Dancing with an AI partner in virtual and mixed reality

Abstract

Consumer extended reality (XR) headsets allow users to interact in both virtual reality (VR) and mixed reality (MR) views. However, there is limited research on how users’ experiences differ between VR and MR. In this study, we explore differences in the context of dancing through a mixed-methods user study (N = 36), in which participants danced with an AI partner in both VR and MR. Qualitative data were gathered through semi-structured interviews, augmented with quantitative self-report data on presence. On average, our users experienced a slightly greater presence in VR, but there were considerable individual differences in preferences for VR and MR. For instance, some users prefer the immersion VR provides and are more comfortable in a virtual environment where they are less aware of their own bodies and dancing. In contrast, others find it safer to move when seeing their surroundings in MR.

1 Introduction

Extended reality (XR) is an umbrella term for virtual reality (VR) and mixed reality (MR). Currently, XR is perhaps the most significant technological foundation enabling new forms of embodied interaction, due to the proliferation of low-cost consumer headsets such as Meta Quest 2 (Meta, 2020) and 3 (Meta, 2023). For instance, researchers have invented and evaluated various types of novel exergames (Karaosmanoglu et al., 2024), locomotion approaches (Krekhov et al., 2018; Cmentowski et al., 2019; Griffin and Folmer, 2019; Kontio et al., 2023; Evin et al., 2020), and avatars (Krekhov et al., 2019) and studied how to transform non-VR games into full-body VR experiences (Kirjonen et al., 2025). On the commercial side, the rhythm game Beat Saber (Games, 2019) is the best-selling Meta Quest title of all time (Hayden, 2025) and has been estimated to correspond to real-life tennis in terms of energy expenditure (VR Exercise Tracker, 2018).

Despite the considerable body of XR research and commercial applications, a topic that remains unclear is how the user experience of movement and physical activity differs between VR and MR, i.e., between a purely virtual view and a view where the virtual content is merged with the real world captured through cameras mounted on a headset.¹ Such MR functionality is supported by the most recent generation of headsets, such as Meta Quest 3 and Apple Vision Pro, but has received relatively little attention from the research community.

Regarding motivation, this study actually began as an interactive dance AI development project. The perhaps surprising choice to study VR–MR differences was made after the research team serendipitously tested the AI prototype in both VR and MR on a Meta Quest 3 headset that we had acquired. We felt that the AI partner appeared somehow more strikingly present and engaging in MR. The difference was difficult to articulate, and we felt that this was a more interesting topic to study than the AI dancer itself.

Our user study results reveal that, on average, participants experienced slightly greater presence in VR, at least as measured using the IPQ. However, preferences between VR and MR were almost evenly split, and we observed noteworthy individual differences. For instance, some users prefer the immersion that VR provides and feel more comfortable and less self-conscious in a virtual environment, where they are less aware of their own body and movements. In contrast, others find it safer to move when seeing their surroundings in MR and appreciate feeling more embodied when seeing their own arms and legs. We believe that our results can inform designers aiming to create XR applications that cater to differing user preferences. Furthermore, the user feedback data we report should be useful for others developing interactive dance partners.

2 Related work

When moving in VR, the user is completely immersed in a digital environment. Previous work has established that this can offer a sense of escapism and privacy, which may encourage self-expression and creativity (Hagedorn et al., 2023; Kang et al., 2023; Bernal and Maes, 2017). In contrast, MR overlays the digital content onto the real-world surroundings. This allows users to remain more grounded and context-aware in their physical surroundings (Li et al., 2019; Gyawali, 2023). However, there is virtually no research exploring differences between VR and MR in dancing, which is often perceived as an activity performed not only for personal enjoyment but also for social interaction and performance.

2.1 Dancing in HCI

Zhou et al. (2021) conducted a systematic literature survey on technology and dancing in Human-Computer Interaction (HCI) research of the last two decades. They discussed how technology is used in dance creation, analysis, and performance. The literature they cover primarily focuses on professional dancing, ranging from live performance to the training of dancers, where technology is used either to assist dancers in their performance, to create novel dance experiences, or to quantify and analyze dance technique and body movements.

Most of the virtual experiences designed for casual use fall under exercise-focused and casual follow-along games, such as Dance Dance Revolution (Hoysniemi, 2006) and Just Dance (Cavalcanti, 2020); in these cases, users are instructed on how to dance and may generally be inexperienced. Research also exists on how to improve such casual “on-the-fly” dance guidance for novice dancers (Laattala et al., 2024).

2.2 Dancing with virtual agents

For this study, we are primarily interested in dancing with an interactive virtual agent. On a high level, technology itself may be treated and interpreted as an agent if its behavior is sufficiently ambiguous, prompting agency attribution, although it may also be interpreted as a tool for the dancer, thereby establishing a different type of relationship (Fdili Alaoui, 2019). Additionally, interaction with performing agents may prompt the exploration of new motions (Bisig et al., 2016; Kirsh, 2010). This applies to both cases: collaborative dancing with an agent perceived as more autonomous and scenarios in which increased sensory and expressive capacities, enabled through the use of tools, enhance the dancer’s ability to interact and self-express.

Our agent is a virtual character in an XR experience about dancing with another person in a casual recreational setting, such as a disco, suitable for people with no prior dance training. Recreational dancing in HCI appears to be addressed and investigated comparatively much less frequently in HCI research, even though there are clear positive use cases and motivations (Piitulainen et al., 2022).

Existing studies on interactive agents in VR in a dancing context also appear to lean toward dance learning or performance, even if extending to the inclusion of novice users. For example, users can practice learning to dance Salsa (Senecal et al., 2020) or engage with a more abstract improvised interaction based on upper-body gestures (Batras et al., 2016). Casual recreational dancing appears to take place primarily in social VR games, such as VRChat (Piitulainen et al., 2022). However, in these environments, users dance with other people rather than with virtual agents. As such, recreational dancing with high-quality interactive agents in VR remains relatively unexplored.

2.3 Animating dance for virtual agents: the promise of AI animation

One reason for the relative rarity of dancing with virtual agents in XR may be that it can pose a difficult computer animation problem. Composing fluid dance movements is difficult using traditional techniques such as motion graphs (Kovar et al., 2023)—i.e., by stitching together separate animation clips—especially if real-time interactivity with a human user is needed, which calls for modern AI and machine learning approaches.

Many AI and machine learning approaches for dance motion synthesis have been published in recent years (Chen et al., 2021; Li et al., 2022; Ghosh et al., 2021; Ghosh et al., 2025), but they are typically not applicable to real-time use on limited computing hardware such as the Meta Quest headsets. In this study, we demonstrate that the motion matching algorithm (Clavet, 2016; Holden et al., 2020), commonly used in the games industry, can provide high-quality, real-time interactive XR dance animation and allows for dance-specific performance optimization.

More generally, AI and machine learning for character animation are advancing rapidly, and for comprehensive reviews of the vast spectrum of methods and approaches, we refer the reader to Kwiatkowski et al. (2022), Mourot et al. (2022), and Abootorabi et al. (2025).

2.4 Comparing VR and MR in dance and exercise

Kirakosian et al. (2021) conducted the only prior study comparing augmented reality (AR) and virtual reality for dancing, specifically pair dance training. They found that practicing dance in VR was perceived as more fun, and participants felt they could better accomplish what they were asked to do. This study was conducted using an Oculus Rift headset (VR, 2012) and a Magic Leap (Magic Leap, 2010) device for hand tracking. These are already quite old technologies, and their limitations, such as limited hand tracking volume, affected the study results.

Outside of dance, there exist more studies comparing VR and MR, focusing on topics such as heritage tourism (Guo Y. et al., 2024), memory recall for data visualization (Hurter et al., 2024), and surgical simulation (Lungu et al., 2021). Relevant to dance and full-body movement, Guo Z. et al. (2024) found that enabling passthrough in VR exergames improved players’ performance and awareness of their surroundings but slightly reduced their sense of immersion. Participants also favored playing in open spaces, and those with higher self-consciousness felt more comfortable using passthrough, highlighting its potential to make VR gaming more accessible in shared environments.

Zhou et al. (2024) found that while doing data analysis, participants felt safer and walked more in AR than in VR. On the other hand, participants were able to concentrate better in VR, felt less frustration, and used more varied postures. Allcoat et al. (2021) also found that in education, VR provided users a higher feeling of positive emotion and presence. VR can also help users judge the dimensions of 3D objects better than other current extended reality technologies (Banerjee et al., 2021).

A key difference between VR and MR is that in MR, one can see one’s own real body. Although this can be an advantage, the ability to inhabit a different body is also part of the appeal of VR. Some users want to look like themselves (Morris et al., 2023), but many also want the options to change their appearance (Javier et al., 2024). Although the avatar’s visuals or visibility do not necessarily affect the user’s sense of presence, they appear to influence social presence (Huang et al., 2023; Kruse et al., 2023). Tang and Bailey (2024) also found that having a less realistic avatar resulted in stronger identification with it and greater interest in some activities. Borgwardt et al. (2023) found that not seeing one’s hands can result in reduced accuracy in some tasks. Virtual avatars also allow easier interactions with other virtual avatars, even in ways that would not be possible in the physical world (Pratticò et al., 2023; Kurzweg et al., 2021; Oh and Kim, 2020).

3 AI dance partner

This section describes our AI dance partner, shown in Figure 1 and the supplemental video. To enable replication and extension of our work, our implementation is available as a Unity game engine project at https://github.com/CarouselDancing/AIDancePartner, including both the source code and the associated dance movement dataset we captured.

FIGURE 1

3.1 Interaction design

We designed the AI dance partner to engage in simple social dancing with the user, focusing on interaction that does not require physical touch, which is challenging to emulate with consumer XR hardware. The goal given to the movement AI was to keep dancing while staying approximately in front of the user, at a distance comfortable for casual social dancing. If the user turned around or moved in the space, the partner should reposition itself accordingly while still dancing.

We also added gaze animation such that the partner attempted to maintain eye contact with the user occasionally, but not so frequently as to be disturbing. Unfortunately, the hardware we had available did not allow tracking of the user’s eyes.

3.2 Visual design

Visual design plays an important role in XR experiences and can also affect users’ presence in both VR (Jicol et al., 2023) and MR (DeMarbre et al., 2024). We decided on simple and stylized character visuals, partly because of limited resources and partly because aiming for highly realistic human characters increases the risk of the uncanny valley (Mori et al., 2012).

3.2.1 Character

We allowed participants to choose between a masculine and feminine dance partner (Figure 2). This feature was added after initial testing, during which some participants reported that it felt awkward or intimidating to dance with a partner of a particular gender.

FIGURE 2

3.3 Movement AI: Motion Matching optimized to dance

From the beginning, it was clear that the main technological problem to be solved in this project was how to synthesize plausible and interactive dance movements for the AI dance partner within the limits of the computing capabilities of consumer XR headsets.

After surveying various machine learning and AI approaches for motion synthesis (e.g., Kwiatkowski et al., 2022; 168), we settled on a data-driven approach called Motion Matching (Clavet, 2016) as the core technical solution. Motion Matching constructs new animation sequences from available data based on requirements such as a target future trajectory for the character’s root. Motion Matching can also be considered a simple generative AI approach, and since its introduction in 2016, it has become a standard approach for many modern games with expressive and complex character animation, e.g., Control, Last of Us, and For Honor (Ilkka Kuusela, 2021; Mach and Zhuravlov, 2021; Clavet, 2016).

The basic premise of Motion Matching is that instead of painstakingly building animation graphs from separate animation clips, one can simply provide it an unstructured motion dataset, such as a single long motion capture clip, where the actor performs all the movements, such as walking, running, and turning to different directions. However, although Motion Matching is highly generalizable in principle, the approach must, in practice, be tailored and optimized for each use case (Mach and Zhuravlov, 2021; van den Heuvel et al., 2021). To share and discuss best practices, various Motion Matching case studies have been presented at game industry conferences (Ilkka Kuusela, 2021; Mach and Zhuravlov, 2021; Clavet, 2016).

In its basic form, Motion Matching simply searches over all the movement data several times per second, comparing user-specified motion requirements to outcome predictions pre-calculated for each frame (Clavet, 2016). Once a best-matching frame is identified, an animation transition is initiated toward the frame using standard techniques such as a linear blend. Notably, a new transition can be initiated before the previous transition has been completed; in effect, a character can be continuously transitioning toward a newly selected target frame. The associated jitter can be smoothed by tuning the transition parameters, such as blend time. In a way, Motion Matching is embarrassingly simple, and it can be somewhat counterintuitive that it works so well.

3.3.1 Optimizing Motion Matching for dance

Naturally, a brute-force search over all motion data can become computationally expensive as more data are added. One way to optimize Motion Matching is to replace the brute-force search with neural networks (Holden et al., 2020). In our case, we observed a simple but previously unreported optimization opportunity: in rhythmic dances such as house or disco, which we focusing on, the set of valid transition target frames can be pruned dramatically, directly translating to improved runtime performance. Specifically, at a current motion data time , we only allow transitioning to times , where is the game’s frame rate, i.e., is the time the animation would progress to without a transition. is the duration of a single bar of music, and is an integer. In addition to greatly reducing the search space, this ensures that the character’s movements always stay on beat. If the captured original motion did not stay on beat, the data were manually annotated with beat information, which was then utilized to modify the search.

3.4 Movement data

To capture the dance movement data needed for Motion Matching, we hired two professional dancers with several years of experience in dance styles such as hip-hop and vogue. We recorded dance motion capture data using both OptiTrack and XSens systems. This included recording house music data of a single dance move at 125 BPM and more diverse disco dance data at 120 bpm.

To guide the capture sessions, we used custom dance cards (Zadziuk, 2016), illustrated in Figure 3. Dance cards can provide a valuable visual guide and memory aid when capturing multiple movements in a single take. We instructed the motion capture actors to dance while moving along following the different trajectories represented by the cards and told them that they could additionally improvise as they wished. Additionally, we recorded some simple walking data for scenarios such as the agent waiting for the user or moving toward the user from a distance.

FIGURE 3

3.5 Software and hardware

The AI dance partner was implemented in the Unity game engine. We focused on developing for the Meta Quest 3 headset, which is well supported by Unity and is currently the dominant low-cost standalone headset, enabling high-quality MR.

3.5.1 Dance environment

For the VR condition, we designed a brightly lit dance studio environment featuring an obstacle-free dance floor and skylights oriented toward a bright, semi-stylized sky (Figure 4, left).

FIGURE 4

Under the MR condition, the participants’ real environment MR was a simple and mostly empty laboratory space with a table, chairs, and large glass windows, along with a one-way mirror between the space and a monitoring room (Figure 4, right).

4 User study

Based on our initial testing within the development team, it appeared that the mixed reality version might elicit higher levels of presence and connectedness with the AI partner and also feel safer to dance in as users are more aware of possible obstacles in the space. On the other hand, mixing the virtual avatar with the real environment might make the MR feel more artificial, given the non-photorealistic appearance of the AI partner. Moreover, dancing in MR may make the participants more aware of the study facilitator and, therefore, may make dancing feel less natural and more awkward. Therefore, our study design avoided making directional hypotheses, and we instead consider our experiment an exploratory, rather than confirmatory, study.

4.1 Sample size

A priori power analysis using G*Power 3 (Faul et al., 2007) was used to determine the total sample size necessary for the study. For two-tailed paired-samples -tests, a sample of 34 participants is required to detect a medium effect size (Cohen’s ), with a type I error rate of 0.05 and 80% power.

4.2 Participants

A total of 36 participants were recruited using different communication channels inside Aalto University. Every participant was compensated with a 20-euro gift card to a campus restaurant. Nine of the participants were between the ages of 18 and 25, twenty-two were between the ages of 26 and 35, four were between the ages of 36 and 45, and one participant was between 46 and 55 years old. A total of 17 participants were male, and 19 were female. All participants were able-bodied, and all except two had at least some VR experience.

4.3 Procedure and data

The study involved dancing in both VR and MR, with the order of conditions counterbalanced. The total study duration was approximately 45 min.

The study began with a research team member informing participants about the study and prompting them to sign a consent form. After putting on the headset, the participant could choose between the two possible AI partner characters. All participants danced to the same song with the AI dance partner in both VR and MR for up to 2 min per condition. The order of the conditions was counterbalanced. Based on internal pilot testing, we concluded that the time would be enough for the participants to compare the conditions. Furthermore, dancing for a longer duration would have risked excessive fatigue and discomfort, considering the potentially diverse levels of fitness and dancing abilities among the participants.

After each condition, the participant removed the headset and completed the IPQ. The order of questions in the IPQ questionnaire was randomized for each participant and condition. Presence was chosen as the primary quantitatively measured construct due to our own experience of the MR condition feeling surprisingly and strikingly more “real,” despite the stylized visual design of the AI dancer. We were not sure how to exactly probe and operationalize the feeling, but presence appeared to be at least one aspect of it. We acknowledge that immersive user experience is multidimensional and that our decision to focus on presence does not imply that it fully captures all experiential differences between VR and MR; rather, it serves as a focused quantitative anchor that is complemented by qualitative analysis to capture the broader experiential nuances observed between conditions.

4.4 Methods

The interviews were recorded, and the recordings were automatically transcribed using OpenAI Whisper (Radford et al., 2023), run locally for data privacy. The transcripts were verified manually. The transcripts were then analyzed using reflexive thematic analysis (Braun and Clarke, 2021). The first author coded the data, and three authors collaboratively reviewed and refined the codes and themes. Two of these authors also facilitated the study and were present during the interviews. In line with reflexive thematic analysis, coding was treated as an interpretive and iterative process rather than a procedure aimed at achieving inter-rater agreement. Discussions among the authors were used to critically examine, refine, and challenge the developing codes and themes. The involvement of multiple authors, including those who facilitated the study, was leveraged as a reflexive resource to surface potential biases and ensure that interpretations remained grounded in the data.

The quantitative data were analyzed using two-tailed t-tests to identify differences between the VR and MR conditions.

5 Results

5.1 Interview data

The thematic analysis resulted in a total of 314 codes, which were grouped into 8 broader themes. Table 1 provides a summary of the themes. The themes are described in more detail in the following subsections. The results are discussed in relation to the research questions in Section 6.

5.1.1 General experience

The participants’ opinions were split on whether VR or MR provided a better overall experience. A total of 19 participants preferred VR, while 17 preferred MR. Many participants understandably thought that MR was more realistic (“…[MR] was realistic because, I mean, I felt that the environment was the same as I [in] was before.”[P11], “But the first one [MR] was largely more realistic.”[P5]), but it was even more common for participants to point out that the VR version felt more real to them (“I would say because I personally feel in the virtual environment, I feel more have the sense of real because the background is not the same as the background in the lab. So I was feeling like I was in the dancing room, and that together with the avatar.” “I would say that the VR one was a bit more realistic, and sometimes I couldn’t distinguish between the real, so I would say that that is the difference.”).

MR was generally perceived as less smooth than VR (“I felt like the animation and the dance was a lot jankier in the real world one, it was a bit jankier.”[P29], “there was a bit of a frame, like a frame lag issue, like the screen was flickering when it was in the actual environment”). This is probably partly caused by the technical limitations of the Meta Quest 3 headset. On the other hand, one participant also pointed out how the not-quite-realistic visuals felt dreamy (“maybe it was just the camera, but the camera made the real world seem somewhat fairy-tale-ish or a bit dreamlike in the pass-through version”[P16]).

5.1.2 Safety and confidence

Many participants felt safer in MR because they could see their surroundings (“I felt more safer in MR because I could see where everything was”[P5], “you can see everything around [you], and it’s much safer”[P22]). On the other hand, many participants felt more confident in VR because of the opposite, i.e., because they could not see their surroundings or their own body (“I did not care where my hands were, what moves I was doing. I felt happy to be there. I felt happy to just dance…”[P5], “So maybe I felt more confident in the second version because I couldn’t judge myself by watching myself moving and dancing.”[P14]). However, some of the participants had the opposite reaction (“with the AR [MR] one, I was more confident …because I knew where I am.”[P15], “in VR, you would feel like you’re totally immersed in it. So maybe VR would, I would feel a bit more safe.”[P7]). It appears that feelings of safety and confidence are intertwined in an experience like ours.

5.1.3 Environment

One way the VR experience was considered better was that the partner avatar appeared less out of place than in MR (“…this like stronger crash between something that I obviously can see that it’s not real and my real surroundings”[P26], “…second one [VR] felt much more immersive in the sense that it felt much more coherent as an experience, that kind of like everything was kind of stylized, saturated, bright, and then like we didn’t have this dissonance between like the real world and the virtual world, and everything felt cohesive.”[P1]) Many also complimented the visuals of the VR environment (“I particularly like the virtual environment, the dancing room …it’s so beautiful.”[P9], “…the space was pretty simple and it was very easy to like…”[P29]), but for some, the space felt artificial (“The second one [VR] felt a bit artificial.”, “first one [VR] was more artificial”[P21]). The MR environment was more often called realistic and even boring (“So that [MR] was one that it felt real. So realistic, definitely.”[P5], “with the see-through version, I have less of this disconnection or two-world phenomena that I feel like I’m both in the virtual world and the real world. So I can push the awareness of these two different places away when I’m in the see-through version. But it’s also maybe, I don’t know, a bit boring.”[P19]). One participant pointed out how they preferred MR because the sounds they heard such as their own steps matched better to the visuals they were seeing.

“…I actually did like the pass-through version a bit more …like I could sometimes hear the way my shoe sounded on the floor but in the virtual world my brains got caught up in the fact that that’s not sound a shoe would make on this type of floor but when it was the pass-through version I was like yeah everything is consistent I don’t feel disconnected from this version of the reality…[P16]”

5.1.4 Embodiment

Many participants complained about the VR version because they could not see their own hands (“One thing I would say, the hands in the virtual environment. I really wanted to see my hands.”[P4], “if I saw my hands in the picture, in the environment, I’m sure that it would definitely had a much more captivating experience for me.”[P14]) or their body in general (“Virtual is more like strange and disembodied. Compared to Passthrough.”[P34], “With the VR version, it was a bit weird to not be able to see your own body.”[P35]).

The real-world environment can also affect the experience in MR. A particular difference between MR and VR related to our laboratory space was that some participants noticed the one-way mirror and could see their own reflection (“…when I saw my own body and especially my own reflection, it was really interesting. I felt much more real in that.”[P34], “I could also see my reflection, which I think add a lot to it, like in terms of like how confident I feel, felt moving around and doing stuff”[P29]). However, some did not find seeing the reflection desirable since it made them self-conscious and having the partner have no reflection can break the immersion (“the fact that I could see my reflection in the mirror disturbed my comfortability”[P32], “having the character in the same space kind of also breaks that [immersion] very quickly when you can’t see the character in the reflection as well”[P29]).

5.1.5 AI partner’s movement

Some participants appreciated the character animation (“I was quite fascinated of how the character was animated, and how you were moving around, and the character was following you.”[P29]), but in general, there was more negative than positive comments.

One of the most common comments from the participants was that the partner’s movement felt unnatural to them (“at some point it felt like she was kind of fishing for attention. Like, hey, I’m coming to your view here again. See me.”[P19], “I thought that the character dances a little robotic, slightly less natural than I would prefer, but still I think it’s very comfortable…”[P24], “There was something slightly uncanny, because when the character is just there being idle, it was, okay, that’s what you kind of like got used to, and suddenly when the dance started, there was like a slight bit of uncanniness.”[P29]).

Many also complained that the movements the partner made were too difficult to follow (“I also felt intimidated myself because I had no dancing experience. And then it like, it also kind of felt like, okay, I’m dancing and I have this like really, really like skilled dancer in front of me and I’m trying to copy their movements. And then I’m like, I’m garbage.”[P1], “AI could be a bit more beginner friendly and then it could be more advanced as once you are accustomed to it.”[P4]). This might have been partly caused by the movement having very little repetition and moving from one move to another fairly quickly (“I could follow the movement, but sometimes it’s a bit too quick, so I couldn’t follow the movement that smoothly.”[P28]). The AI partner also sometimes moved too close to the participant, which made them uncomfortable (“ I tried to get away from the avatar and looked in a different direction and then turned, and the other kind of just came up on me. That was kind of sudden and frightening”[P30], “…when I come back, and she’s disappeared because she kind of came to me and we are a little bit overlapped”[P8]).

5.1.6 Connection with the AI partner

Even with the limited interactions, some participants felt a connection to the partner (“(”Definitely say so, like to some extent. Definitely.)[P1], “the eye contact was some at some moment very intensive. And some at some point, like I did see that I wasn’t interacting with the avatar, but it felt surprisingly real.”, “when I move around the place, I think the character is also trying to face the front side of me to some degree. So yeah, I feel the connection.”[P28]).

Some participants also pointed out the lack of interaction as a specific positive since they did not feel judged (“I think I might feel more secure dancing with the virtual person because I know she won’t judge me.”[P18], “It felt more in the sense that I was in a club and, you know, there were random people dancing.”[P5], “…I tend to feel somehow judged because I know that I’m not moving properly…Whereas if the character is fictional, I know that I can move…You care about like having fun, enjoying the moment because you don’t have the social pressure of being good at it.”[P26], “…there wasn’t like, for me, not like a personal or emotional connection to the avatar, but I kind of liked that part about it.”[P35]). A participant also mentioned that feeling awkward with the partner made it feel more real:

“I think I, in this sense, I think the character seems pretty real to me because I kind of feel awkward dance with her like I’m dancing with a real person.[P18]”

On the negative side, most participants did not feel much connection with the partner (“ I didn’t feel the connection. Right. It was actually more negative, it’s not even neutral to me. Opposite of connection, because there was no feedback”[P13], “I felt quite a lot that we were doing our own things. The character was doing what she was doing. Maybe in the first run I was trying to connect with the character a bit more, but I didn’t find a connection between us.”[P19]). One participant said that they decided to ignore the partner altogether (“I stopped sort of paying attention to the avatar because I didn’t care about him anymore…”[P34]).

Many participants described the partner avatar to feel like a stranger (“So I would I would say it’s kind of between the stranger and a friend. Yeah, but more closer to stranger.”, “he just felt like a rando that I happened to be in the same party”). This is, of course, also true in the sense that they had not danced with the avatar before, as one participant pointed out:

“…I think if it’s an avatar that you know because this is a stranger right like it’s difficult to have that kind of emotions going through so yeah it felt a little weird you need time to know the character and have some emotional connection or something …you needed some time to build connections.[P31]”

Some participants explained that they felt more like audience members than active participants in dancing (“I was looking at the character dancing and appreciating the moves, and not as much me dancing with the character.”[P32], “…I was just kind of observing him do his little show”[P33]).

Some people thought that in the MR version, they were able to connect better with the partner (“the first one [MR] I would say I had more of a connection because it was in my space that I’m familiar with. So there, the character resonated more than in a completely virtual environment.”[P4], “felt I would engage more with the mixed reality …Because it doesn’t feel like a video. When it’s all virtual, it seems like not real. Right. I can sense it’s not real. So mixed reality is like we are both in each other’s world somehow.”[P13], “I was in one world and I knew how this world works. So I don’t have to kind of pay attention or try to see anything. I don’t have to explore the surroundings when I’m in the see-through version.”[P19]). In the same way, some thought that it is easier for them to connect with the partner in VR (“ Like if there was a relationship, I could say that in the second [VR] version, it was improved. So yeah, I would say that.”[P26], “the virtual one I was much more immersed, so I felt like I let myself go by the character that I was watching, while in the other version I didn’t focus as much on the character.”[P32], “I felt like I was dancing with a friend of a friend, so not completely stranger, but not a friend, and in the other world, in the other version [MR], it felt like I was dancing with a complete stranger”).

5.1.7 Interaction suggestions

A majority of the participants wanted to be able to interact more with the partner avatar. One of the most common suggestions was that the avatar should greet them before dancing (“…would be nice if the character could say something in the beginning or maybe wave a hand to me, just say okay, we’ll start now, maybe set the mood.”[P24], “…maybe try to like fist bump or shake your hand first or something like that…”[P29]). Another common wish was that the partner should encourage them dance more and give them positive feedback (“she could talk to me or something or give me some more instruction or encourage me to dance.”[P28], “…maybe like some words of encouragement or like some gesture.”[P34], “And it also kind of is maybe verbal, I don’t know, encouragement or something like that might be nice”[P35]). Regarding the dancing style and movements, a common suggestion was that the partner’s movements should match their intensity (“…maybe he would start to do a few things to get me to dance and then when I danced more, he would dance more”[P33]) or try to mirror their moves (“it would be nice to if it could mimic the moves. Yeah. That would make a huge difference.”[P36], “…maybe if she imitates your movements or something, somehow it will be cool as well.”[P25]).

Multiple participants said that they would like the avatar to speak (“Maybe some kind of verbal communication could be nice or some kind of maybe that could be one way of kind of, like I said previously, to kind of help kind of anticipate the kind of like the next.” [P35], “If she can speak, then it would be good”). However, one participant also mentioned how they do not want to talk with the avatar (“Not talking, just dancing.”[P17]).

Most participants suggested different non-verbal ways of interacting with their partner such as gesturing, pushing, or holding hands. Only one participant suggested that they would also want to control the partner with a menu (“I’m pretty comfortable in the gaming environment where you just click on menus and stuff. So if there is like a Like of exhaustive possibilities then”[P30]).

Many participants referred to the dance partner as a kind of trainer (“I’m a complete beginner, I would say more of a trainer, trainee relationship”[P4],“kind of like the coaching or like tutoring role…”[P8]). Many also explained how they tried to mimic the partner’s moves (“I’m not much of a dancer, honestly, but it was kind of nice to kind of, you could kind of try to do more or less what she was doing. And it felt, I guess it felt nice.”[P35],“I was trying to mimic the first time his moves”[P33]).

5.1.8 AI partner’s visuals

By a large margin, the most commented part of the partner’s visuals was the face, especially the eyes. Some participants thought that the eye contact was good (“…I feel like it’s cute because we will have eye contact with this really good one.”[P8], “It was interesting to see the eye connection the character got once in a while”[P19]), and one even said that they would like a bit more eye contact (“…a bit more eye contact, a bit more, you know…”[P5]). However, most participants felt that the eye contact and movement were unrealistic (“ her eyes wasn’t really so realistic. I could see it’s a completely virtual character. I don’t remember. Did she have any eye blinked or not? I don’t remember it really.”[P14], “it was a bit like creepy at times because you were like staring at your face. Yeah. Okay. Constant eye contact while you were doing all stuff. Yeah. So it felt a bit weird.”[P23]). Many also pointed out that the avatar did not show emotions (“ it’s kind of like not even like displaying emotions like joy or excitement, but just like stone cold expression, just dancing, dancing, dancing…”[P1], “Like there was no face expression, I guess. It was yes, it should be some like smiling”[P6]).

5.2 Presence data

A paired-samples two-tailed -test for the IPQ questionnaire indicated a statistically significant difference showing that the VR version elicited slightly higher presence (, Cohen’s , MR average of all subscales = 2.892, MR , VR average of all subscales = 3.565, and VR ).

To further investigate the difference, we conducted Bonferroni-corrected pairwise t-tests between the VR and MR conditions for each of the three IPQ subscales. The results are illustrated in Figure 5. The VR condition elicited statistically significantly higher involvement [, Cohen’s , AR (, ), and VR (, )] and spatial presence [(, Cohen’s , AR (, ), and VR (, )]. There was no statistically significant difference in experienced realism (, Cohen’s , AR(, ), and VR(, ).

FIGURE 5

The interview data provide further nuance to the results. The concepts of presence and immersion are intertwined (Berkman and Akan, 2024), and many commented that VR felt more immersive (“I would definitely say that the second one [VR] felt much more immersive. The avatar itself felt like it like behaved pretty identically…”[P1], “…I mean, obviously in VR, we get a bit more immersed in the virtual environment…”[P2]), while MR visuals and the ability to see oneself helped with realism and embodiment (“…because it was like here, it felt like I was like real, everything was real…”[P25], “…it felt way more real. It felt more, way more connected to my body.”[P34]).

It should be noted that four participants, by accident, only partially completed the IPQ, and the data from these participants were not used in the IPQ analyses. To achieve the total sample size necessary for the study, we had three more participants (two women and one man aged 26–35) complete the study and fill out the questionnaire while omitting the semi-structured interview, resulting in a final sample size of 35 for the IPQ data. Due to the counterbalanced within-subjects design of our study, the extra participants should not bias the results.

6 Discussion

In the following subsections, we review and contextualize our findings on both the VR–MR differences and the user experience of our AI dance partner.

6.1 Comparing virtual and mixed reality

Agreeing with previous research in contexts other than dance (Allcoat et al., 2021; Gronowski et al., 2024), our results indicate that users feel more present while dancing in virtual reality. However, there appears to be a split between the users’ overall preferences for these two technologies. VR seems to be somewhat more immersive than MR and helps the user move more confidently. MR, on the other hand, appears to feel safer and more realistic to the user and keeps the user more aware of their own body. This might suggest that MR is more suitable for use cases where a user needs to learn new dance moves. One way to further probe the differences could be to quantitatively measure how much and how expressively participants dance in VR and MR.

We further observed that MR is more dependent on the physical space, and the experience may be affected by factors such as real mirrors, which can make users overly self-conscious. In our case, this occurred for some users due to the one-way mirror in the user study space. On the other hand, previous research has found real and virtual mirrors to be helpful in practicing dance and other movements (Hämäläinen, 2004; Zhou et al., 2023; Treffer et al., 2024).

VR appears to reduce public self-consciousness (Fenigstein et al., 1975) by decoupling users from their visible physical self. This aligns with research on the Proteus effect, showing that avatar-mediated interaction alters self-perception and reduces social inhibition (Yee and Bailenson, 2007), and with findings that VR dancers experience reduced self-consciousness through avatar anonymity (Piitulainen et al., 2022). However, this same decoupling diminishes body awareness (Döllinger et al., 2023), which was preserved in our MR condition—explaining why some participants preferred MR for its greater sense of embodiment and movement awareness.

VR can provide a more coherent visual experience, in which the environment and avatars’ appearances do not conflict with each other. This has also been noted in previous work (Collins et al., 2017; Alhakamy and Tuceryan, 2020) and can affect the users’ feeling of presence (DeMarbre et al., 2024). Steptoe et al. (2014) found that stylizing the visual input might help solve this problem. In dancing, physical interactions and occlusions are important aspects of the environment–character cohesion; these are easier to get right in VR, where an explicit model of the environment geometry is available. In MR, the model needs to be constructed using computer vision techniques, which are presently limited, at least with low-cost mass-market headsets.

There have been multiple studies showing that VR can help form and foster social connections (Agnew et al., 2022; Wei et al., 2023). In our case, there were no clear differences in the sense of connection with the AI dance partner between MR and VR, and the results varied considerably between users. Some participants felt a stronger connection when the AI partner appeared to be part of the physical world, while others felt more connection when they were able to fully immerse themselves in the partner’s virtual environment. This might have also been caused by limited interaction possibilities between the user and the avatar. When designing the study, we considered two minutes to be sufficient for participants to evaluate their experience, but aspects such as connection to the partner avatar might have been stronger with a longer dance session. In addition, the participants were mostly university students, which limits the generalizability of the study’s results.

The seemingly contradictory preferences expressed in the interviews suggest that responses to VR and MR are strongly shaped by individual differences rather than solely by technologies. Although our interview data indicate that factors such as prior dance experience, comfort with being observed, spatial awareness, and sensitivity to one’s physical surroundings influence the experience, the study was not designed to connect preferences to specific personal traits; presently, we are lacking more detailed per-participant data that would allow further analyses. For individual participants, there were likely carryover effects between the conditions, although we did not carry out detailed carryover analyses. Due to counterbalancing, carryover effects should not affect the quantitative results.

The difference between higher presence scores in VR and the nearly split qualitative preferences, with many participants describing MR as more “real,” is also a notable finding. Our results suggest that participants’ notion of “realism” does not to fully align with the dimensions of presence evaluated using the IPQ. Although VR increased involvement and spatial presence, qualitative accounts indicate that MR’s perceived realism was often grounded in continuity with the physical world, bodily awareness, and perceived safety rather than immersive absorption. This implies that MR’s “connection to reality” may reflect qualities that are not well captured by traditional presence scales, highlighting a limitation of existing instruments in mixed-reality contexts.

6.2 The AI dance partner

Our dance partner’s interaction design received considerable criticism. Many participants did not move much while dancing, which failed to trigger interactive repositioning of the partner and therefore resulted in little to no sense of connection with the AI partner. Multiple participants commented that the partner should respond to their movements and adapt to their dancing intensity. Future work should focus on AI partners that adjust to the user’s skill and intensity levels or provide users with ways to select the desired behavior. On the other hand, some people enjoyed that the avatar focused on its own dancing and did not “judge” the participant’s performance.

Many participants considered the partner avatar a trainer whose movements they were supposed to follow, although this was not instructed. For this, it would be important for the trainer to encourage and provide feedback on their performance.

The face of the AI partner felt emotionless to participants. Many also found that the partner’s gaze to be unnatural. Aligning with previous research (Uono and Hietanen, 2015; Broda and de Haas, 2024), the appropriate amount of eye contact appears to be highly subjective. How eye contact is implemented can also affect user behavior (Mello et al., 2024). To improve the facial and gaze interaction of the AI partner, one would need to evaluate different eye behaviors and add facial expressions that respond to the user’s behavior. This, however, would require hardware capable of tracking the user’s face and eyes. Another solution could be to move away from realistic graphics and provide more abstract avatars with less facial and eye detail. The best possible solution, of course, would be to provide both options and allow users to choose as our participants exhibited varying preferences between stylized and realistic graphics.

It is important to note that several of the critiques raised by participants are specific to the current implementation of the AI partner, and the partner’s behavior might have been the biggest factor affecting participants’ experience. However, these design limitations were held constant across both VR and MR conditions, suggesting that the observed differences between the two media are unlikely to be solely attributable to the partner’s behavior. Moreover, participants’ contrasting responses often related explicitly to media-dependent characteristics, such as environmental awareness, perceived safety, and embodiment, rather than to the partner itself. Nevertheless, future work should validate these findings using alternative partner designs to further differentiate VR- and MR-specific effects from agent-specific design choices.

The largely critical responses to the AI partner also warrant reflection on its role as a research instrument. Although the partner was intended as a neutral stimulus held constant across conditions, its limited interactivity and expressive capacity clearly shaped participants’ engagement and expectations.

Many dancers felt that the partner’s movements were too advanced and that the avatar transitioned between them too quickly. One reason for this might have been that the dance data we used were captured from professional dancers. In an early prototype, we used dance data captured from one of the researchers who was less skilled (supplemental video at 45s)—in hindsight, these data might have worked better with novice users. In our GitHub repository (https://github.com/CarouselDancing/AIDancePartner), we provide both movement datasets and AI partner versions.

6.3 Design implications

To synthesize our findings into an actionable format, we propose the following design suggestions for XR dance experiences and AI dance partners. These aim to crystallize our current understanding of the domain, grounded in the study as a whole and drawing on the quantitative presence data, thematic analysis of interviews, and our observations during the user study sessions and the iterative prototyping and design process. We note that individual suggestions are typically informed by a subset of interview responses rather than by statistically generalizable quantitative results; they should, therefore, be understood as empirically grounded starting points for design rather than definitive prescriptions.

7 Conclusion

In this study, we explored the differences in user experience between virtual reality and mixed reality in the context of dancing with an AI partner. Through a mixed-methods user study involving qualitative semi-structured interviews and quantitative presence data, we identified different preferences and experiences among users. Although VR was found to elicit a higher sense of presence on average, preferences for VR or MR were nearly evenly split, with individual differences influenced by factors such as confidence, perceived safety, and the desire for immersion. We found that the main causes for these feelings varied among participants, sometimes in contradictory ways.

Additionally, we demonstrate that Motion Matching can be a viable generative AI animation approach for creating real-time AI dance partners on limited consumer XR hardware especially as the algorithm’s search space for suitable next movement states becomes considerably smaller when movements are synchronized to the beat of the music. As demonstrated in the supplemental video, the adaptive dancing movements synthesized via Motion Matching are fairly smooth and plausible. However, only dancing is not enough, and our user study indicates that various aspects of our dance partner need improvement. Perhaps most notably, a larger database of dancing movements should be collected so that the partner can adapt to the user’s dance intensity and skill; an AI partner that is too intense or skilled compared to the user can appear intimidating.

The insights gained from this study emphasize the importance of tailoring XR experiences to accommodate diverse user needs and preferences. Designers should consider how VR and MR can complement different scenarios, particularly in activities involving full-body movement such as dance. We hope that our AI dance partner’s open-source implementation and open dance motion-capture data can enable further innovation and research in this domain.

References

• 1

  AbootorabiM. M.GhahroodiO.ZahraeiP. S.BehzadaslH.MirrokniA.SalimipanahM.et al (2025). Generative ai for character animation: a comprehensive survey of techniques, applications, and future directions. arXiv Preprint arXiv:2504, 19056.

  • Google Scholar
  • View reference in article

• 2

  AgnewC. R.CarterJ. J.ImamiL. (2022). Forming meaningful connections between strangers in virtual reality: comparing affiliative outcomes by interaction modality.

  • Google Scholar
  • View reference in article

• 3

  AlhakamyA.TuceryanM. (2020). Real-time illumination and visual coherence for photorealistic augmented/mixed reality. ACM Comput. Surv. (CSUR)53, 1–34. 10.1145/3386496

  • CrossRef
  • Google Scholar
  • View reference in article

• 4

  AllcoatD.HatchardT.AzmatF.StansfieldK.WatsonD.von MühlenenA. (2021). Education in the digital age: learning experience in virtual and mixed realities. J. Educ. Comput. Res.59, 795–816. 10.1177/0735633120985120

  • CrossRef
  • Google Scholar
  • View reference in article

• 5

  BanerjeeN. T.BaughmanA. J.LinS.-Y.WitteZ. A.KlausD. M.AndersonA. P. (2021). Side-by-side comparison of human perception and performance using augmented, hybrid, and virtual reality. IEEE Trans. Vis. Comput. Graph.28, 4787–4796. 10.1109/TVCG.2021.3105606

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 6

  BatrasD.GuezJ.JégoJ.-F.TramusM.-H. (2016). “A virtual reality agent-based platform for improvisation between real and virtual actors using gestures,” in Proceedings of the 2016 virtual reality international conference, 1–4.

  • Google Scholar
  • View reference in article

• 7

  BerkmanM. I.AkanE. (2024). “Presence and immersion in virtual reality,” in Encyclopedia of computer graphics and games (Springer), 1461–1470.

  • Google Scholar
  • View reference in article

• 8

  BernalG.MaesP. (2017). “Emotional beasts: visually expressing emotions through avatars in vr,” in Proceedings of the 2017 CHI conference extended abstracts on human factors in computing systems, 2395–2402.

  • Google Scholar
  • View reference in article

• 9

  BisigD.PalacioP.RomeroM. (2016). “Piano&dancer-interaction between a dancer and an acousticinstrument,” in 19th generative art conference GA2016.

  • Google Scholar
  • View reference in article

• 10

  BorgwardtM.BouekeJ.SanabriaM. F.BonfertM.PorzelR. (2023). Vrisbee: how hand visibility impacts throwing accuracy and experience in virtual reality. Ext. Abstr. 2023 CHI Conf. Hum. Factors Comput. Syst.314, 1–7.

  • Google Scholar
  • View reference in article

• 11

  BraunV.ClarkeV. (2021). One size fits all? What counts as quality practice in (reflexive) thematic analysis?Qual. Res. Psychol.18, 328–352. 10.1080/14780887.2020.1769238

  • CrossRef
  • Google Scholar
  • View reference in article

• 12

  BrodaM. D.de HaasB. (2024). Individual differences in human gaze behavior generalize from faces to objects. Proc. Natl. Acad. Sci. U. S. A.121, e2322149121. 10.1073/pnas.2322149121

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 13

  CavalcantiA. B. (2020). The effects of social exergames on mental wellbeing and how game designers can enhance them. Queensland: University of the Sunshine Coast. Ph.D. thesis.

  • Google Scholar
  • View reference in article

• 14

  ChenK.TanZ.LeiJ.ZhangS.-H.GuoY.-C.ZhangW.et al (2021). Choreomaster: choreography-oriented music-driven dance synthesis. ACM Trans. Graph. (TOG)40, 1–13. 10.1145/3450626.3459932

  • CrossRef
  • Google Scholar
  • View reference in article

• 15

  ClavetS. (2016). Motion matching and the road to next-gen animation (game developer conference 2016 presentation). Available online at: https://www.gdcvault.com/play/1023280/Motion-Matching-and-The-Road (Accessed January 16, 2025).

  • Google Scholar
  • View reference in article

• 16

  CmentowskiS.KrekhovA.KruegerJ. (2019). Outstanding: a perspective-switching technique for covering large distances in vr games. Ext. Abstr. 2019 CHI Conf. Hum. Factors Comput. Syst., 1–6.

  • Google Scholar
  • View reference in article

• 17

  CollinsJ.RegenbrechtH.LanglotzT. (2017). Visual coherence in mixed reality: a systematic enquiry. Presence26, 16–41. 10.1162/pres_a_00284

  • CrossRef
  • Google Scholar
  • View reference in article

• 18

  DeMarbreE.HendersonJ.TeatherR. J. (2024). “Investigating presence across rendering style and ratio of virtual to real content in mixed reality,” in Proceedings of the 2024 ACM symposium on spatial user interaction, 1–8.

  • Google Scholar
  • View reference in article

• 19

  DöllingerN.WolfE.BotschM.LatoschikM. E.WienrichC. (2023). “Are embodied avatars harmful to our self-experience? The impact of virtual embodiment on body awareness,” in Proceedings of the 2023 CHI conference on human factors in computing systems, 1–14.

  • Google Scholar
  • View reference in article

• 20

  EvinI.PesolaT.KaosM. D.TakalaT. M.HämäläinenP. (2020). “3pp-r: enabling natural movement in 3rd person virtual reality,” in Proceedings of the annual symposium on computer-human interaction in play, 438–449.

  • Google Scholar
  • View reference in article

• 21

  FaulF.ErdfelderE.LangA.-G.BuchnerA. (2007). G* power 3: a flexible statistical power analysis program for the social, behavioral, and biomedical sciences. Behav. Research Methods39, 175–191. 10.3758/bf03193146

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 22

  Fdili AlaouiS. (2019). “Making an interactive dance piece: tensions in integrating technology in art,” in Proceedings of the 2019 on designing interactive systems conference, 1195–1208.

  • Google Scholar
  • View reference in article

• 23

  FenigsteinA.ScheierM. F.BussA. H. (1975). Public and private self-consciousness: assessment and theory. J. Consulting Clinical Psychology43, 522–527. 10.1037/h0076760

  • CrossRef
  • Google Scholar
  • View reference in article

• 24

  GamesB. (2019). Beat saber. Available online at: https://store.steampowered.com/app/620980/Beat_Saber/ (Accessed April 30, 2026).

  • Google Scholar
  • View reference in article

• 25

  GhoshA.CheemaN.OguzC.TheobaltC.SlusallekP. (2021). “Text-based motion synthesis with a hierarchical two-stream rnn,” in ACM SIGGRAPH 2021 posters, 1–2.

  • Google Scholar
  • View reference in article

• 26

  GhoshA.DabralR.GolyanikV.TheobaltC.SlusallekP. (2025). “Remos: 3d motion-conditioned reaction synthesis for two-person interactions,” in European conference on computer vision (Springer), 418–437.

  • Google Scholar
  • View reference in article

• 27

  GriffinN. N.FolmerE. (2019). “Out-of-body locomotion: vectionless navigation with a continuous avatar representation,” in Proceedings of the 25th ACM symposium on virtual reality software and technology, 1–8.

  • Google Scholar
  • View reference in article

• 28

  GronowskiA.ArnessD. C.NgJ.QuZ.LauC. W.CatchpooleD.et al (2024). The impact of virtual and augmented reality on presence, user experience and performance of information visualisation. Virtual Real.28, 133. 10.1007/s10055-024-01032-w

  • CrossRef
  • Google Scholar
  • View reference in article

• 29

  GuoY.LuS.ShenM.HuangW.YiX.ZhangJ. (2024a). Differences in heritage tourism experience between vr and ar: a comparative experimental study based on presence and authenticity. ACM J. Comput. Cult. Herit.17, 1–21. 10.1145/3648001

  • CrossRef
  • Google Scholar
  • View reference in article

• 30

  GuoZ.WangH.DengH.XuW.BaghaeiN.LoC.-H.et al (2024b). Breaking the isolation: exploring the impact of passthrough in shared spaces on player performance and experience in vr exergames. IEEE Trans. Vis. Comput. Graph.30, 2580–2590. 10.1109/TVCG.2024.3372114

  • Pubmed Abstract
  • CrossRef
  • Google Scholar
  • View reference in article

• 31

  GyawaliD. (2023). Mixed reality: the interface of the future. arXiv Preprint arXiv:2309.00819.

  • Google Scholar
  • View reference in article

• 32

  HagedornL. J.De RooijA.AlimardaniM. (2023). Virtual reality and creativity: how do immersive environments stimulate the brain during creative idea generation?Ext. Abstr. 2023 CHI Conf. Hum. Factors Comput. Syst., 1–7.

  • Google Scholar
  • View reference in article

• 33

  HämäläinenP. (2004). “Interactive video mirrors for sports training,” in Proceedings of the third nordic conference on human-computer interaction, 199–202.

  • Google Scholar
  • View reference in article

• 34

  HaydenS. (2025). Meta shares the 50 best-selling quest games of all time. Available online at: https://www.roadtovr.com/meta-quest-50-best-selling-games-2025/(Accessed April 24, 2025).

  • Google Scholar
  • View reference in article

• 35

  HoldenD.KanounO.PerepichkaM.PopaT. (2020). Learned motion matching. ACM Trans. Graph. (TOG)39, 53. 10.1145/3386569.3392440

  • CrossRef
  • Google Scholar
  • View reference in article

• 36

  HoysniemiJ. (2006). International survey on the dance dance revolution game. Comput. Entertain. (CIE)4, 8–es. 10.1145/1129006.1129019

  • CrossRef
  • Google Scholar
  • View reference in article

• 37

  HuangT.LiY.LiangH.-N. (2023). “Avatar type, self-congruence, and presence in virtual reality,” in Proceedings of the eleventh international symposium of Chinese CHI, 61–72.

  • Google Scholar
  • View reference in article

• 38

  HurterC.RogowitzB.TruongG.AndryT.RomatH.GardyL.et al (2024). Memory recall for data visualizations in mixed reality, virtual reality, 3d and 2d. IEEE Trans. Vis. Comput. Graph.

  • Google Scholar
  • View reference in article

• 39

  Ilkka KuuselaV. R. (2021). Motion matching and the road to next-gen animation. Proc. Gdc. Available online at: https://www.gdcvault.com/play/1027146/Animation-Summit-Take-CONTROL-of (Accessed January 18, 2025).

  • Google Scholar
  • View reference in article

• 40

  JavierM.WheelerS.NiQ.DrivasM.WongK. G.FitzsimmonsS.et al (2024). “The alt-self: investigating the inclusivity of self-avatar representations in social vr,” in Companion proceedings of the 2024 annual symposium on computer-human interaction in play, 131–138.

  • Google Scholar
  • View reference in article

• 41

  JicolC.ClarkeC.TorE.DakinR. M.LancasterT. C.ChangS. T.et al (2023). Realism and field of view affect presence in vr but not the way you think. Proc. 2023 CHI Conf. Hum. Factors Comput. Syst., 1–17.

  • Google Scholar
  • View reference in article

• 42

  KangJ.CaoN. T. B.WangY.ZhaoM.YouK. H. (2023). A study on avatar self-expression in vr environment. J. Internet Comput. and Serv.24.

  • Google Scholar
  • View reference in article

• 43

  KaraosmanogluS.CmentowskiS.NackeL. E.SteinickeF. (2024). “Born to run, programmed to play: mapping the extended reality exergames landscape,” in Proceedings of the 2024 CHI conference on human factors in computing systems (New York, NY, USA: Association for Computing Machinery). 10.1145/3613904.3642124

  • CrossRef
  • Google Scholar
  • View reference in article

• 44

  KirakosianS.DaskalogrigorakisG.MaravelakisE.ManiaK. (2021). “Near-contact person-to-3d character dance training: comparing ar and vr for interactive entertainment,” in 2021 IEEE conference on games (CoG) (IEEE), 1–5.

  • Google Scholar
  • View reference in article

• 45

  KirjonenM.SlezasA.DuttaS.LaattalaM.Hm̈l̈ïnenP. (2025). “Souls-vr: dodge-rolling in virtual reality,” in Proceedings of the 20th international conference on the foundations of digital games (New York, NY, USA: Association for Computing Machinery). FDG ’25.

  • Google Scholar
  • View reference in article

• 46

  KirshD. (2010). Thinking with external representations. AI and Society25, 441–454. 10.1007/s00146-010-0272-8

  • CrossRef
  • Google Scholar
  • View reference in article

• 47

  KontioR.LaattalaM.WelschR.HämäläinenP. (2023). “I feel my abs”: exploring non-standing vr locomotion. Proc. ACM Human-Computer Interact.7, 1282–1307. 10.1145/3611069

  • CrossRef
  • Google Scholar
  • View reference in article

• 48

  KovarL.GleicherM.PighinF. (2023). Motion graphs. Seminal Graph. Pap. Push. Boundaries2, 723–732.

  • Google Scholar
  • View reference in article

• 49

  KrekhovA.CmentowskiS.EmmerichK.MasuchM.KrügerJ. (2018). “Gullivr: a walking-oriented technique for navigation in virtual reality games based on virtual body resizing,” in Proceedings of the 2018 annual symposium on computer-human interaction in play, 243–256.

  • Google Scholar
  • View reference in article

• 50

  KrekhovA.CmentowskiS.EmmerichK.KrügerJ. (2019). “Beyond human: animals as an escape from stereotype avatars in virtual reality games,” in Proceedings of the annual symposium on computer-human interaction in play, 439–451.

  • Google Scholar
  • View reference in article

• 51

  KruseL.MostajeranF.SteinickeF. (2023). “The influence of virtual agent visibility in virtual reality cognitive training,” in Proceedings of the 2023 ACM symposium on spatial user interaction, 1–9.

  • Google Scholar
  • View reference in article

• 52

  KurzwegM.ReinhardtJ.NabokW.WolfK. (2021). “Using body language of avatars in vr meetings as communication status cue,” in Proceedings of Mensch und Computer 2021, 366–377.

  • Google Scholar
  • View reference in article

• 53

  KwiatkowskiA.AlvaradoE.KalogeitonV.LiuC. K.PettréJ.van de PanneM.et al (2022). A survey on reinforcement learning methods in character animation. Comput. Graphics Forum41, 613–639. 10.1111/cgf.14504

  • CrossRef
  • Google Scholar

• 54

  LaattalaM.PiitulainenR.AdyN. M.TamarizM.HämäläinenP. (2024). “Wave: anticipatory movement visualization for vr dancing,” in Proceedings of the CHI conference on human factors in computing systems, 1–9.

  • Google Scholar

• 55

  LiZ.AnnettM.HinckleyK.SinghK.WigdorD. (2019). “Holodoc: enabling mixed reality workspaces that harness physical and digital content,” in Proceedings of the 2019 CHI conference on human factors in computing systems, 1–14.

  • Google Scholar

• 56

  LiB.ZhaoY.ZhelunS.ShengL. (2022). Danceformer: music conditioned 3d dance generation with parametric motion transformer. Proc. AAAI Conf. Artif. Intell.36, 1272–1279. 10.1609/aaai.v36i2.20014

  • CrossRef
  • Google Scholar

• 57

  LunguA. J.SwinkelsW.ClaesenL.TuP.EggerJ.ChenX. (2021). A review on the applications of virtual reality, augmented reality and mixed reality in surgical simulation: an extension to different kinds of surgery. Expert Review Medical Devices18, 47–62. 10.1080/17434440.2021.1860750

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 58

  MachM.ZhuravlovM. (2021). Available online at: https://s3.amazonaws.com/nd.images/research/tlou2/motion_matching_in_tlou2_nd_final.pptx (Accessed July 14, 2024).

  • Google Scholar

• 59

  Magic Leap (2010). Magic leap. Available online at: https://www.magicleap.com/ (Accessed April 30, 2026).

  • Google Scholar

• 60

  MelloB.WelschR.VerbokkemM. C.KnierimP.DechantM. J. (2024). “Navigating the virtual gaze: social anxiety’s role in vr proxemics,” in Proceedings of the CHI conference on human factors in computing systems, 1–15.

  • Google Scholar

• 61

  Meta (2020). Meta quest 2. Available online at: https://www.meta.com/fi/quest/products/quest-2/ (Accessed April 30, 2026).

  • Google Scholar

• 62

  Meta (2023). Meta quest 3. Available online at: https://www.meta.com/fi/quest/products/quest-3/ (Accessed April 30, 2026).

  • Google Scholar

• 63

  MilgramP.KishinoF. (1994). A taxonomy of mixed reality visual displays. IEICE Trans. Inf. Syst.77, 1321–1329.

  • Google Scholar

• 64

  MoriM.MacDormanK. F.KagekiN. (2012). The uncanny valley [from the field]. IEEE Robotics and Automation Magazine19, 98–100. 10.1109/mra.2012.2192811

  • CrossRef
  • Google Scholar

• 65

  MorrisM. E.RosnerD. K.NuriusP. S.DolevH. M. (2023). ““I don’t want to hide behind an avatar”: self-representation in social vr among women in midlife,” in Proceedings of the 2023 ACM designing interactive systems conference, 537–546.

  • Google Scholar

• 66

  MourotL.HoyetL.Le ClercF.SchnitzlerF.HellierP. (2022). A survey on deep learning for skeleton-based human animation. Comput. Graph. Forum41, 122–157. 10.1111/cgf.14426

  • CrossRef
  • Google Scholar

• 67

  OhJ.KimG. J. (2020). “Effects of gestural exaggeration to user experience in virtual reality,” in Proceedings of the 25th international conference on 3D web technology, 1–4.

  • Google Scholar

• 68

  PiitulainenR.HämäläinenP.MeklerE. D. (2022). “Vibing together: dance experiences in social virtual reality,” in Proceedings of the 2022 CHI conference on human factors in computing systems, 1–18.

  • Google Scholar

• 69

  PratticòF. G.ChecoI.ViscontiA.SimeoneA.LambertiF. (2023). “Designing hand-held controller-based handshake interaction in social vr and metaverse,” in Proceedings of the 16th ACM SIGGRAPH conference on motion, interaction and games, 1–6.

  • Google Scholar

• 70

  RadfordA.KimJ. W.XuT.BrockmanG.McLeaveyC.SutskeverI. (2023). “Robust speech recognition via large-scale weak supervision,” in International conference on machine learning, Honolulu, Hawaii, July 23–29, 2023, 28492–28518.

  • Google Scholar

• 71

  SenecalS.NijdamN. A.AristidouA.Magnenat-ThalmannN. (2020). Salsa dance learning evaluation and motion analysis in gamified virtual reality environment. Multimedia Tools Appl.79, 24621–24643. 10.1007/s11042-020-09192-y

  • CrossRef
  • Google Scholar

• 72

  SteptoeW.JulierS.SteedA. (2014). “Presence and discernability in conventional and non-photorealistic immersive augmented reality,” in 2014 IEEE international symposium on mixed and augmented reality (ISMAR) (IEEE), 213–218.

  • Google Scholar

• 73

  TangJ.BaileyJ. O. (2024). “Exploring the effects of fantasy level of avatars on user perception and behavior,” in Adjunct proceedings of the 37th annual ACM symposium on user interface software and technology, 1–3.

  • Google Scholar

• 74

  TrefferA.ClarkA.LukoschS. (2024). “Teaching dance with mixed reality mirrors,” in 2024 IEEE international symposium on mixed and augmented reality (ISMAR) (IEEE), 971–980.

  • Google Scholar

• 75

  UonoS.HietanenJ. K. (2015). Eye contact perception in the west and east: a cross-cultural study. Plos One10, e0118094. 10.1371/journal.pone.0118094

  • Pubmed Abstract
  • CrossRef
  • Google Scholar

• 76

  van den HeuvelJ.CunliffeJ.ParkerE. (2021). “The right foot in the wrong place: half-life character locomotioncharacter locomotion in half-life: alyx,” in ACM SIGGRAPH 2021 talks, 1–2.

  • Google Scholar

• 77

  VrO. (2012). Oculus rift. Available online at: https://oculusvr.com/ (Accessed April 30, 2026).

  • Google Scholar

• 78

  VR Exercise Tracker (2018). Beat saber. Available online at: https://vrhealth.institute/portfolio/beat-saber/ (Accessed April 30, 2026).

  • Google Scholar

• 79

  WeiX.GuY.KuangE.WangX.CaoB.JinX.et al (2023). “Bridging the generational gap: exploring how virtual reality supports remote communication between grandparents and grandchildren,” in Proceedings of the 2023 CHI conference on human factors in computing systems, 1–15.

  • Google Scholar

• 80

  WobbrockJ. O.KientzJ. A. (2016). Research contributions in human-computer interaction. Interactions23, 38–44. 10.1145/2907069

  • CrossRef
  • Google Scholar

• 81

  YeeN.BailensonJ. (2007). The proteus effect: the effect of transformed self-representation on behavior. Hum. Communication Research33, 271–290. 10.1111/j.1468-2958.2007.00299.x

  • CrossRef
  • Google Scholar

• 82

  ZadziukK. (2016). Motion matching - ’dance card’ breakdown. Available online at: https://youtu.be/_Bd2T7uP9VA?si=XRsHlfIUDK3fySaP (Accessed July 14, 2024).

  • Google Scholar

• 83

  ZhouQ.ChuaC. C.KnibbeJ.GoncalvesJ.VellosoE. (2021). “Dance and choreography in hci: a two-decade retrospective,” in Proceedings of the 2021 CHI conference on human factors in computing systems, 1–14.

  • Google Scholar

• 84

  ZhouQ.GrebelL.IrlittiA.MinaaiJ. A.GoncalvesJ.VellosoE. (2023). “Here and now: creating improvisational dance movements with a mixed reality mirror,” in Proceedings of the 2023 CHI conference on human factors in computing systems, 1–16.

  • Google Scholar

• 85

  ZhouX.BatmazA. U.WilliamsA. S.SchreiberD.OrtegaF. R. (2024). I did not notice: a comparison of immersive analytics with augmented and virtual reality. Ext. Abstr. CHI Conf. Hum. Factors Comput. Syst., 1–7.

  • Google Scholar