”It Brought the Model to Life”: Exploring the Embodiment of Multimodal I3Ms for People who are Blind or have Low Vision

BLV people face challenges accessing graphical information, which impacts education opportunities (Butler et al., 2017), makes independent travel difficult (Sheffield, 2016), and causes disengagement with culture and the creative arts (Bartlett et al., 2019). These barriers can lead to reductions in confidence and overall quality of life (Keeffe, 2005).

Graphical information can be made available in formats that improve non-visual access. Traditionally, accessible graphics – known as raised line drawings or tactile graphics – have been used to assist BLV people in accessing information. Tactile graphics are frequently used to facilitate classroom learning (Aldrich and Sheppard, 2001; Rosenblum and Herzberg, 2015) and orientation and mobility (O&M) training (Blades et al., 1999; Rowell and Ungar, 2005). Work has been conducted on the development of interactive tactile graphics, including the NOMAD (Parkes, 1994), IVEO (ViewPlus, [n. d.]), and Talking Tactile Tablet (Miele et al., 2006; Inc, [n. d.]). These devices combine printed tactile overlays with touch-sensitive surfaces, enabling BLV users to explore the graphics tactually and access audio labels by interacting with predefined touch areas. However, as these systems rely on printed tactile graphics, their scope is limited to two-dimensional content.

3D-printed models are an increasingly common alternative to tactile graphics. They enable a broader range of material to be produced, particularly for concepts that are inherently three-dimensional in nature. In recent years, the production cost and effort of 3D-printed models have fallen to levels comparable to tactile graphic production. 3D-printed models are increasingly being applied in various accessible graphic areas: mapping and navigation (Gual et al., 2012; Holloway et al., 2018, 2019b, 2022; Nagassa et al., 2023); special education (Buehler et al., 2016); art galleries (Karaduman et al., 2022; Butler et al., 2023); books (Kim and Yeh, 2015; Stangl et al., 2015); mathematics (Brown and Hurst, 2012; Hu, 2015); graphic design (McDonald et al., 2014); science (Wedler et al., 2012; Hasper et al., 2015); and programming (Kane and Bigham, 2014). Compared to tactile graphics, 3D-printed models have been shown to improve tactual understanding and mental model development among BLV people (Holloway et al., 2018). However, as with traditional tactile graphics, the provision of written descriptions or braille labelling presents challenges. The limited space on models and the low-fidelity of 3D-printed braille can significantly impact the utility and readability of labels (Brown and Hurst, 2012; Taylor et al., 2015; Shi et al., 2016).

To address labelling challenges, limitations on the type of content that can be produced, and to create more engaging and interactive experiences, there is growing interest in the development of interactive 3D-printed models (I3Ms). By combining 3D-printed models with low-cost electronics and/or smart devices, many I3Ms now include button or touch-triggered audio labels that provide verbal descriptions of the printed model (Landau, 2009; Shi et al., 2016; Reichinger et al., 2016; Giraud et al., 2017; Götzelmann et al., 2017; Holloway et al., 2018; Ghodke et al., 2019; Davis et al., 2020). Such I3Ms have been applied across various BLV-accessible graphic areas, including: art (Holloway et al., 2019a; Iranzo Bartolome et al., 2019; Butler et al., 2023); education (Ghodke et al., 2019; Shi et al., 2019; Reinders et al., 2020); and mapping and navigation (Götzelmann et al., 2017; Holloway et al., 2018; Shi et al., 2020). I3Ms with audio labels are especially useful for BLV users who are not fluent braille readers. Stored as text and synthesised in real-time, audio labels are easier to update compared to labels on non-interactive models. Many I3Ms also support multiple levels of audio labelling (Holloway et al., 2018; Shi et al., 2019; Reinders et al., 2020), extracted through unique button presses or touch gestures, enabling them to convey far more information than the written descriptions supplied alongside tactile graphics and non-interactive models.

I3Ms are inherently multimodal. Multimodality can improve the adaptability of a system (Reeves et al., 2004), and when modalities are combined, they can increase the resolution of information the system conveys (Edwards et al., 2015) and enable more natural interactions (Bolt, 1980). For BLV users, combining modalities has been shown to improve confidence and independence (Cavazos Quero et al., 2021). Modality adaptability allows BLV users to choose interaction methods based on context, ability, or effort. For example, a user may be uncomfortable engaging in speech interaction in public due to privacy concerns (Abdolrahmani et al., 2018), opting instead to use button or gesture-based inputs. Richer resolutions of information can be achieved when combining modalities, e.g.,  the tactile features of a 3D-printed model with haptic vibratory and auditory outputs. Combining modalities is critical to overcoming the ‘bandwidth problem’, in which BLV users’ non-visual senses cannot match the capacity of vision, necessitating their combined use (Edwards et al., 2015).

While early I3Ms primarily relied on button or gesture-based triggered audio labels, recent research has explored integrating speech interfaces and conversational agents. This shift is driven by research finding that BLV people find voice interaction convenient (Azenkot and Lee, 2013), along with widespread adoption (Pradhan et al., 2018) and high usage (Abdolrahmani et al., 2018) of conversational agents among BLV users. For instance, Quero et al. (Cavazos Quero et al., 2019) combined a tactile graphic of a floor plan with a conversational agent that focused on indoor navigation; however, voice interaction was performed through a connected smartphone rather than the graphic itself. Other works have developed voice-controlled agents to guide BLV users in exploring 3D-printed representations of gallery pieces (Iranzo Bartolome et al., 2019; Cavazos Quero et al., 2018). These systems, however, have primarily focused on basic command-driven interactions more analogous to voice menus rather than conversational dialogue. Shi et al. (Shi et al., 2017, 2019) proposed incorporating conversational agents into I3Ms to allow BLV users to expand their understanding of the modelled content.

Recent research into I3Ms has begun to explore modalities beyond audio and touch. Quero et al. (Cavazos Quero et al., 2018) designed an I3M representing an art piece that integrated localised audio, wind, and heat output. However, participants faced challenges in interpreting the semantic mapping of modalities, e.g.,  whether heat represented the morning sun or the shine of starlight. In our previous work, we found that BLV users desired I3Ms that combined touch, haptic vibratory feedback, and conversational dialogue (Reinders et al., 2020). Additionally, we co-designed an I3M with BLV co-designers to explore how these modalities could create natural interactions (Reinders et al., 2023), inspired by the ‘Put-That-There’ paradigm (Bolt, 1980). This work led to five I3M design recommendations, including: support interruption-free tactile exploration; leverage prior interaction experience with personal technology; support customisation and personalisation; support more natural dialogue; and tightly coupled haptic feedback. These studies motivated our current work, with participants finding the I3M engaging, and several beginning to embody it.

Dourish (Dourish, 2001) presents a seminal view that embodiment “denotes a form of participative status”, where embodied interaction in natural forms of communication is influenced both by physical presence and context. They posit that this perspective applies to “spoken conversations just as much as to apples or bookshelves”. Within the design of agent-based systems, embodiment is largely understood as the use of different modalities – e.g., voice, visual output, gestures, gaze – to imbue machines with human-like behaviours and characteristics, making them appear more ‘alive’; with an enhanced perception of social presence that is capable of approximating human-human social interaction (Cassell, 2000; Lester et al., 1997; Lankton and Tripp, 2015; Biocca, 1999). Such agents are often described as being more ‘lifelike’ or possessing ‘lifelikeness’ (Lester et al., 1997; Cassell et al., 1999; Cassell and Vilhjálmsson, 1999; Cassell and Thorisson, 1999; Lester et al., 1999).

Research into embodiment has predominantly focused on sighted users. As systems become more embodied, users’ perceptions of social presence can increase, motivating users to treat them more favourably (Reeves et al., 2004). Lankton et al. (Lankton and Tripp, 2015) described how social presence can make systems appear more sociable, warm, and personal, while Cassell (Cassell, 2001) noted that embodying technology allows users to locate intelligence, illuminating what would otherwise be an ‘invisible computer’. Importantly, embodied agents exhibiting higher levels of social presence and perception have been shown to enhance user perception of engagement (Shamekhi et al., 2018; Luger and Sellen, 2016; Heuwinkel, 2013; Cassell, 2001) and trust (Bickmore and Cassell, 2001; Shamekhi et al., 2018; Bickmore et al., 2013; Minjin Rheu and Huh-Yoo, 2021).

In the HCI community, efforts to embody intelligent agents have focused on enhancing social perception through conversational, visual, and physical attributes. Conversational embodiment includes mimicking human voices (Cassell, 2001), small talk (Liao et al., 2018; Pradhan et al., 2019; Cassell, 2001; Shamekhi et al., 2018; Bickmore and Cassell, 2001), greetings (Cassell, 2001; Luria et al., 2019; Shamekhi et al., 2018; Lester et al., 1997), and conversational turn-taking (Cassell, 2001; Kontogiorgos et al., 2020). Lester et al. (Lester et al., 1997) described the persona effect, proposing that social presence can increase when agents exhibit personality, making them appear more lifelike. Many conversational agents, like Siri, incorporate attributes of conversational embodiment.

Visually embodied agents, which are often also conversationally embodied, associate systems with virtual avatars or characters, many of which have faces (Cassell, 2001; Kontogiorgos et al., 2020; Bickmore et al., 2013), and are capable of gesturing (Cassell, 2001; Bickmore and Cassell, 2001) and gaze (Kontogiorgos et al., 2020; Shamekhi et al., 2018; Bickmore and Cassell, 2001). Visual embodiment extends beyond the visual feedback that mainstream conversational agents emit, e.g., rings of light or animations used to indicate that agents are processing or ‘thinking’. Depending on the task, visually and conversationally embodied agents can improve social perceptions (Shamekhi et al., 2018; Luria et al., 2019; Cassell, 2001; Nowak and Biocca, 2003), trust (Bickmore and Cassell, 2001; Bickmore et al., 2013; Minjin Rheu and Huh-Yoo, 2021; Shamekhi et al., 2018), and engagement (Shamekhi et al., 2018; Cassell, 2001).

Embodied agents can extend beyond virtual embodiment and include physical bodies (Kontogiorgos et al., 2020; Luria et al., 2017). Lura et al. (Luria et al., 2017) found that users’ situational awareness was higher when using a physically embodied robot to perform smart-home tasks compared to an unembodied voice agent. Robots capable of emitting human-like warmth have been associated with increased perceptions of friendship and presence (Nie et al., 2012). Recent research has explored the use of haptic vibratory feedback to create lifelike cues, such as heartbeats (Borgstedt, 2023) and handshaking (Bevan and Stanton Fraser, 2015). Physically embodied agents can be perceived as having higher social presence than non-physically embodied agents (Kidd and Breazeal, 2004). Additionally, they have been found to be more forgivable during unsuccessful interactions; however, depending on their realism, thi can become distracting in high-stakes scenarios (Kontogiorgos et al., 2020).

The design of embodied agents has traditionally focused on the perception of sighted users. However, in the last decade, work has begun exploring how human-human conversational cues can be converted into non-visual formats for BLV users. Many of these efforts utilise haptic belts/headsets (Rader et al., 2014; McDaniel et al., 2018) or AR glasses (Qiu et al., 2016, 2020) to convey body movements like head shaking, nodding, and gaze. Despite this, the impact of agent embodiment, and specifically embodied I3Ms, on BLV users’ perceptions remains unstudied.

In our previous work, we found that a number of participants desired I3Ms that felt more lively and human (Reinders et al., 2023). Participants suggested integrating a conversational agent with a personality, incorporating haptic vibratory feedback to make the I3M feel alive, and enabling speech to originate directly from the model itself. This desire for more human-like interactions aligns with findings with conversational agents. Choi et al. (Choi et al., 2020) reported that many BLV users valued human-like conversation with conversational agents as critical in relationship building, while Abdolrahmani et al. (Abdolrahmani et al., 2018) observed that BLV users preferred agents they could talk with as if they were other people rather than pieces of technology. Karim et al. (Karim et al., 2023) recommended that agents have customisable personalities, recognising that such features may not be relevant in all scenarios, such as group settings. Collins et al. (Collins et al., 2023), however, found hesitance among BLV users towards embodied AI agents in VR applications. Whether these perspectives and desires extend to I3Ms is unknown.

Further impetus for studying I3M embodiment comes from the links between embodiment, engagement, and trust with embodied agents that have been previously identified for sighted users. Trust and engagement are especially critical for BLV users, as the usefulness of accessible graphics, aids, or tools depends on users’ willingness to engage with and accept/rely on the information they provide. This directly determines the extent to which users rely on these tools (Phillips and Zhao, 1993; Wu et al., 2017; Abdolrahmani et al., 2018). Therefore, it is crucial to explore whether I3Ms can be effectively embodied and whether embodiment fosters greater trust and engagement between users and their I3Ms.

To investigate conversational and physical embodiment, we created two I3Ms – (1) the Egyptian Pyramids and (2) the Saturn V Rocket [Figure 3]. These subjects were selected because they represent the types of materials commonly found in museums, galleries, or in science or history classes. Additionally, they facilitated the design of models with multiple components that could be individually picked up, detached, and manipulated, which has been shown to increase engagement (Reinders et al., 2020). Each I3M can be configured in two states – High Embodied Mode (HEM) or Low Embodied Mode (LEM) – based on five non-visual design factors (Sections 3.2 &  3.3).

The base of each I3M was constructed from laser-cut acrylic, serving as a stand to hold the constituent components of the I3M and to house a Raspberry Pi, speaker, and microphone (Figure 2). The Pi powered each I3M and handled the following responsibilities:

• •

  Maintaining a Wi-Fi connection with each I3M component and operating as a message broker (MQTT) to facilitate messaging between the Raspberry Pi and each component.

• •

  Managing the embodied state of the I3M by setting the design factors to operate in either HEM or LEM mode.

• •

  Controlling speech and auditory output through the connected speaker and microphone, using the Picovoice Porcupine wake word library and Google Cloud Speech-to-Text and Text-to-Speech for speech input and synthesis.

• •

  Connecting to the I3M’s conversational agent, built using Google Dialogflow, and integrating ChatGPT to perform external searches triggered by Dialogflow’s fallback intent.

Each I3M had four 3D-printed components. For the Saturn V Rocket, these were the Stage A, Stage B, and Stage C modules, and the Launch Tower. For the Egyptian Pyramid, these included the Sphinx, Great Pyramid, Pyramid of Menkaure, and Pyramid of Khafre. Each component had an ESP32 microcontroller embedded in the print, providing localised touch and movement sensing and haptic vibratory output. The microcontroller had integrated Wi-Fi, capacitive touch-sensing GPIO pins, and was connected to a touchpoint 3D-printed with conductive filament, a 3.7V 400mAh lithium polymer battery, an MPU6050 accelerometer and gyroscope, a DRV2605L haptic motor controller, and a vibrating haptic disc.

In earlier work, we observed that I3M conversational agents should speak with high-quality, human-like voices (Reinders et al., 2023). When designing our I3Ms, the researchers made the decision that voice quality should be independent of embodiment state, to prevent preferences for high-quality voices from overpowering other design factors.

I3Ms are inherently designed to be perceived physically, i.e.,  picked up and tactually observed or manipulated. We sought to explore ways in which the tangible nature of I3Ms could be enhanced by physically embodying presence.

We designed a controlled user study to explore and understand whether I3Ms configured with different embodied design factors influence BLV end-users’ perceptions of model embodiment, engagement, and trustworthiness. The study utilised both the Saturn V Rocket I3M and Egyptian Pyramids I3M, each configurable into two states – High Embodied Mode (HEM) and Low Embodied Mode (LEM) – as described in Section 3. We hypothesised that:

• •

  H#1: HEM I3Ms are perceived as more embodied than LEM I3Ms

• •

  H#2: HEM I3Ms are perceived as more engaging than LEM I3Ms

• •

  H#3: HEM I3Ms are perceived as more trustworthy than LEM I3Ms

Twelve BLV participants were recruited from our lab’s participant contact pool (Table 1). This sample size ($n=12$) falls within the range commonly seen in BLV accessibility studies, which often involve anywhere between 6-12 participants (Holloway et al., 2022; Nagassa et al., 2023; Shi et al., 2017, 2020) due to the low incidence of blindness in the general population and associated recruitment challenges (Butler et al., 2021).

Participants ranged in age from 27 to 78 years ($\mu$ = 50, $\sigma$ = 16.6). Nine participants self-reported as totally blind, while three reported being legally blind with low levels of light perception. Participants also varied in their prior experience with tactile graphics: Many reported substantial exposure and confidence ($n=5$), some reported some use but lacked confidence ($n=4$), and others reported limited or no exposure ($n=1$ and $n=2$, respectively).

Familiarity with 3D-printed models was slightly less common. All participants regularly used conversational agents, including Google Assistant ($n=11$), Siri ($n=10$), Alexa ($n=4$), and ChatGPT ($n=4$). These interfaces were accessed on various devices, such as smartphones ($n=12$), smart speakers/displays ($n=11$), smartwatches ($n=6$), computers ($n=7$) and tablets ($n=5$).

We used a series of questionnaires and asked participants to subjectively rate the I3Ms in order to investigate our hypotheses.

For Hypothesis H#1, to measure how embodied participants perceived the I3Ms, we utilised the Godspeed Questionnaire Series (GQS) (Bartneck et al., 2009). Originally devised to measure users’ social perceptions of robot and agent-based systems, GQS subscales such as  anthropomorphism and intelligence have recently been applied to measure aspects of embodiment and liveliness in conversational agents  (Shamekhi et al., 2018) and robots with human-like abilities (Kontogiorgos et al., 2020). We selected four subscales – anthropomorphism, animacy, likeability, and intelligence – as we felt each provides insight into components of embodied sociability. For example, anthropomorphism captures the attribution of human-like characteristics, animacy reflects perceptions of liveliness, likeability gauges formation of positive impressions, and intelligence focuses on perception of ability. Additionally, we also explored model embodiment by asking participants for their subjective perceptions by expressing the concept of ‘embodiment/embodied’ using the terms ‘lively/liveliness’ to ensure clarity. We felt this terminology would have more meaning to participants, and aligns with previous works that have used the similar terms ‘lifelike/lifelikeness’ to describe embodied agents (Lester et al., 1997; Cassell et al., 1999; Cassell and Vilhjálmsson, 1999; Cassell and Thorisson, 1999; Lester et al., 1999).

To explore Hypothesis H#2, we used two engagement measures – the User Engagement Scale [Short Form] (UES-SF) and the Playful Experiences Questionnaire (PLEXQ). The UES-SF measures user engagement as the depth of a user’s perceived investment with a system (O’Brien, 2016). It consists of 12 five-point Likert items across four subscales – focused attention, perceived usability, aesthetic appeal, and reward (O’Brien et al., 2018). The UES-SF has been widely applied across HCI to measure engagement in contexts such as interactive media (Carlton et al., 2019), video games (Wiebe et al., 2014), and 3D-printed building plans for BLV people (Nagassa et al., 2023). We used all four subscales. To complement the UES-SF, we also used PLEXQ, which measures playfulness, pleasurable experiences, and playful engagement (Boberg et al., 2015). PLEXQ is commonly used to assess how engaging games and game-like experiences are (Bischof et al., 2016; Cho et al., 2024). We used eight subscales that we felt were most relevant to I3Ms – captivation, challenge, control, discovery, exploration, humor, relaxation, and sensation.

To supplement these measures of perceived engagement, we also captured two behavioural metrics (time spent and interactions performed during tasks), as more time spent with the model and more interactions may indicate greater immersion and enjoyment (O’Brien and Lebow, [n. d.]; Doherty and Doherty, 2018). Note that comparing these between the LEM and HEM conditions was meaningful as the length of responses in both conditions was similar and every interaction type was available in both modes.

To investigate Hypothesis H#3 and measure trust, we utilised the Human-Computer Trust Model (HCTM). The HCTM conceptualises trust as a multifaceted construct, encompassing users’ perceptions of the perceived risk, benevolence, competence, and reciprocity during interactions. These perceptions can influence users’ reliance on a system and their likelihood of continued use (Siddharth Gulati and Lamas, 2019). Previous work has demonstrated the utility of HCTM in assessing trustworthiness in conversational agents like Siri (Gulati et al., 2018), machine learning systems (Guo et al., 2022), and other human-like technologies such as large language models (Salah et al., 2023) and chatbots (Degachi et al., 2023).

Based on pilot study feedback, we made minor adjustments to specific subscale items in the UES-SF and PLEXQ measures. For instance, the aesthetic appeal (UES-SF) and sensation (PLEXQ) subscales were adjusted, as concepts of “attractiveness”, “aesthetics”, and “visuals” held little meaning to our pilot user in BLV contexts. They recommended adding the phrase “to my senses” to these items. Per  (O’Brien et al., 2018)’s guidance on modifying the UES-SF, we did not report an overall UES score, instead focusing on individual components of engagement. These adjustments also motivated our decision to add our own questions gathering participants’ subjective ratings of the I3Ms, supplementing the validated scales and providing additional nuance and insight. The modified UES-SF and PLEXQ, along with GSQ and HCTM, are provided in the Appendices.

Our user study used a within-subject design. All participants were exposed to (one) LEM and (one) HEM-configured I3M. To control for bias related to model type (Rocket/Pyramids) and design factor state (LEM/HEM), the order in which the I3Ms were presented, along with their associated LEM/HEM configuration, was counterbalanced. The activities that participants completed with each I3M remained the same, regardless of its LEM/HEM state, and could be completed in either mode without significant difficulty.

Each user study session lasted approximately two hours and included at least one researcher being present. Sessions began with the researcher providing an overview of the research project, and were divided into the following stages:

1. (1)

   Training. Participants were guided through a 10-minute training exercise, which allowed them to familiarise themselves with how I3Ms operate. We designed and built an I3M that represented a non-descript sphere for training. Participants were first asked to explore the I3M tactually when it was turned off, before being taught how to extract basic information from the I3M using touch gestures, and asking questions through the conversational interface. The training I3M did not operate in either LEM or HEM mode and would, using a low-quality synthesised voice, only respond by confirming when interactions had been successfully performed (e.g.,  “Double Press”, “Recording query”). Operating outside of LEM/HEM states was a deliberate design decision to allow training of basic interaction functionality without biasing future LEM/HEM exposure.
   

2. (2)

   Exposure to LEM/HEM I3Ms.

   1. (a)

      Activities. Participants were introduced to their first I3M, configured in one of the LEM/HEM states, and completed a walkthrough activity. The I3M identified and described each component it included, with participants given the opportunity to tactually explore the I3M throughout. These walkthroughs were carefully curated so that participants encountered the majority of system functionality specific to the HEM/LEM design factors¹¹1The only exception to this was the small talk component of the introductions and small talk design factor, which, unlike all other design factors that were explicit, required user initiation (see Section 3.2.1).. After the walkthrough, participants were given up to five minutes to explore the models, during which they could interact with the I3M in any way they wished. Participants were then asked to complete four researcher-directed information-gathering tasks (e.g.,  finding out how long it took to build the Great Pyramid, what happened to the nose of the Sphinx, or the significance of the Pyramids). Participants could access this information using their choice of either touch gesture interaction or using the model’s conversational agent. Before concluding, participants were given up to an additional five minutes for a free play task designed to mimic undirected, real-world use. They were instructed to discover something interesting about the modelled concept that they were not aware of prior.

   2. (b)

      Questionnaire Scales. Participants were taken through the questionnaire scales – GSQ, PLEXQ, UES-SF, and HCTM. In addition, participants were asked to rate how lively, engaging, and trustworthy the I3M was, using 5-point Likert scales. On average, these took 15 minutes to complete.
      

   3. (c)

      Remaining Model. Participants would then complete 2(a) and 2(b) again with the second I3M, configured in the remaining LEM/HEM state. Participants spent an average of 30 minutes with each I3M.
      

3. (3)

   Semi-Structured Interview. At the end of the session, participants were asked questions about specific interactions with the LEM/HEM I3Ms and were asked to rank how lively, engaging and trustworthy each model was. We also asked about the role each design factor played, and whether they impacted perceptions of how lively, engaging, and trustworthy the I3Ms were. On average, it took 20 minutes to answer these questions.

All sessions were video-recorded and subsequently transcribed. Collected data included responses to scale questions, semi-structured interview questions, and participant comments made during task completion. The time participants spent completing tasks, as well as the number of interactions they performed, were also recorded. Descriptive statistics were calculated on all subscale responses – GSQ, UES-SF, PLEXQ, and HCTM – as well as for interview questions that ranked the I3Ms and individual embodiment design factors.

For data that did not follow a normal distribution (e.g., our scale data), we conducted non-parametric statistical tests, specifically Wilcoxon signed-rank tests. Binomial tests were performed on the rankings of the I3Ms and the impact ratings of individual embodiment design factors. Paired t-tests were conducted on data that was normally distributed (e.g., time taken to complete tasks). We opted for one-tailed tests because our hypotheses were explicit in nature (described in Section 4.1). This approach was deliberate, allowing us to dedicate more power to detecting effects in one direction.

The analysis should be interpreted in light of both the exploratory nature of our user study and our small sample size ($n$ $=$ 12). This influenced our approach in two ways. First, from the outset, due to our small sample size, we decided not to use $p$ $<$ 0.05 as the sole determinant of significance (Shamekhi et al., 2018; Cramer and Howitt, 2004), instead marking results $p$ $<$ 0.05 as significant (^∗) and 0.05 $<=$ $p$ $<=$ 0.1 as marginally significant (^∧). Second, given the exploratory nature of our work, in order to reduce the risk of overlooking meaningful results (false negatives), we chose not to apply corrections for multiple comparisons (although this does increase the risk of false positives).

Participants felt that HEM I3Ms were more lively, with an increased perception of embodiment compared to the LEM I3Ms. When ranking the I3Ms on liveliness, 8/12 selected HEM I3Ms (Section 5.1.2). This was also statistically significant, with most participants providing emphatic explanations for their selections, supporting H#1.

Results from the Godspeed Questionnaire Series also supported H#1 and allowed us to explore different dimensions and key aspects of embodiment, adding nuance to our understanding of how the I3Ms appeared embodied (Table 2). All GQS subscales were statistically significant. It is particularly noteworthy that the anthropomorphism and animacy subscales were rated more positively for HEM I3Ms, as these subscales deal with the attribution of human-like behaviours and perception of life (Bartneck et al., 2009), which are key components of embodiment. The likeability and intelligence subscales were also perceived more favourably for HEM I3Ms.

Overall, our findings and participant comments help to support H#1 that BLV users do perceive I3Ms configured with HEM design factors as more embodied, ‘present’, and ‘lively’. These findings align with work on the embodiment of other interfaces in non-BLV contexts, including robots (Kontogiorgos et al., 2020) and conversational agents (Luria et al., 2017; Shamekhi et al., 2018; Luria et al., 2019), which has found that imbuing machines with human-like behaviours can enhance their perceived embodiment and sense of presence, creating a sense of ‘being there’ (Shamekhi et al., 2018).

Participants rated higher levels of engagement with HEM I3Ms compared to LEM I3Ms. When ranking which I3M was more engaging, the overwhelming majority (10/12) selected HEM (Section 5.2.2). These differences were statistically significant, supporting H#2.

We used the User Engagement Scale and Playful Experiences Questionnaire (Table 4) to explore different dimensions of engagement to add to and help contextualise our understanding of how engaging participants found the I3Ms. The UES-SF indicated that HEM I3Ms had higher average scores than LEM I3Ms across 3/4 subscales. Of these subscales, two were statistically significant, and one was marginally significant. These subscales focus on the extent to which users feel absorbed in an interaction (focused attention) and the usability/negative affect experienced during interactions (perceived usability), both of which are critical dimensions of engagement (O’Brien et al., 2018). The aesthetic appeal subscale, which the researchers adjusted to be more meaningful in BLV contexts – encompassing aesthetics beyond those purely visual – was also statistically significant. This is particularly noteworthy, with HEM-configured I3Ms providing more engaging sensory experiences. The reward subscale, centred on valued experiential outcomes, was not significant with respect to our hypothesis, and was the only subscale where LEM I3Ms received higher mean scores. This may indicate that the additional presence and feedback of HEM I3Ms could, in some circumstances, reduce initial curiosity.

It is our belief that several PLEXQ subscales may have had reduced meaning to participants, potentially as a result of the controlled nature and limited time exposure of the study. For example, subscales that focus on finding something hidden (discovery) or unwinding through playful experiences (relaxation) may have been less relevant. Interestingly, one subscale, exploration, elicited higher mean values for LEM I3Ms, possibly for similar reasons to the UES-SF reward subscale. Despite this, several PLEXQ results added support to H#2, with subscales related to excitement (sensation), enjoyment/amusement (humor), and power (control) being more favourably observed with HEM I3Ms.

Across all tasks, participants spent more time interacting with the HEM I3Ms and performed a greater number of interactions compared to the LEM I3Ms. We feel this was particularly noteworthy during the undirected free play task, which was designed to mimic real-world use. Time spent and the number of interactions were statistically significant and marginally significant, respectively. These metrics have previously been used as behavioural measures of engagement in HCI (O’Brien and Lebow, [n. d.]; Doherty and Doherty, 2018), and provide further evidence that participants were more engaged with the HEM I3Ms. They also complement the UES-SF, which ties engagement to the depth of a user’s investment with a system (O’Brien, 2016).

Based on emphatic participant discussions and clear preferences when directly asked, our study provides evidence supporting H#2 that HEM I3Ms are perceived as more engaging. Our scale data adds important nuance to this understanding, with key UES-SF and PLEXQ subscales showing that HEM I3Ms were more favourably perceived. Although ratings for both HEM and LEM I3Ms were generally high, which aligns with works showing that BLV users find 3D-printed models and I3Ms engaging (Nagassa et al., 2023; Reinders et al., 2020; Shi et al., 2019), our results suggest that HEM I3Ms are perceived as even more engaging. More broadly, our findings align with prior work demonstrating that interfaces imbued with more human-like behaviours can positively impact end-user perceptions of engagement (Lester et al., 1997; Cassell et al., 1999; Shamekhi et al., 2018).

Our participants were mixed on whether there were any discernible differences in the trustworthiness of HEM and LEM I3Ms. Hesitance was observed when participants were asked to rank which I3M was the more trustworthy model, with 9/12 indicating no major discernible difference between them (Section 5.3.2). Participants expressed indecisiveness, often basing their interpretation of trust solely on the believability of the information output from the models, rather than their interactions with the I3Ms.

On the other hand, results from the Human-Computer Trust Model (Table 7) suggest that HEM I3Ms may lead to greater trust, as all four HCTM subscales were either significant ($=3$) or marginally significant ($=1$). Results did, however, reveal only very minor differences in mean scores between HEM I3Ms and LEM I3Ms. Despite this, these scales relate to the willingness of users to engage with a system despite the possible risks (perceived risk), whether a system possesses the functionalities needed to depend on it (competence), and a willingness to spend more time using it when support situations arise (reciprocity). The benevolence subscale, which was marginally significant, focuses on whether users believe that a system has the abilities required to help them achieve their goals.

Despite the statistical significance of the HCTM results, we believe that, overall, our results provide only mixed support for Hypothesis H#3. It is the researchers’ view that the HCTM questionnaire and subjective ratings may have been interpreted differently by participants. While the scale data appears to have successfully captured different dimensions of trust based on interactions with the HEM/LEM models, many participants, when asked about the concept of trust subjectively (Section 5.3.2), focused solely on the believability of the information provided by the models, independent of their HEM/LEM state. They tended to prioritise believability of information over considerations as to whether their interactions with the models and their behaviours influenced perceptions of reliance or competence. Our findings also suggest that the HEM design factors tested do not appear to play a strong role in influencing trust. As visually embodied conversational agents have been shown to be subjectively more trustworthy in non-BLV contexts (Heuwinkel, 2013; Bickmore et al., 2013; Sidner et al., 2018; Shamekhi et al., 2018; Gulati et al., 2018), it is clear that more research is needed in order to better understand the impact of embodiment on trustworthiness of I3Ms for BLV users, particularly in real-world situations.

Broadly speaking, our embodiment design factors and their implementations were well-received, contributing to how embodied and engaging the HEM I3Ms were perceived. However, connections to trust were mixed. These are discussed below in order of importance.

DF#4: Embodied Vibratory Feedback. Participants shared their enthusiasm regarding haptics. Referring to the liveliness of the HEM I3Ms, P3 described how the use of haptics made them feel “more three-dimensional… realistic”. Participants also discussed how haptics made the HEM I3Ms more engaging, including P5, “[haptics] added an extra sense of interaction”. These emphatic reactions align with other work, where haptics added life to I3Ms (Reinders et al., 2023), and support the growing body of research exploring how haptics can create physically embodied, lifelike cues (Nie et al., 2012; Bevan and Stanton Fraser, 2015; Borgstedt, 2023).

Our study supports the larger idea that I3Ms are able to be perceived as more embodied and are more engaging. This follows from H#1, where our findings support that BLV people perceive HEM I3Ms as more embodied and lively, and H#2, where HEM I3Ms were found to be more engaging. In the context of I3Ms, it appears that model embodiment increases end-user engagement, specifically with I3Ms using the set of embodied design factors we implemented. However, the impact of embodiment on trust (H#3) remains less clear, as BLV people generally appear to trust I3Ms regardless of embodiment.

Prior research has shown that the embodiment of an interface can increase both engagement (Shamekhi et al., 2018; Luger and Sellen, 2016; Heuwinkel, 2013; Cassell, 2001) and trust (Bickmore and Cassell, 2001; Shamekhi et al., 2018; Bickmore et al., 2013; Minjin Rheu and Huh-Yoo, 2021). Our findings support the relationship between embodiment and engagement, but stop short of any impacts on trust. With the bulk of this previous research focusing on conversational agents and robots, our work extends this relationship to include contexts relevant to both I3Ms and with people who are BLV.

The relationship between embodied I3Ms and their impact on engagement is particularly significant in learning contexts, allowing BLV students or self-learners to engage in deeper, more meaningful experiences where the I3M can assist in teaching and testing their knowledge. Embodied agents have been used in learning environments, and have been found to enrich the learning experiences of students (Schroeder et al., 2013). Pedagogic agents often rely on embodied designs, including combinations of visual and conversational embodiment (Lester et al., 1997; Moreno et al., 2001; Schroeder et al., 2013). While prior research on non-embodied I3Ms in learning contexts has highlighted positive feedback from both teachers and BLV students (Shi et al., 2019), we believe that understanding how I3Ms can be effectively embodied could further enhance engagement in educational settings, and facilitate broader adoption.

Based on the positive reception amongst participants of our embodied I3Ms, we propose extending our existing I3M design recommendations (Reinders et al., 2023) to include embodiment:

• •

  Support more embodied experiences: I3Ms that support human-like behaviours and characteristics in their design can make them appear more human, lively, and engaging to use. This may involve combining aspects of physical and conversational embodiment, e.g.,  introductions and small talk, embodied voices, embodied narration, embodied haptic vibratory feedback, and location of speech output.