{part}^2=0.187$. Post hoc tests (Bonferroni) show a tendency toward significantly lower enjoyment of the task in the real interaction $(M=2.85,SD=1.05)$ compared to the virtual interaction $(M=1.95,SD=0.77,p=0.098)$. Also, the effect of the type of interaction tends to be significant in this case, $F(2,27)=2.61,p=0.092,{\eta }_{part}^2=0.162$. Post hoc tests (Bonferroni) show significantly lower enjoyment in interaction task T6, compared to task T2 $(p=0.029)$ and T3 $(p=0.014)$. Interaction task T6 also shows a tendency toward lower enjoyment compared to task T1 $(p=0.083)$, and task T5 shows a tendency toward lower enjoyment compared to task T3 $(p=0.052)$. The values of enjoyment across the interaction tasks are shown in Figure 10. All other effects do not reach the level of significance (see

Table 8. Enjoyment of the situation: Differences regarding the types of interaction, sequence effect, and learning effect—results of the mixed ANOVAs.

Dependent Variable	F	df	dferror	p	$\mathit{\eta }$2part.
Enjoyment of the situation
Sequence effect & type of interaction
type of interaction	1.70	2	48	0.194	0.066
type of interaction*sequence	1.64	10	48	0.123	0.255
sequence	0.95	5	24	0.470	0.165
Learning effect
interaction task 1	6.20	2.84	76.65	<0.001	0.187
interaction task*type of interaction 1	0.75	5.68	76.65	0.604	0.053
type of interaction	2.61	2	27	0.092	0.162

¹ Mauchly’s test of sphericity indicates that the assumption of sphericity is violated—sphericity correction: Greenhouse–Geisser.

).

3.2.6. Performance Measures

This chapter presents an analysis of the performance data. First, Cochran’s Q tests were used to check whether an effect of learning could be detected. Subsequently, a Kruskal-Wallis test was carried out to evaluate differences between the types of interaction. The dependent variables are the frequencies of the particular performance measures. It should be noted that no tests were calculated for the performance measures regarding a left position or the correct performed visual inspection because all recorded values at all times indicated that the position was not left and the visual check was always correct. For the performance measure “Error-free return to the robot”, Chochran’s Q test does not show a significant result, ${\chi }^2\left(5\right)=7.00,p=0.444$. In the interaction tasks two (correct: 28, incorrect: 2) and five (correct: 29, incorrect: 1), the cube was not returned to the robot correctly at all times. The analysis of differences between the types of interaction also shows no significance regarding the performance measure “Error-free return to the robot”, $H\left(2\right)=2.15,p=0.754$.

4. Discussion

4.1. Research Question 1

The results of the analyses for research question 1 “Does human–robot interaction in a virtual environment lead to human reactions different from those observed in real-life interactions?” reveal no significant differences for these two types of interactions. In this case, this applies to both cardiovascular and electrodermal activity. Since these measurements are influenced by physical, mental, and emotional stress depending on the measured parameter [42,43], it can be assumed that humans do not experience and perceive the two situations significantly differently. These findings can be compared with the results of a meta-study by [44], which examined psychophysiological reactions to fear in virtual reality and found a fear response fairly similar to that observed in reality. The results regarding research question 1 provide an initial indication of the suitability of virtual realities for implementing a training program in the context of this study. However, some limitations must be pointed out, which, if taken into account, may provide a more comprehensive picture of the situation. In this study, only the first movement of the robot was used as the evaluation period for the psychophysiological measures. Although this is identical in all conditions (less than 15 s), it represents a rather short comparison period for the tonic measures used here, since the latency of electrodermal activity typically ranges between one and two seconds [43]. It may be possible to gain even deeper insights by additionally evaluating phasic measures of the electrodermal activity [43] at specific relevant points in time (e.g., the start of the robot).

4.2. Research Question 2

The results of the analyses for research question 2 “Are interactions with a robot in virtual reality as realistic as possible, or are they subjectively and differently perceived as interactions with a real robot?” show that with regard to the dimensions valence, arousal, and dominance of the Self-Assessment-Manikin [32], no significant differences were found between the virtual and real interaction. These findings are supported by the results of the questionnaire on the emotional state as well as by the non-significant results regarding the measured trust [12] and the measured enjoyment of the situation, which also indicates perceived experiences that are not different among the participants. Only the dimensions, usefulness and satisfaction, based on [33], showed significant differences. Here, both the perceived usefulness and the perceived satisfaction were greater during virtual interaction. This effect may be attributable to the novelty of the virtual environment. Additionally, acceptance could be influenced by the reduced perceived risk associated with virtual interaction, as observed in the study by [45]. However, in the current study, perceived risk was only roughly assessed using the applied trust scale. Since this did not reveal any significant differences, further research should include a separate assessment of the perceived risk in order to determine a possible correlation. Although the emotional demands appear not to be significantly different in both types of interaction, differences in acceptance showed that the question cannot be answered completely uniformly. Nevertheless, conclusions can be drawn about the suitability of virtual realities for the design of training courses. Given the comparable emotional demands, levels of trust, and enjoyment observed, it can be concluded that virtual reality is fundamentally suitable for representing real human–robot interactions. Both objective measures and subjective assessments now indicate that VR provides an effective approximation of real-world scenarios/human–robot interactions. The higher values in the acceptance dimensions may also indicate that virtual realities are well suited for the implementation of training scenarios. Among other things, higher acceptance levels suggest a greater willingness to engage in initial interactions with a robot in a virtual environment. The fundamental relationships between acceptance and intention to use are, among others, demonstrated in the publication by [46]. This initial reduction in inhibitions may also ease subsequent training on real robots and lead to greater overall acceptance. Although this shows that virtual interaction does not differ significantly from real interaction, it should be noted that mixed interaction was considered the most suitable. This points to the relevance of the modalities used (see Section 4.4).

4.3. Research Question 3

Learning effects in both psychophysiological measures (see Section 3.1) and subjective assessments (see Section 3.2) were examined to answer the third research question “Is there a difference in the learning effect of human–robot interaction between a virtual environment and a real interaction after multiple repetitions?”. With regard to psychophysiological measures, significant results were found for the skin conductance level and heart rate. The skin conductance level (SCL) increased as the task progressed (see Figure 4). If a learning effect were present, a decrease in SCL would be expected over time, as this would indicate lower stress levels [43]. In the present experiment, the opposite was the case. One explanation could be that the higher values may have been caused by the very high temperature overall (exp. in summer), which continued to rise during the experiment. Since the main function of the sweat glands is temperature regulation [43], an increase in activity is to be expected here. This is also supported by the fact that none of the other parameters of the electrodermal activity showed significant effects of repeated measurements. These parameters are less susceptible to temperature fluctuations, as they are amplitude parameters that occur individually and not a basic skin conductance value. To avoid such temperature-dependent results, a room with a constant temperature should be selected for future studies. Significant differences were also shown for the heart rate. In addition to the significant effect of the interaction task, the main effect type of the interaction and the interaction of the two main effects were significant. For T2-T6, the heart rate was significantly lower in the real interaction than in the two conditions with head-mounted displays. Additionally, a significantly lower heart rate was observed in the first contact task (T1) compared to many other interactions for the virtual and mixed conditions. This also does not correspond to an expected classical learning effect, in which the heart rate should decrease with decreasing physical or mental demand [42]. The significant interaction suggests that the difference between the interaction tasks does not apply to all types of interaction with the robot. Therefore, it is noticeable that the difference in the first contact task only appears to occur in the conditions with head-mounted displays. One possible explanation for this is the orientation reactions of humans. When an event occurs that attracts a person’s attention (orientation reaction), the heart rate drops for a certain period of time [47]. On the other hand, regarding a real interaction with the robot, the situation seems to be predictable, it may be more unpredictable for virtual situations. It can be assumed that, since it was not initially known how the robot would behave in the conditions with head-mounted displays, it is possible that a greater orientation reaction and thus a greater drop in heart rate was triggered during the respective first contact task. In the real interaction, the real robot leaves little room for interpretation as to what a movement actually looks like. Such heart rate behavior during first contact is also described in the second experiment of a study by [48]. In the case of subjective measures, most measures showed no learning effects, as one would have expected during familiarization with the robot. If there were a learning effect, a decrease in stress would also be expected here. Regarding the dimensions valence, arousal, and dominance of the Self-Assessment-Manikin [32], the acceptance scale based on [33], and the trust score [12], no significant effects were found regarding the interaction tasks. Only for the scale of the emotional state was a significant main effect interaction task shown, which could be associated with a learning effect. Interaction task five showed a significantly more positive emotional state than all previous performances. This can be interpreted as habituation to the situation and thus as a learning effect. Although there were significant differences in the enjoyment of the interaction tasks, the enjoyment was descriptively lower at later stages than at the beginning (see Section 3.2.5). This can be explained by the monotony of the repetition itself and was confirmed by feedback from the participants at the end of the study. They referred to a lack of variety and a high degree of monotony. Overall, regarding the third research question, it was shown that with the exception of the emotional state, there appears to be no learning effect in the classic sense. Consequently, no differences between the types of interaction were found. However, it should be noted that in this case, no conclusions should be drawn about the suitability of virtual realities for designing training in the context of this work. The task used was not designed as a training task. It was selected from a study by [19], as this task had already provided indications of differences in human reactions to different distances and movement speeds of the robot. If a learning effect is to be demonstrated, a more application-oriented task that could occur in real training situations should be selected for future studies.

4.4. Research Question 4

The analysis regarding research question 4 “Does the level of detail of the digital twin (e.g., modalities used) have an influence on objective human reactions and the subjective evaluation?” shows no significant differences for neither cardiovascular activity nor electrodermal activity. As already discussed in Section 4.1, this circumstance overall suggests no significant different physical, mental, and emotional stress levels [42,43]. A similar pattern can also be observed when looking at the subjective measures. In this case, however, neither the usefulness nor the satisfaction of the acceptance scale based on [33] showed significant differences between the mixed interaction and one of the other two types of interaction. The only significant difference was found in the dimension dominance of the Self-Assessment-Manikin [32] between the real and mixed interactions. The mixed interaction was less dominant than the real interaction. A higher degree of dominance represents greater personal control over the scenario [32]. For this question, this may indicate that virtual reality and the use of realistic sensory impressions can be used to create a situation in which people feel they have greater control over events than in reality or in virtual reality without additional sensory impressions. This may support the relevance of a high level of detail. If this results in a higher perceived level of control, individuals may feel more comfortable in the situation. The fourth question must therefore be answered in a differentiated manner. The level of detail does not seem to have any effect on objective human reactions at first. However, it does have a significant effect on the subjective evaluations. With regard to the questionnaires used, this difference is only visible in terms of perceived dominance. In the case of the level of detail, further research is needed to obtain a final assessment of these effects and the extent to which the use of certain details is possible or even necessary. In a more extensive comparison, for example, with a different interaction task, it may be possible to demonstrate the effects of the level of detail on the psychophysiological measures and the subjective questionnaires.

4.5. Limitations and Outlook

Fundamentally, it can be stated that the method used was effective and appropriately selected to answer the research questions. As already mentioned in the previous chapters of the discussion, it may be useful to supplement the results with additional subjective measures in some areas. However, the results of the performance measures confirm that the interaction task was suitable for all types of interaction with the robot. As described in Section 3.2.6, no significant differences were found between the individual interaction tasks, as well as between the types of interaction. Therefore, the difficulty and feasibility do not seem to vary under different conditions. The very high success rate and the very low number of errors generally indicate that the task was feasible. In further investigations of this kind, care should be taken to select a suitable room. Ideally, it should be acoustically shielded and kept at a constant temperature by air conditioning. This would further minimize external influences [49]. The use of wireless VR technologies in this study had both advantages and disadvantages. Advantages include greater freedom of movement, a wider range of possible movements, and, as a result, less disruption to the study process. However, disadvantages arise, among other things, from the heat already mentioned. Participants reported in interviews that wearing a head-mounted display was very uncomfortable at warm room temperatures. It is very important to address and therefore control these technical aspects in future studies. Despite the relatively small sample size and the homogeneity of the sample in terms of age and educational background, the results provide valuable basic knowledge about the possibilities of training human–robot interaction in virtual realities. However, this study only covers the group of potential career starters. Future studies should also consider older employees who may need to be retrained to work with robots.

The present results can be used in many ways as a basis for future research. As the theoretical background of this study has shown, there is a need for training to reduce anxiety in human–robot interaction. Based on this study, implementing such training using virtual reality seems very useful and should definitely be explored further. A possible next step is the development of various training tasks, which should then be assessed. By combining the information gained from this study with basic principles from the literature, it may be possible to find ideal interaction tasks for training scenarios.

5. Conclusions

The presented study examined the suitability of virtual realities for implementing a training program to reduce and prevent fears and other negative emotions in human–robot interaction. The potential VR training is intended to increase acceptance and trust. Four research questions were therefore formulated. The first finding relates to the similarities observed between human reactions when interacting with a robot in reality and in a virtual scenario. That was made possible by recording psychophysiological measures. This initial finding already suggests that the stress factors associated with human–robot interaction can be transferred to a virtual scenario. The results of a similar evaluation, but for subjective measures, largely confirm this. However, some differences in acceptance were found, which may be due to the novelty of virtual interaction.

In addition, the question arises as to whether a learning effect could already be observed in this very rudimentary type of interaction. Such a learning effect was not shown by the results of this study. Although isolated effects related to the iterations of the task did occur, these differences were, in most cases, inverse to the expected direction. Only the scale for the emotional state showed a more positive value in later interactions than in earlier ones. However, this was the only indication of a learning effect. Choosing a different interaction task with a higher training factor might reveal other effects. Therefore, future studies shall investigate this aspect. However, the findings on learning effects have no bearing on the suitability of virtual realities, as the task was not designed to determine such an effect.

Furthermore, the question arose as to how varying levels of detail in the virtual environment influence participants’ reactions and subjective evaluations. No direct effects on psychophysiological measures could be found. Except for perceived dominance, the subjective measures also did not show any significant differences. However, when looking at the interview results, completely different feedback from the participants can be observed. The results indicate that differences arising from methodological approaches—particularly regarding the modalities employed—may have obscured detectable variations, underscoring the need for further research in this area.