Design space taxonomy and investigation of gesture-based transparency cues in human-robot interaction

1 Introduction

Robots are increasingly common in people's daily lives and workplaces (Dzedzickis et al., 2022; Licardo et al., 2024), ranging from vacuum or lawn mower robots at home to robotic waiters in restaurants or robots in the industry: More than 500,000 industrial robots and around 200,000 professional service robots are installed worldwide every year (International Federation of Robotics, 2025). For all use cases, whether in production, medical services, agriculture, or cleaning, it is key that users have an adequate understanding of the robotic system to maintain trust and engage in responsible, safe, and appropriate interactions (Schött et al., 2023). Therefore, robotic systems per design need to be transparent to their users, that is, they need the ability to communicate their internal states, goals, behavior, and intentions in a way that users can understand (Miller, 2019).

However, it is no trivial task to design transparent robotic systems adequately and effectively in a way that matches users' needs. On the one hand, the transparent design of robotic systems has shown the potential to enhance trust, decision quality, user understanding, and safety perceptions (Akalin et al., 2023; Nesset et al., 2021). On the other hand, too much information or transparency cues that do not match user expectations can also backfire, leading to confusion or mistrust and jeopardizing user experience (UX) (Lyons, 2013; Nunes and Jannach, 2017). It is therefore crucial that transparency is carefully designed, seamlessly and holistically tailored to specific products, and that the extent and modality of transparency cues match user needs (Diefenbach et al., 2022).

Previous research on robot transparency has, for the most part, relied on “monomodal visual or verbal explanations” (Schött et al., 2023), which are not integrated into the system per design. This modality, however, is limited in the kind of information it can convey (e.g., it depends on users speaking the system's language) and is not suitable for all situations alike (e.g., text-based explanations require users to explicitly pay attention and interrupt the workflow, or verbal explanations may be hard to understand in noisy surroundings; Kerzel et al., 2022).

Another promising approach is gesture-based transparency cues as a form of robot actions that mirror non-verbal cues such as gestures, gaze, and spatial behavior of human interactions. This modality of communication makes use of the unique opportunities for interaction that robots gain (e.g., compared to purely digital technology) through their physical embodiment, matching the modality in which the system primarily operates. Additionally, because of the parallels to human interactions, such cues might seem natural, subtle, unintrusive, and intuitively understandable, without increasing cognitive load (Mead and Salem, 2012; Nukovic et al., 2024; Wang et al., 2022; Zhao et al., 2026). Thus, gesture-based transparency cues would function independently of verbal communication, which could be especially valuable in industrial applications where speech is impractical or impossible (Wang et al., 2022) and for robots without a textual or visual interface, which is commonly used for presenting robot explanations (Anjomshoae et al., 2019).

However, empirical investigations of gesture-based robot transparency are scarce (Schött et al., 2023). Therefore, this article contributes to a stronger basis for designing robot gestures that support understanding and transparency perception in human-robot interaction (HRI). Our study aims to explore the following general research questions:

• Can robot gestures explain goal-deviating robotic behavior, i.e., create transparency in HRI?
• What different options are there to design gesture-based transparency cues, i.e., what are the levels and dimensions of the design space?
• How might such differences in the specific design of robot gestures affect users' experience, in particular their understanding and perception of technology transparency?

In the following sections, we briefly discuss the theoretical foundations of our research, focusing on related work on technology transparency and gestures in HRI. This is followed by a pilot study that demonstrates the need for a systematic design space for gesture-based transparency of robots. We go on to propose a multidimensional design space taxonomy that defines relevant categories and questions for such design and, thus, can serve as a conceptual and practical foundation. After this, we apply the design space taxonomy to an exploratory user study with a kitchen aid system that includes a robotic arm. We empirically explore the impact of different gesture types on user understanding and perception of technology transparency and thus investigate the extent to which different robot gestures derived from the initial design space taxonomy can explain robotic behavior.

2 Related work

2.1 Designing for technology transparency

Based on previous conceptualizations (DIN Deutsches Institut für Normung e.V., 2023; Hein and Diefenbach, 2025; Hein et al., 2026; Miller, 2019), we define technology transparency as “users having the information they want about the technology” (Hein et al., 2026) – an umbrella term for a growing and interdisciplinary field related to terms like XAI, interpretability, or explainability of technology (Alarcon and Capiola, 2025; Graziani et al., 2023), which does not yet have a consistent and clear terminology (Brennen, 2020; Felzmann et al., 2020). Technology transparency consists of different facets, for example, regarding what a technology “currently does, how it reaches its outcomes, and why” (Hein et al., 2026). Applied to robotic interactions, transparency of robotic systems has been described as the capacity to reveal their decision-making processes in a way that is comprehensible to users (Theodorou et al., 2017).

Technology transparency can be beneficial for establishing user trust in the system, accountability, autonomy, and control (Felzmann et al., 2020; Scharowski et al., 2023), a greater understanding and positive impression of the system, as well as easier use (Chazette and Schneider, 2020), more positive affect, and higher willingness to use the technology (Papenmeier et al., 2022; Shin, 2020). However, technology transparency needs to be carefully implemented in the design process, as not all explanations (and modalities of explanations) are automatically perceived as transparency-increasing and understandable. Poorly designed transparency cues can seem overwhelming, mentally taxing, cause a negative impression, impair the use, hinder understanding, or consume unnecessarily much time to handle (Bitzer et al., 2023; Chazette and Schneider, 2020; Deters et al., 2023). In summary, the success of transparent technological systems hinges on the right fit between transparency cues, context, and user needs (Felzmann et al., 2020).

Designing transparent systems, thus, is a delicate process that requires careful thought and evaluation, yet few articles offer concrete and systematic guidelines for doing so (Diefenbach et al., 2020; Hein et al., 2023). Moreover, existing investigations of technology transparency often focus on digital technologies with traditional graphical interfaces like mobile apps or AI applications (Guidotti et al., 2019; Mohseni et al., 2021) but neglect robotic systems and their unique opportunities for designing transparency (e.g., using gestures or spatial relations as means of communication). Making use of these unique options, this article proposes and explores an initial systematic design space taxonomy for developing transparent robotic systems through robot gestures.

2.2 The role of gestures in robotic design

Drawing on foundations of non-verbal communication between humans, gestures can be defined as spontaneous hand movements that are “intentionally expressive” (Kendon, 2004) rather than purely instrumental (Kendon, 2004; Novack and Goldin-Meadow, 2017). Thus, they convey a specific, purposeful message to the recipient. In this regard, four types of gestures can be distinguished in human communication (McNeill, 1992): iconic gestures, which depict concrete objects or actions (e.g., shaping an object in the air with the hands); metaphoric gestures, which illustrate abstract concepts (e.g., size or weight); deictic gestures, which involve pointing to refer to an object or location; and beat gestures, which are rhythmic movements aligned with the flow and emphasis of speech.

Human non-verbal communication has been applied to HRI before, for example, with regard to gaze or facial expressions (Janowski and André, 2021; Leoste et al., 2024; Urakami and Seaborn, 2023). Regarding gestures, previous research has focused on robots reacting to human gestures (Liu and Wang, 2018; Shrinah et al., 2024; Terreran et al., 2023; Wang et al., 2022) as well as robots themselves gesturing to accompany speech (Breazeal et al., 2005; Ham et al., 2015; Huang and Mutlu, 2013; Wolfert et al., 2022), to express personality traits (Niu et al., 2022; Zabala et al., 2021), to make robotic behavior more human-like and emotionally expressive (de Wit et al., 2023), or to increase likability (Zinina et al., 2020).

However, research on the effects of robot gestures on users' transparency perception and understanding is limited. While gestures can be effective in communicating motion intentions (Dragan et al., 2013; Holladay et al., 2014; Pascher et al., 2023; Sheikholeslami et al., 2018; Zhao et al., 2026) or internal states such as emotions or thought processes (Hirschmanner et al., 2021; Kerzel et al., 2022; Rossi et al., 2021; Takayama et al., 2011), little is known about gestures to explain robotic behavior that deviates from the user's interaction goal. A previous study (Kwon et al., 2018) showed that robot gestures differ in their ability to communicate a robot's goal and the reason for its incapacity to perform a task. However, to our knowledge, there has been little further research on how robot gestures impact transparency perceptions and understanding after erroneous, unintended, or otherwise goal-deviating behavior – though this aspect is especially crucial as handling goal-deviating robot behavior is important for shaping users' experience of the system, and the need for transparency is especially high after a goal impairment (Hein and Diefenbach, 2025).

3 Initial exploration of the design space

3.1 Pilot laboratory study

Can robot gestures explain goal-deviating robotic behavior? In this pilot study, we investigated whether gesture-based explanations can improve the understanding and perceived transparency of a robotic system. This investigation served two key purposes: first, to test the general idea of robot gestures to enhance the system's transparency; and second, to provide a basis for extracting different dimensions of the design space for robot gestures. As an exemplary use case, we used a non-anthropomorphic robotic arm that operates in a smart kitchen environment. Based on the emergence of robots designed to help individuals with everyday tasks in homes or personal environments, such as carrying or fetching objects (e.g., Samsung Bot Handy, Korstanje, 2024; Toyota Human Support Robot; Toyota Blog, 2021), the smart kitchen environment was built as part of a larger research project at our university, with the opportunity to program different robotic arm gestures. This allowed us to test the effects of gestures on understanding and transparency perception in a less easily understandable setting for users (as gestures from a non-humanoid robot may be harder to implement and understand than those from a humanoid robot; Nukovic et al., 2024) and thereby generate a first, general design space taxonomy that is applicable and adaptable across diverse robot appearances.

3.1.1 Materials and methods

The pilot study was approved by the university's ethics committee (ethics vote: EK-MIS-2025-0352-FT-d01), preregistered, and all participants provided written informed consent before participation. It was conducted as a laboratory study in a smart kitchen that features a robotic arm to fulfill users' voice-controlled requests (for a detailed description, see Section 4.1.1). The robotic arm was remote-controlled by a trained research assistant, resembling a Wizard of Oz design (Hein et al., 2024; Kelley, 2018). As an exemplary interaction with the smart room, users should order two cans of Dr Pepper soda from the room, which were supposed to be served by the robotic arm. However, not aligned with the participants' goal, the robotic arm served only one Dr Pepper soda and then an orange lemonade, because the second can of Dr Pepper soda was already empty. Thus, the information that one can was empty as a reason for serving the wrong drink was an issue of transparency to be communicated through the robot (or not). Following a between-subjects design, participants were randomly allocated to one of two experimental conditions. In the no-explanation condition, the robot did not use a gesture to explain why it served the incorrect drink. In the explanation condition, the goal-deviating delivery was accompanied by a gesture-based explanation, i.e., the robot brought the empty can of Dr Pepper soda to the participant, showed it to them, disposed of it in the bin, and only then served the orange lemonade.

After interacting with the robot, participants completed an online questionnaire, which measured understanding of the reason for the room's behavior, perception of technology transparency (Cronbach's α = 0.83), need for technology transparency (Cronbach's α = 0.91), awareness of the explanation, and secondary outcomes that will not be reported further (see Sections 4.1.4 and 4.1.5 for all measures and the OSF repository for the data set and analysis script). All participants could receive course credit or 6€ as study compensation.

Study participants were recruited through university platforms, mailing lists, and personal outreach. Initially, 50 participants completed the pilot study. After excluding 4 participants for incorrect responses to an attention check item and 4 participants for failing to recall crucial details of the study procedure (i.e., the beverages ordered and served), the final sample consisted of 42 participants (n = 22 in the no-explanation condition, n = 20 in the explanation condition). Participants' ages ranged from 18 to 74 years (M = 27.45, SD = 11.68). 52.4% of the participants identified as male, 45.2% as female, and 2.4% as diverse. 50.0% had a university degree, 42.9% a university entrance qualification, and 2.4% a secondary school certificate. 85.7% of the sample were students, 7.1% employees, 4.8% pensioners, and 2.4% self-employed.

3.1.2 Results and discussion

More participants indicated perceiving an explanation for the robot's behavior in the explanation condition (30% yes, 5% not sure) than in the no-explanation condition (100% no). Nonetheless, the percentage of people in the explanation condition who did not consciously notice the room's explanation (65%) was surprisingly high. However, more participants in the explanation condition (85%) than in the no-explanation condition (5%) correctly inferred the reason why the robot served a different drink than ordered.

Perception of technology transparency did not differ significantly between the no-explanation condition (M = 4.85, SD = 1.12) and the explanation condition (M = 4.48, SD = 1.01), t(40) = 1.13, p = 0.263, d = 0.35. Similarly, there was no significant difference in perceived transparency of the reasons for the system's behavior (“why” facet of technology transparency) between the no-explanation condition (M = 4.32, SD = 1.78) and the explanation condition (M = 4.85, SD = 1.81), t(40) = –0.96, p = 0.344, d = 0.30. Across both conditions, participants stated that no explanation for the robot's behavior (n = 13) and receiving the wrong beverage (n = 12) negatively affected the perceived transparency of the system. However, a significant but opposite effect was found for the “personal data” facet of transparency, showing that participants in the explanation condition perceived it as less transparent what personal data is used (M = 3.15, SD = 2.08) than those in the non-explanation condition (M = 4.46, SD = 1.87), t(40) = 2.14, p = 0.039, d = 0.66.

Participants reported a high need for technology transparency, significantly higher than the scale midpoint in both the no-explanation condition (M = 5.71, SD = 1.20), t(21) = 6.72, p < 0.001, d = 1.43, and the explanation condition (M = 5.54, SD = 1.29), t(19) = 5.33, p < 0.001, d = 1.19. However, there was no significant difference between need for technology transparency in both conditions, t(40) = 0.47, p = 0.644, d = 0.14.

In summary, though participants who saw a gesture-based explanation reported the reason for the robot's behavior more accurately than without the explanation, the system was not subjectively perceived as more transparent. First, only around a third of participants who saw the gesture-based explanation reported that the system explained its behavior. This could hint at gesture-based transparency cues being indeed more subtle cues than explicit, language-based explanations (Hein et al., 2024). Second, perceived technology transparency in the explanation condition did not differ significantly from the no-explanation condition, even showed a lower value for the explanation condition on one transparency facet and there was a heightened need for technology transparency in both conditions. As the subjective experience of transparency seems crucial in shaping user attitudes (e.g., trust, feelings of control or security) and behavior (e.g., willingness to use) (Hein and Diefenbach, 2025), this raises the question of how gesture-based explanations should then be designed to not only effectively communicate the actual reason for a robot's behavior but also create matching transparency perception.

3.2 The design space taxonomy of gestures for transparency in HRI

As shown in the pilot laboratory study, the design of gestures for transparency is no trivial task. Therefore, we aimed to take a more systematic view of the design space and explore how specific design options for robot gestures might affect users' transparency perception and understanding. As a starting point for future robot gesture design, we propose a 12-dimensional design space taxonomy, specifying how explanation content can be realized in a robot gesture (Eiband et al., 2018; Hein et al., 2023).

To develop the design space taxonomy, we combined inductive and deductive approaches in an iterative process. First, three researchers with psychological, human-computer interaction, and computer science backgrounds independently and inductively generated a broad set of explanatory gestures for different use cases based on everyday human-to-human interaction (e.g., pointing, waving, or mimicking object use), also considering the pilot study setting. Rather than directly deriving dimensions, we first compared the generated gestures to identify recurring characteristics that varied independently across examples (e.g., proximity to the user, change of location, number of objects involved). These characteristics were iteratively formed into dimensions by merging them when they described the same underlying property (e.g., technical characteristics of executing the same gesture, such as speed and broadness of movements), and discarding them when they were either too specific to individual gestures (e.g., retaining general repetition of a movement instead of the specific pointing frequency) or could not be meaningfully varied independently of other characteristics (e.g., object visibility was dependent on the proximity of the gesture and the used objects). This process resulted in an initial set of candidate dimensions describing complementary aspects of explanation gestures.

Then, these candidates were deductively refined based on previous work on non-verbal communication in HRI and human interactions. Specifically, we integrated research efforts that have distinguished gestures by robot proximity and movement (Leoste et al., 2024), technical features (Janowski and André, 2021), temporal sequences of steps, or directness of the gesture (Urakami and Seaborn, 2023). Additionally, the aforementioned classification system for human gestures that differentiates iconic, metaphoric, deictic, and beat gestures was considered (e.g., comparing iconic with deictic gestures indicates that concrete objects may be used or not) (McNeill, 1992). Candidate dimensions were renamed, and leading questions to describe them were revised where conceptual overlap was identified (e.g., user proximity was renamed to spatial relations because it encompasses the gesture's position relative to multiple relevant reference points rather than only the user).

Finally, the resulting taxonomy was discussed with additional experts to critically examine its conceptual coherence and the relevance of each dimension. During these discussions, three complementary levels of the dimensions solidified according to the primary aspect of the gesture they describe: the communicational level, capturing what explanation content is communicated; the spatial level, capturing where the explanation is performed relative to relevant entities; and the motor level, capturing how the gesture is executed. While individual gestures may instantiate several dimensions simultaneously (e.g., a gesture close to the user may be more direct), the dimensions describe analytically distinct properties that can vary independently across gestures and therefore provide complementary perspectives for characterizing explanation gestures (e.g., gestures close to the user may also be less direct when they refer to a symbolic meaning than directly showing an object). Table 1 summarizes the resulting initial design space taxonomy together with relevant questions for each dimension.

Table 1. The design space taxonomy of gesture-based transparency cues in robotic systems with guiding questions for each dimension.

4 Main study

Aiming to explore the proposed design space taxonomy, we applied it to design five gesture-based explanations for robot behavior that vary along the design space dimensions. Specifically, we compared the effects of these different robot actions on users' understanding and perception of technology transparency. We deliberately confronted participants with goal-deviating robot behavior in the same setting as in the pilot study to induce the need for an explanation, because explanations for robotic behavior become especially important in cases of erroneous or other goal-deviating actions (Das et al., 2021; Felzmann et al., 2019). Additionally, the need for transparency seems to be higher after encountering a barrier to goal achievement and a mismatch between expectations and the system's behavior (Hein and Diefenbach, 2025).

4.1 Materials and methods

4.1.1 Procedure

The study was approved by the university's ethics committee (ethics vote: EKMIS-2025-0403-FT-d01), preregistered on the OSF platform, and written informed consent was obtained from participants. Aiming for a larger and more diverse sample than in the laboratory pilot study, we implemented this study as an online survey using the Unipark software.

Participants should imagine being in a smart kitchen that is controlled by a central room intelligence (Diefenbach et al., 2020; Hein et al., 2024) which connects all appliances and smart home devices (e.g., the coffee machine, lights, and audio system). Instead of addressing each device separately, users speak to the kitchen as a whole, and the room intelligence automatically routes their commands to the corresponding component. A key component of this room is a robotic arm (rail-mounted Kinova Gen3 6DoF robotic arm, Kinova Robotics, 2019; with a RobotiQ 2F-140 gripper, Robotiq Inc., 2025; and 24V servo motors, controlled by an Elmo Gold Duo digital servo drive, Elmo Motion Control Ltd., 2025), which can perform simple tasks for users and assist them in their use of the kitchen (see Figure 1). For this study, participants watched a video of the robotic arm without sound (to ensure standardized viewing experiences despite differences in participants' audio playback conditions in the online setting), in which its movements were remotely controlled by a research assistant according to a fixed interaction script.

When starting the survey, participants read an introduction to the study and the aforementioned setting, learning general information on the room intelligence. Participants were then told to imagine they had just returned from work and were sitting in this kitchen with a friend. They should imagine feeling thirsty and therefore ordering two cans of Dr Pepper soda from the room intelligence, which should be delivered by the robot arm – one for themselves, and one for their friend. Next, participants viewed one out of six videos (depending on the experimental condition, see the OSF repository for all videos), which showed the robot arm's reaction to their request. The videos were on average 93 seconds long (SD = 10). They were filmed from the user's perspective, who sat at the kitchen table with the friend to their left side (see Figure 1), watching the robotic arm fulfill the request. The robot was shown to correctly deliver a can of Dr Pepper soda for the friend, but subsequently served the user with the wrong can (orange lemonade) from the kitchen shelf (see Figure 2). Depending on the experimental condition, this goal-deviating delivery was either accompanied by one of five different gesture-based explanations or left unexplained (see Section 4.1.2). Afterwards, participants completed a questionnaire on the study's measures (see Sections 4.1.4 and 4.1.5) and provided their contact details to receive course credit or to participate in a raffle of two 50€ vouchers as study compensation.

4.1.2 Experimental design

The study followed a between-subjects design with five conditions that featured different gesture-based explanations for goal-deviating robot behavior and one control condition (n = 37) without explanation (see Figure 3). Participants were randomly allocated to one of these six experimental conditions. The different gesture-based explanation conditions represented variations across the design space dimensions, as described in further detail below. More specifically, the following gesture-based explanations were implemented:

• Show condition (n = 38): The robot moves the empty can toward the table where the user is seated and tilts it so that the open lid is clearly visible to the user.
• Point condition (n = 37): The robot takes out the empty can and points to the lid without bringing it to the table.
• Shake condition (n = 45): The robot moves the empty can toward the table where the user is seated, turns it upside down, and shakes it.
• Trash condition (n = 40): The robot throws the empty can into a bin marked with a recycling symbol.
• Trash-implied condition (n = 38): The robot moves the empty can toward the recycling bin and hints at throwing it in, but does not actually release it.

In line with previous recommendations (Han et al., 2021), these gestures were designed to be immediate, concise, and human-like. They aimed to communicate the same reason for the robot's behavior (an orange lemonade was delivered because the second can of Dr Pepper soda was empty), but differed on dimensions of the design space taxonomy as to how an explanation was conveyed via robot gestures. Thereby, we especially focused on two dimensions that were prominent in our taxonomy development, as well as in HRI and human gesture literature (Leoste et al., 2024; McNeill, 1992; Urakami and Seaborn, 2023): the directness of the gesture and its spatial relations. Figure 4 schematically illustrates how the gesture-based explanation conditions reflect both dimensions. In the show condition, the object being central to the explanation (i.e., the empty can) is explicitly brought into the user's immediate space, creating a direct gesture in close proximity. Remaining close to the user but making the explanation inferable through symbolic motion rather than an explicit visual reference, the shake condition serves as an example for an indirect gesture in close proximity. Relying on a typical deictic gesture, the robot directly points at the open lid in the point condition as a reason for not delivering it, but remains far away from the user in the kitchen, representing a rather far and direct gesture. Lastly, throwing the can into the bin is a symbolic gesture far from the user, suggesting an indirect gesture in lower proximity. The trash-implied condition was additionally explored as an iconic gesture because it is even more indirect by merely suggesting a symbolic action. Conversely, we refrained from using a more extreme variation of the spatial relations dimension because an even higher distance from the user would render the gesture difficult for participants to see, especially in the video format, not representing a meaningful explanatory gesture design.

Importantly, while each condition was designed with respect to the focal dimensions, some secondary dimensions covary due to the physical constraints of the robotic system and the aim to produce ecologically valid gestures (see Table 2). More precisely, we prioritized robot gestures that align with typical human gesturing behavior over the definite causal interpretability of the results, as this approach seemed better suited to demonstrating the practical significance of gesture-based transparency design and to examining the transferability of human gestural behavior as a valuable starting point (Nukovic et al., 2024), using our design space taxonomy. Therefore, while the conditions are motivated by systematic variation in the design space taxonomy, they do not constitute a full factorial manipulation of individual dimensions; rather, they instantiate ecologically plausible combinations of dimension configurations, allowing us to exploratively investigate how common gesture strategies are perceived by users. Nevertheless, the dimensions modality (none other than the gesture), integration (none, each explanation stands on its own), user response (none), repetition (none), and technical execution (e.g., speed of the movement) were kept consistent across conditions to increase comparability and sensibly display gestures in the video format.

Table 2. Manifestation of the design space dimensions that covary across the gesture-based explanation conditions.

4.1.3 Participants

The sample was recruited via university platforms, mailing lists, personal outreach, and the SurveyCircle platform. Initially, 311 participants provided all necessary responses to the focal measures of perception of technology transparency and understanding. Subsequently, 30 participants were excluded for failing an attention check (“Please select the response option ‘Strongly agree' here.”), 29 for not recalling crucial study details (the original and received order of beverages, see Section 4.1.1), and 17 for both reasons. Therefore, the final sample consisted of 235 participants, exceeding the power analysis estimate of 216 participants (see preregistration for details).

Participants were between 18 and 65 years old (M = 28.16, SD = 9.64). 46.4% of the participants identified as female, 36.2% as male, and 2.1% as diverse. 41.7% had a university degree, 30.6% a university entrance qualification, 3.8% a PhD, 3.4% a secondary school certificate, 3.0% a vocational training, and 0.4% a habilitation as the highest level of education. 52.8% of the sample were students, 23.4% employees (including public officials), 3.4% self-employed, 1.3% job seekers, 1.3% pensioners, and 0.4% trainees. 15.3% of the participants did not report demographic information.

4.1.4 Primary measures

Perception of technology transparency was measured using the TechTra-P Scale (Hein and Diefenbach, 2025), which includes 10 items on a scale from 1 (strongly disagree) to 7 (strongly agree; Cronbach's α = 0.90), and an open question asking about decisive experiences for the transparency perception. Each item addresses a different facet of technology transparency (e.g., how a technological product works, how to use it). Of these facets, the focal interest of this study was in the “why” facet of technology transparency, that is, the perception of transparency concerning the reasons behind the technology's behavior (Hein and Diefenbach, 2025). This facet was measured by the second item of the TechTra-P Scale.

Moreover, and building on this “why” facet, we included an open-ended item (“Describe why you think the room behaved the way it did.”) to receive in-depth insights into users' understanding of the robotic system's behavior.

4.1.5 Secondary measures

As an additional outcome, we measured participants' need for technology transparency using the TechTra-N Scale (Hein and Diefenbach, 2025). This scale includes 10 items on a scale from 1 (strongly disagree) to 7 (strongly agree) as well as two open-ended questions asking what should be transparent about the technological product and what does not need to be transparent.

Moreover, we assessed several interaction-related variables as secondary measures: Perceived attractiveness of the room intelligence was measured using the AttrakDiff Mini (Hassenzahl et al., 2003; Hassenzahl and Monk, 2010), which consists of 10 items rated on a 7-point semantic differential scale, and the interaction principles using a 20-item questionnaire adapted from ISONORM 9241/110-S (Pataki et al., 2006) rated on a 7-point semantic differential scale. In addition, the current positive and negative affect was examined by the Positive and Negative Affect Schedule (PANAS; Watson et al., 1988), consisting of 10 items and a scale from 1 (very slightly or not at all) to 5 (extremely). Lastly, we measured intuitive use using the INTUI questionnaire (Ullrich and Diefenbach, 2010) with 17 items assessed on a 7-point semantic differential scale and intention to use using five items, each rated on a 5-point scale from 1 (strongly disagree) to 5 (strongly agree).

As control variables, we also included users' experience with robots with five items and their attitude toward robots with one item, rated on a scale from 1 (strongly disagree) to 7 (strongly agree) (Heerink et al., 2009; Koverola et al., 2022; Nomura, 2014). One question asked about which kinds of robots participants have experience with (i.e., humanoid, animal, industrial, or other types).

To ensure adequate data quality, there was also an attention check item and two items checking whether participants accurately recalled the intended and the actual order of beverages. All items, including those not reported in the main text, and Cronbach's alpha values (α = 0.54–0.91) can be found in Supplementary Table S1. In particular, the results regarding the gut feeling subscale of the INTUI as well as the conformity with user expectations and error robustness subscales of the ISONORM 9241/110-S should be interpreted with caution, as the subscales' Cronbach's alpha values were below 0.60 and could not be substantially improved by omitting single items.

4.1.6 Data analysis

The quantitative analyses were performed using R (R Core Team, 2025) (see data set and script in the OSF repository). We conducted one-way ANOVAs with the experimental conditions as the independent variable and perception of technology transparency (or one of the secondary measures) as the dependent variable. If the ANOVA revealed a significant effect, Tukey's Honestly Significant Difference (HSD) Tests (Tukey, 1949) with p-values adjusted for multiple comparisons were applied to identify which specific conditions differed significantly from one another. To test the effects' robustness, we also calculated exploratory ANCOVAs, including the experience with and attitude toward robots. Qualitative answers to open questions (e.g., users' understanding) were analyzed using qualitative content analysis (Kuckartz, 2014). Two researchers inductively developed and then applied the category system, resolving potential disagreements through discussion until consensus was reached. Because responses could be assigned multiple categories when appropriate, inter-rater reliability was calculated separately for each category, with Cohen's kappa ranging from κ = 0.85 to 0.94. Moreover, we jointly analyzed the focal measures of users' understanding and perception of transparency by an exploratory, independent two-sample t-test to compare participants who identified the correct reason for the robot's behavior with those who gave an incorrect reason, with respect to their perceived transparency.

4.2 Results

4.2.1 Qualitative understanding (primary measure)

The open-ended item, building on the “why” facet of perception of technology transparency, provided insights into how participants understood the robot's behavior and if they understood why the robot served the wrong drink. In the control condition, only 2.7% of participants identified the reason for the wrong delivery, namely that the second can was empty. In contrast, participants seeing an explanation gesture partly identified the intended explanation. Yet, the ratio of participants who correctly understood the gesture varied across the different gesture-based explanation conditions: It was highest in the show condition (44.7%) and the trash condition (32.5%), followed by the trash-implied condition (26.3%). Conversely, few participants correctly interpreted the gesture-based explanation in the shake condition (6.7%) and the point condition (2.7%). Thus, although these variations were found, less than half of the participants identified the correct reason for receiving the wrong drink in all conditions.

Across the conditions, the most frequently mentioned incorrect reasons for the robot's behavior included system malfunctions (e.g., difficulties in understanding the users' request, n = 80 mentions), causes related to the robot's programming (e.g., inability to distinguish between beverages, n = 32), incorrect handling of the system by the user (e.g., unclear instructions, n = 30 mentions), problems with the second can other than it being empty (e.g., being expired or damaged, n = 29 mentions), or autonomous behavior by the robot (e.g., detecting a change in user preferences and reacting accordingly, n = 23 mentions). 36 participants expressed confusion or stated not knowing why the system behaved the way it did.

Yet, the frequency with which participants cited these incorrect reasons differed between the experimental conditions (see Table 3). Notably, participants in the control, point, and trash conditions clearly attributed the robot's behavior to a malfunction more frequently than to any other reason. In contrast, the pattern in the show and shake conditions was not as pronounced, with smaller differences in the frequencies between the incorrect reasons. Furthermore, the shake condition was the only one in which malfunctions were not the most frequently indicated reason; instead, it was other problems with the can (aside from being empty). Regarding the trash-implied condition, although malfunctions were also mentioned the most, participants felt more often confused compared to the other experimental conditions.

Table 3. Frequency of incorrect reasons mentioned for the robot's behavior by experimental condition.

4.2.2 Quantitative perception of technology transparency (primary measure)

Regarding overall perception of technology transparency (M = 3.48, SD = 1.31), a one-way ANOVA revealed a significant difference between the experimental conditions, F(5, 229) = 3.58, p = 0.004, η² = 0.07. Contrasting single conditions, perception of technology transparency was higher in the show condition (M = 4.07, SD = 1.24) than the shake condition (M = 3.05, SD = 1.06), diff = 1.02, p = 0.005, d = 0.89 (see Figure 5, left bars). No further differences between conditions were statistically significant. The largest, yet non-significant, difference in perception of technology transparency to the control condition (M = 3.86, SD = 1.24) was observed for the shake condition, diff = 0.81, p = 0.051, d = 0.71. Similarly, although perception of transparency was descriptively higher in the show than the control condition, this effect was not statistically significant, diff = 0.21, p = 0.982, d = 0.17. Controlling for experience with robots and attitude toward robots in an exploratory ANCOVA did not change these results (see Supplementary Table S2). In addition, further exploratory analyses revealed that participants who correctly understood the reason for the robot's behavior perceived the robot as significantly more transparent (M = 3.98, SD = 1.27) than participants who did not (M = 3.36, SD = 1.27), t(66.87) = 2.96, p = 0.004, d = 0.49.

Regarding specifically the “why” facet of perceived technology transparency (M = 3.01, SD = 1.87), a one-way ANOVA revealed a significant difference between the experimental conditions, F(5, 229) = 5.31, p < 0.001, η² = 0.10. Perception of transparency regarding the reasons for the robot's behavior was higher in the show condition (M = 4.18, SD = 1.93) than the shake condition (M = 2.47, SD = 1.53), diff = 1.72, p < 0.001, d = 1.00, point condition (M = 2.86, SD = 1.86), diff = 1.32, p = 0.020, d = 0.70, trash condition (M = 2.58, SD = 1.82), diff = 1.61, p = 0.001, d = 0.86, and trash-implied condition (M = 2.66, SD = 1.65), diff = 1.53, p = 0.004, d = 0.85 (see Figure 5, right bars). Again, despite the descriptive difference and small to medium effect sizes, comparisons with the control condition were not statistically significant (p ≥ 0.151) with diff = 0.98, p = 0.187, d = 0.58 for comparing the control and shake condition and diff = 0.67, p = 0.678, d = 0.35 for comparing the control and show condition. Moreover, controlling for experience with robots and attitude toward robots in an exploratory ANCOVA did not alter the results (see Supplementary Table S2).

At the item level, differences were observed in three further facets of perception of technology transparency (see Supplementary Table S3). First, participants reported greater perceived transparency of the system's data basis in the show condition (M = 4.29, SD = 1.78) than the shake condition (M = 2.87, SD = 1.49), diff = 1.42, p = 0.005, d = 0.87, point condition (M = 2.97, SD = 1.82), diff = 1.32, p = 0.020, d = 0.73, and trash condition (M = 3.05, SD = 1.83), diff = 1.24, p = 0.030, d = 0.69. Second, perceived transparency of the system's predictability was higher in the control condition (M = 3.49, SD = 1.73) than the shake condition (M = 2.13, SD = 1.47), diff = 1.35, p = 0.004, d = 0.85; as well as higher in the show condition (M = 3.76, SD = 1.73) than the shake condition, diff = 1.63, p < 0.001, d = 1.02, and trash condition (M = 2.67, SD = 1.51), diff = 1.09, p = 0.048, d = 0.67. Third, participants perceived the transparency of the system's usefulness as higher in the control condition (M = 4.46, SD = 1.82) than in the shake condition (M = 3.11, SD = 1.71), diff = 1.35, p = 0.019, d = 0.77.

4.2.3 Need for technology transparency, attitude, and experience (secondary measures)

Descriptive statistics of all secondary outcomes overall and per experimental condition can be found in Supplementary Table S4. There was no significant difference between the experimental conditions in participants' need for technology transparency, F(5, 218) = 1.03, p = 0.401, η² = 0.02. Across all conditions, responses to the open-ended questions indicated a need to understand the robot's reasons for behaving in a certain way (n = 79), what data it uses and how (n = 62), or what it can do and how it can be useful (n = 51). Frequent concerns about predictability (n = 32), error handling (n = 31), reliability (n = 23), and safety (n = 18) were mentioned, particularly in the kitchen environment, where malfunctions could have severe consequences. Participants did not report a need for technical details such as internal algorithms (n = 78) or reasons for why/how the system works at a detailed level (n = 56) as long as it works reliably.

A one-way ANOVA revealed a significant difference between the experimental conditions in the robot's conformity with user expectations (ISONORM 9241/110-S), F(5, 203) = 3.53, p = 0.004, η² = 0.08. Conformity with user expectations was significantly lower in the trash condition (M = 3.63, SD = 1.30) than the control condition (M = 4.46, SD = 0.99), diff = 0.83, p = 0.044, d = 0.71, and show condition (M = 4.47, SD = 0.94), diff = 0.84, p = 0.039, d = 0.73. However, these results should be treated with caution, given a Cronbach's alpha of the subscale of 0.58. Apart from that, there were no significant differences between the experimental conditions on the other secondary outcomes (see Supplementary Table S4).

5 Discussion

Aiming to facilitate and inspire the design of intuitively transparent robotic systems, we propose a taxonomy to describe the design space for gesture-based robotic explanations. The design space taxonomy is based on insights into gestures in human communication and HRI, as well as an initial pilot study in which a gesture-based explanation did not significantly increase perceived technology transparency compared to a control condition without explanation. It features 12 dimensions, which are clustered into three levels, focusing on communicational aspects of explanations, the physical space in which gestures take place, and technical attributes of the robotic systems involved.

In a user study, we exploratively applied the design space taxonomy to design five gesture-based explanations for goal-deviating robotic behavior, focusing especially on the directness and spatial relations of the gestures. Participants watched a robotic arm in a smart kitchen serve a wrong drink (because the ordered can was empty). Following a between-subjects design, the robot's actions were either accompanied by one of five gesture-based explanations (shake, show, point, trash, or trash-implied condition) or no explanation gesture (control condition). In the show condition, the highest ratio of users correctly understood the gesture. Similarly, perception of technology transparency was highest in the show condition, eliciting statistically significant differences with large effect sizes relative to the other explanation conditions but not relative to the control condition. Contrarily, in the shake condition, some facets of transparency perception were rated lower than in the control condition.

These observed differences in users' understanding and transparency perception highlight the importance of careful, systematic design of gesture-based robotic explanations: On the one hand, gestures may indeed be helpful in making robot actions subtly and intuitively understandable, underscoring the relevance and potential of gesture-based transparency design, as argued in other work (Nukovic et al., 2024; Zhao et al., 2026). On the other hand, the participants' still insufficient understanding of the robots' behavior across conditions and the lack of differences relative to the control condition emphasize how difficult it is to design effective robot gestures, which aligns with previous research (Hirschmanner et al., 2021; Zhao et al., 2026). Moreover, the findings suggest that gesture-based explanations operate on at least two partially separable levels: shaping (1) users' inference of a robot's behavior (i.e., their understanding), and (2) their subjective experience that the system has meaningfully explained itself (i.e., their transparency perception). As improvements in understanding did not translate into significant increases in transparency perceptions in our investigations, these two outcomes were shown not to be necessarily as aligned as one might assume.

One possible explanation for this pattern of findings may be our chosen futuristic setting with an industrial robot arm, making the implementation of effective gesture-based explanations more difficult than with a humanoid robot in an everyday scenario (Nukovic et al., 2024). This may have led to the differences in the understanding of the robot's behavior not being strong enough to yield statistically significant differences in transparency perception. Moreover, as indicated by users' qualitative responses, most incorrect explanations of the robot's actions referred to technical reasons - particularly malfunctions - both overall and in the most experimental conditions. Users' focus on technical causes of the robot's behavior may stem from their lack of familiarity with the futuristic robotic system and uncertainty about its capabilities, which may have reduced transparency perceptions across all experimental conditions (Hein and Diefenbach, 2025). Yet, it seems that for gestures close to the user (show and shake) that are clearly visible, other attributions, such as object- or user-related issues (e.g., other problems with the can or incorrect handling), may also play a greater role. Nevertheless, the limited improvements over the control condition and the understanding ratios could further suggest that gesture-based explanations alone may not always provide sufficient transparency perception and understanding. Perhaps multimodal cues that combine gestures with complementary explanation modalities such as speech, visualizations, or projections may, at least in some settings, be better suited to explaining robot behavior (Kerzel et al., 2022).

However, although the results cannot be interpreted causally, they point to direct evidence-presenting gestures in proximity to the user as a promising explanation approach, while other explanation gestures may be ineffective or hinder rather than promote understanding and transparency perception (e.g., indirect, symbolic gestures). This may be because symbolic gestures require users to infer the meaning and could therefore introduce ambiguity or competing interpretations, whereas the show gesture directly presented relevant and clearly visible evidence (i.e., the empty can) and may thus have reduced interpretative burden and potential uncertainty (Urakami and Seaborn, 2023). The frequency of incorrect reasons mentioned by participants supports this assumption, even though differences between the experimental conditions were small: Feelings of confusion were less often reported in the conditions with direct gestures (show and point), while confusion was most frequently cited in the trash-implied condition as the condition with the most indirect gesture. The comparison between the show and point conditions further suggests that directness itself may take different forms. While both gestures directly referred to the causally relevant object, the show gesture more clearly presented the explanatory evidence by making the open lid more directly visible to participants. This difference may explain why the show gesture consistently outperformed the point gesture and suggests that directly presenting explanatory evidence may constitute a stronger transparency cue than merely directing attention toward it. Against the background of these interpretative discussions, our results indicate that the research field could benefit from a systematic framework for further exploration and therefore underlines the value of our proposed design space taxonomy as a starting point.

5.1 Theoretical implications

By varying the design of gesture-based explanations along dimensions of the design space taxonomy, we demonstrated that gesture attributes derived from the design space taxonomy can affect perceived technology transparency in general and users' understanding of the reasons for the system's behavior in particular. We thus show the potential of the design space taxonomy for exploring gesture-based explanatory designs in robotic systems. Additionally, no secondary outcomes, except for conformity with user expectations, significantly differed across the experimental conditions. Although these findings should be interpreted with caution due to low internal consistencies of some secondary outcomes (i.e., Cronbach's alpha of conformity with user expectations, error robustness, and gut feeling was below 0.60), they imply that technology transparency is distinct from other UX constructs (Hein et al., 2026), and that variations along the dimensions of the design space taxonomy do indeed primarily affect understanding and technology transparency.

However, results support the notion that not all explanatory cues are necessarily better for user understanding and experience (Diefenbach et al., 2022; Felzmann et al., 2019): Gesture-based explanations of robotic behavior may enhance users' understanding (see the show condition), but they might also be ineffective and even have the potential to mislead or confuse users (e.g., approximately one-third of the participants indicated perceiving the behavior as a malfunction) and reduce perceived technology transparency (see the shake condition). Thus, poorly designed gesture-based explanations may actively induce an incorrect understanding of the system's actions and lower perceived technology transparency. Moreover, the varying effects on individual facets of transparency suggest that while gestures may be tailored to enhance one focal facet (such as the “why” facet in our studies), they may undermine perceptions of other transparency facets. So far, the field lacks a comprehensive theoretical explanation for such differences in how robot gestures are perceived. However, we hope that the design space taxonomy serves as a useful starting point for systematic research into how gesture-based transparency cues should be designed to improve users' perception and understanding of technological behavior.

In this regard, our findings also suggest that users' understanding and transparency perception of explanatory gestures are related but separate outcomes. This reflects previous basic research showing differences between subjective and objective assessments (e.g., Rozenblit and Keil, 2002) and insights from transparency research that UX-related and mental model-related outcomes of explanatory designs may diverge, forming two perspectives (Hein et al., 2025). The differentiation has implications for transparency research more broadly: Users' understanding and subjective transparency perception should be treated as different variables, and evaluations of transparency interventions should consider monitoring both measures, as optimizing one does not necessarily optimize the other.

As another starting point, the user study offers preliminary indications for the idea that direct gestures near the user, including a clear evidence presentation (show condition), may be more effective than indirect, symbolic actions (e.g., shake condition) or actions performed further away (e.g., trash condition). Accordingly, the results for users' understanding and their transparency perception are consistent in the show condition in that, among the tested conditions, the highest proportion of participants attributed the correct reason to the robot's actions, and participants also reported the highest subjective transparency perception. This is in line with previous research, in which participants preferred concise and direct explanations of robotic behavior (Han et al., 2021; Mualla et al., 2022). Nonetheless, the findings should be regarded as exploratory, and users in the main study generally reported low to medium perceived technology transparency and a high need for transparency, especially regarding the safety-critical kitchen environment. Thus, our findings further highlight the relevance of research into the subtleties of designing transparent robotic systems that align with user needs.

5.2 Practical implications

Compared to other technological systems, robots are distinguished by their physical embodiment and ability to interact directly with their environment, providing designers with unique opportunities to create intuitively understandable, transparent robotic systems (e.g., by using gestures as explanations). The design space taxonomy introduced offers an initial, preliminary tool for exploring these opportunities, structuring different dimensions along which explanatory gestures in robotic systems can be designed. Instead of prescribing specific design solutions, it offers a systematic starting point for conceptualizing, comparing, and discussing alternative gesture-based explanatory cues. We hope this design space can inspire the development of gesture-based explanations across different application domains, including industrial robots (where transparency can support safe and effective collaboration, yet written explanations are often impractical) as well as domestic and service robots (where unobtrusive transparency may foster user trust and positive user experiences). However, further validation across different robot embodiments, interaction contexts, cultural settings, and real-world physical interactions is needed before the taxonomy can be considered a broadly applicable design framework.

Previous research has recommended including transparency considerations from the start of product development rather than as an afterthought (Felzmann et al., 2020; Hein et al., 2023). The design space taxonomy may support this process, encouraging developers to explore explanatory gestures at early stages of robotic development. This could be done, for example, as follows:

• Think of content which might be unclear to users and would benefit from being explained (e.g., why the system acts the way it does, how to use it, or which data it uses; Hein et al., 2026), referring for instance to typical use cases of the system and user stories.
• To explore gesture-based explanations of the robotic system, vary their features along the different design space dimensions to develop candidate and potentially promising gesture designs.
• Evaluate the users' understanding of the gestures and the fit between perceived transparency of the product and users' need for transparency (e.g., using the TechTra Scales; Hein et al., 2026). Thereby, attention should be paid not only to one key transparency facet (e.g., why a specific output occurred) but also to other facets that may be unintentionally negatively affected by the gesture (e.g., the predictability of the robotic system).
• If needed, refine the gesture-based explanations based on these evaluation results by iteratively revisiting the previous steps until your desired level of users' understanding and transparency satisfaction is reached. Rather than assuming that the design space taxonomy will automatically produce effective explanations, designers should treat its initial form as a structured means of generating and evaluating alternative design solutions.

Importantly, our results indicate that users' understanding and transparency perception may differ. Users may correctly interpret a robot's behavior without necessarily experiencing it as transparent, or perceive a system as transparent even with partial misunderstanding. Thus, it seems essential to consider both outcomes and, depending on the context, probably place greater emphasis on one of them. For instance, users' actual understanding could be especially considered when misunderstandings could have serious consequences (e.g., collisions with the robot), and users' transparency experiences could be more important when aiming to optimize a product from a UX perspective. In general, further research is needed to determine which manifestations of the design space dimensions may consistently support users' understanding and transparency perception across different robotic platforms and interaction contexts - and whether such consistent principles are even attainable.

5.3 Limitations and future research

Some limitations of this work should be addressed by further studies. First, in the main study, users did not interact with the robotic system. This setup was chosen to obtain a larger and more diverse sample and to maintain high internal validity, as the users viewed standardized videos of the robot that were comparable across experimental conditions. However, the video format, as opposed to personal interaction with the robot (like in the pilot study), might have made it harder for users to confidently evaluate the technology. Thus, the generalizability to real-life interactions should be investigated in further research. For this purpose and as a next step, researchers could use the Wizard of Oz method (Hein et al., 2024; Kelley, 2018) also applied in our studies to create realistic scenarios and further test different gesture designs before developing them as an automated feature of a robot and employing them in the field. This would initially reduce the required resources, enable the identification of potentially effective designs, and lay a sound basis for field studies that require more extensive preparation, such as initiating collaborations with employing institutions. Nevertheless, versatile real-life applications are crucial, as the setting or context may affect users' transparency experiences (Hein and Diefenbach, 2025), such as that the perception of gestures might vary in industrial settings with time constraints or background noise (Sauppé and Mutlu, 2014; Schött et al., 2023). In this regard, future research might also investigate how cultural contexts affect the interpretation of robot gestures, for example, by distinguishing between universally understood robot gestures (e.g., pointing) and culturally specific ones (e.g., the signal for “OK”, Kita, 2009).

Second, the way users see and interpret robot gestures might further differ for other robotic systems, such as more humanoid robots that are more strongly anthropomorphized than the robot arm we used (Nukovic et al., 2024). Following on from our setting, it would be interesting to investigate robotic systems that use similar arms but as parts of larger, more anthropomorphic robots (Fitter et al., 2020; Tröbinger et al., 2021). Such robots offer new opportunities to apply the design space taxonomy because they can, for example, move more independently or use two hands to gesture. In addition, even more humanoid robots, such as NEO (1X Technologies, 2026) or Nadine (Ramanathan et al., 2019), employed in different use cases (e.g., home assistance, manufacturing, service, elderly care) should be explored to further develop the taxonomy and compare results across robot appearances and applications.

Third, as just using gestures without additional explanatory cues led to the mixed pattern of findings described, combining them with further modalities, such as sounds, visualizations, or language-based explanations, also seems promising for future investigations (Aryania et al., 2026; Kerzel et al., 2022; Zhao et al., 2026). For example, while we chose not to include sound in the videos due to the online context, hearing the robot grab the empty can in our study may have made its behavior easier to understand, because the typical sound of lightly squeezing an empty can might directly indicate that the false delivery was due to the can's emptiness. In addition, gestures may be more suitable for conveying some facets of technology transparency (e.g., reasons or next steps) than others (e.g., data usage), further rendering the involvement of other modalities a useful addition - in particular as the second most frequently mentioned information needed by participants was what data the robotic system uses and how.

Lastly, readers should note that the design space taxonomy should not be understood as exhaustive, but rather as an inspiration and starting point for designing transparent robotic systems. As further dimensions may be useful to consider, we encourage product designers to add or omit dimensions as needed for a specific robotic system. Similarly, each dimension of the proposed design space taxonomy has more possible operationalizations than those exemplary tested in the study, and the dimensions' relative importance might vary in different user scenarios. In this regard, although the dimensions were designed to be distinct, the extent to which they are interlinked may depend on how they are implemented in practice. Thus, we call for future research to explore additional dimensions and further operationalizations of existing dimensions of the proposed design space taxonomy. This may also contribute to further understanding the dynamics between users' understanding and transparency perceptions and to identifying gesture designs that may increase both outcomes sufficiently.

6 Conclusion

In summary, this article developed and empirically explored an initial design space taxonomy for gesture-based explanations of robotic behavior. The proposed taxonomy comprises 12 dimensions as a starting point for systematically generating, varying, and evaluating robot gestures. An exploratory user study showed the potential of this design space taxonomy by comparing different gesture-based robotic explanations, revealing that some gestures may enhance users' understanding, while others might be ineffective or even counterproductive for understanding and transparency perceptions. Specifically, despite the exploratory nature of this study, direct gestures in close proximity to the user, including an explicit evidence presentation, might be more effective at explaining the reasons for robotic behavior than indirect, distant gestures with symbolic actions. Yet, our findings also highlight the difficulty of gesture-based design because of still insufficient user understanding and a lack of differences relative to a control condition without explanation. Further complicating this difficulty, the results point to users' understanding and perception of transparency as separate outcomes that may not align and therefore should be considered simultaneously when designing explanatory robot gestures. Thus, the results not only point to the need for careful use of gestures for conveying transparency in robotic behavior, but also to the theoretical and practical relevance of a systematic design space taxonomy for this purpose. We therefore offer an initial taxonomy to facilitate further systematic research into gesture-based robotic behavior by varying gesture design along the dimensions of the taxonomy. At the same time, the design space taxonomy may serve as a tool for thinking about and designing robotic transparency in a structured way in practice.

Correction note

This article has been corrected with minor changes. These changes do not impact the scientific content of the article.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher's note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fcomp.2026.1869586/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Curtis Gittens, The University of the West Indies, Cave Hill, Barbados

Reviewed by: Tim Schreiter, Technical University of Munich, Germany

Yi Zhao, The University of Sydney, Australia