Comparative effects of cooperative and competitive social models on motivation and physiological engagement in VR exergaming

1 Introduction

1.1 Global context and motivation

Physical inactivity remains a pressing global health challenge, contributing to chronic disease and reduced quality of life (World Health Organization, 2022). Although digital exergames can increase short-term physical activity and energy expenditure (Althoff et al., 2016; Bailey and McInnis, 2011; Graves et al., 2008; Peng et al., 2011; Gao et al., 2015), sustaining engagement over time remains difficult, particularly when novelty effects diminish and ongoing adherence is not supported by social, motivational, or contextual reinforcement (Collado-Mateo et al., 2021; Harris, 2019; Kaos et al., 2019). Recent work further indicates that cognitive–motor exergames can support functional and cognitive rehabilitation, particularly among older adults (Altorfer et al., 2021), highlighting the broader health potential of interactive exercise technologies.

Virtual-reality (VR)–based social exergaming, which combines immersive exercise with real-time social interaction, has been proposed as a means to address this motivational sustainability problem by transforming solitary workouts into socially meaningful experiences (Hibbs et al., 2024; IJsselsteijn et al., 2006; Neumann et al., 2018; Tarr et al., 2018). However, social interaction in exergaming can take different forms—such as cooperation or competition—which are theorized to activate distinct motivational pathways (e.g., relatedness versus competence). Despite growing interest in immersive VR applications in sport and exercise (Neumann et al., 2018), systematic comparisons of these social mechanisms remain limited. Without directly examining how cooperative and competitive dynamics differentially influence motivation, physiological engagement, and performance, it remains unclear how to design socially adaptive VR exercise systems that encourage repeated participation.

Unlike traditional fitness applications, VR exergaming can elicit presence, embodiment, and interpersonal connection—factors that may support motivation and future adherence when combined with appropriate design and evaluation (IJsselsteijn et al., 2006; Tarr et al., 2018; Oh et al., 2018). Recent work further suggests that immersive exergames can foster both enjoyment and physiological engagement, potentially reducing perceived exertion even under higher workload (Barathi et al., 2024).

1.2 Research gap

Despite growing evidence that social interaction may enhance exergaming outcomes (Kaos et al., 2019; Staiano et al., 2012), existing research often treats “social interaction” as a broadly beneficial design feature rather than distinguishing how different social structures produce different motivational, physiological, and experiential effects. This is a critical limitation because cooperative and competitive dynamics are not interchangeable. Cooperative exergaming has been associated with social support, self-efficacy, intrinsic motivation, and continued play, whereas competitive exergaming can increase exercise intensity but may also depend on individual competitiveness and may produce less positive experiences for some users (Staiano et al., 2012; Marker et al., 2015; Song et al., 2013). Prior work has also shown that cooperative and competitive exergame contexts can differentially shape intrinsic motivation and psychological attractiveness, suggesting that the social form of play is not merely an implementation detail but a central motivational mechanism (Marker et al., 2015; Song et al., 2013; Shaw et al., 2016).

Therefore, the need for systematic comparison goes beyond the fact that few studies have directly contrasted these modes. Without examining cooperation and competition within the same immersive VR exercise task, participant sample, and measurement protocol, it remains difficult to determine whether observed benefits are driven by general co-presence, relatedness-oriented cooperation, competence-oriented competition, or task-specific challenge effects. This distinction is especially important in VR cycling, where greater physical output may not necessarily correspond to greater enjoyment, lower perceived exertion, or stronger motivational value beyond a single session.

Although prior studies have examined social facilitation, competitive elements, and avatar-based feedback in exergaming (IJsselsteijn et al., 2006; Shaw et al., 2016; Park and Catrambone, 2007), many have focused on single-mode interventions, non-immersive systems, or isolated outcome measures. Even studies that directly explore virtual competitors or trainers indicate that competition and cooperation can produce different effects on exercise performance, motivation, and enjoyment, but further work is needed to connect these effects to psychophysiological engagement and social presence within immersive VR environments (Barathi et al., 2024; Shaw et al., 2016; Barathi et al., 2018). As a result, the field still lacks clear evidence on when cooperative models should be used to support enjoyment and relatedness, when competitive models should be used to promote effort and competence, and how these mechanisms might be adaptively balanced in future socially intelligent VR exercise systems.

1.3 Research objectives and contributions

To address these gaps, we designed a controlled VR cycling experiment with three system configurations: a Solo self-paced baseline; a Cooperative-framed condition in which participants rode beside a replay-based self- or friend-modelled virtual pacing partner; and an adaptive AI-Competitive condition in which a computer-controlled virtual agent adjusted its pace according to the participant’s physiological state. The Cooperative-framed condition emphasized companionship and shared pacing but did not implement a strongly interdependent team outcome.

The study compares these three complete VR exercise implementations using power output, physiological activation, intrinsic motivation, perceived exertion, and qualitative experience. Its contribution is therefore not a factorial isolation of cooperation and competition as single psychological variables. Instead, it evaluates how replay-based cooperative framing and adaptive competitive pacing are associated with different short-term experiential and performance profiles, and it identifies design considerations for future studies that more tightly match pacing, feedback, and goal structure across conditions.

Terminology (social-exercise implementation). In this paper, a social-exercise implementation refers to the combined goal framing, virtual counterpart, pacing behaviour, and feedback used in a condition. The term Cooperative-framed denotes companionship and shared pacing with a replay-based self- or friend-modelled virtual partner; it does not imply a strongly interdependent team task or a shared performance outcome. The term AI-Competitive denotes a computer-controlled virtual agent whose pace was adaptively regulated using the participant’s heart-rate response. These labels describe bundled implementations rather than isolated causal manipulations of cooperation and competition.

2 Related work and theoretical framework

2.1 Self-determination theory (SDT) and social cognitive theory (SCT)

Self-Determination Theory (SDT) (Deci and Ryan, 1985; Ryan and Deci, 2017) proposes that high-quality motivation is supported by autonomy, or the experience of volition and choice; competence, or feeling effective and capable; and relatedness, or feeling socially connected and cared for. Social Cognitive Theory (SCT) (Bandura, 1986) complements this account by emphasizing observational learning, self-efficacy, and social reinforcement. In VR exercise, social cues may therefore influence perceived capability and connection, but the presence of a virtual counterpart does not by itself establish cooperation.

Within VR settings, these motivational theories intersect with principles from Social Facilitation Theory, which proposes that the perceived presence of others—whether represented through live partners or virtual avatars—can elevate arousal and performance effort, especially on relatively simple or well-learned tasks (Park and Catrambone, 2007; Guerin, 1993; Zajonc, 1965). This link is particularly important for exergaming because social presence does not only make the experience feel more socially meaningful; it can also change how users regulate physical effort. In cooperative environments, enhanced relatedness may increase commitment to a shared goal, create a sense of accountability to the partner, and reduce the subjective salience of fatigue by making exertion feel socially supported rather than individually burdensome. Competitive contexts, by contrast, primarily stimulate competence, challenge appraisal, social comparison, and performance striving, which may increase exertion through goal-directed pressure and the desire to outperform or keep up with an opponent (Marker et al., 2015; Song et al., 2013; Shaw et al., 2016; Teixeira et al., 2012). Building on these insights, socially adaptive exergames may dynamically modulate challenge and interpersonal cues to balance exertion and enjoyment (Kaos et al., 2019; Barathi et al., 2024).

Research in gamification further identifies feedback mechanisms, goal structures, and social comparison features as design elements that support the psychological needs articulated by SDT, thereby reinforcing motivation during interactive exercise (Barata et al., 2015; Domínguez et al., 2013; Koivisto and Hamari, 2019).

2.2 Social presence

Social presence refers to the perceived sense of sharing an interaction space with another entity, and it shapes how social cues influence motivation and performance in VR environments (Tarr et al., 2018; Oh et al., 2018; Blascovich et al., 2002). In immersive exercise, heightened social presence may contribute to enjoyment, perceived competence, and cycling intensity, but these relationships depend on the specific interaction design. Psychophysiological measures such as pupil responses and skin conductance can complement self-report measures of presence and arousal, although they should not be treated as direct measures of social presence on their own (Cacioppo et al., 2007; Boucsein et al., 2012; Barathi et al., 2020). Designs that enable cooperative exertion—such as group-based swimming games and mixed-reality co-play scenarios—have reported ways in which shared activity can support relatedness and engagement (Choi et al., 2016; Kalaitzidou et al., 2022; Lehtonen et al., 2019). Together, these findings suggest that carefully orchestrated avatar behaviors and interaction rhythms can strengthen social presence and may amplify motivational benefits derived from facilitation.

Friend-modelled opponents or partners introduce an additional relational layer to social presence: by linking the virtual avatar to a known peer, users may experience stronger identification, relatedness, accountability, and perceived challenge. Empirical evidence indicates that this friend association can enhance both performance and intrinsic motivation relative to self- or stranger-models operating at the same enhanced pace (Barathi et al., 2024), underscoring social presence as a possible mechanism for improving exertion and engagement in VR exercise. The present study leverages this mechanism in the Cooperative condition by allowing participants to cycle alongside a self-modelled or friend-modelled cooperative avatar, rather than a generic virtual partner. This design tests whether relationally meaningful co-presence can support exertion through relatedness, shared pacing, and perceived accountability, while also enabling comparison with the AI-Competitive condition, where exertion is expected to be driven more strongly by competence, challenge, and adaptive performance pressure.

2.3 Social facilitation in VR

Social facilitation theory (Zajonc, 1965) posits that the perceived presence of others can amplify effort, arousal, and performance, particularly on simple, familiar, or well-learned tasks. In immersive VR, this effect is especially relevant because social presence can be simulated through avatars, virtual agents, co-present users, or evaluative audiences, even when no physical partner is present. As a result, VR provides a controlled way to examine how the perceived presence, behaviour, and responsiveness of others shape user performance and experience across interactive tasks.

Prior VR research has shown that virtual others can influence attention, arousal, self-presentation, and task behaviour by making users feel observed, accompanied, or socially evaluated. In collaborative or co-present virtual environments, avatars may increase accountability and task engagement, while evaluative or competitive agents may intensify performance monitoring and goal-directed effort. These effects suggest that social facilitation in VR is not limited to physical exercise; rather, it reflects a broader mechanism through which perceived co-presence changes how users allocate attention, regulate effort, and respond to challenge.

In VR exergaming, these mechanisms become physically consequential. Social facilitation may manifest as players pedalling harder, sustaining higher power output, or persisting longer when they perceive co-actors, rivals, or an evaluative audience. Prior VR and exergaming research indicates that virtual humans, competitive agents, social play, and immersive co-presence can influence exercise performance, persistence, and social connection, although the strength and direction of effects vary by task and design (Hibbs et al., 2024; Tarr et al., 2018; Barathi et al., 2024; Shaw et al., 2016; Park and Catrambone, 2007).

Further evidence by Barathi et al. (Barathi et al., 2024) suggests that pre-existing positive relationships—such as exercising with a friend-modelled avatar—can strengthen facilitation effects by increasing accountability and perceived interpersonal connection. This perspective indicates that pairing users with partners of similar ability and existing social ties may improve facilitative outcomes while reducing the risk of disengagement; social-loafing mechanisms remain a design risk that should be tested directly (Karau and Williams, 1993; Latané et al., 1979).

Facilitation in immersive contexts is intertwined with affective arousal and attentional focus. Beyond performance outcomes alone, psychophysiological markers can help distinguish whether increased effort reflects socially driven engagement, evaluative pressure, or generic task intensity. In VR exergaming, physiological studies have used markers such as skin conductance, pupil responses, and gaze behaviour to index challenge and engagement during high-intensity bouts (Cacioppo et al., 2007; Boucsein et al., 2012; Barathi et al., 2020). Live physiological co-visualization can further heighten co-presence and joint focus (Kang et al., 2016).

2.4 Cooperation vs. competition

Cooperative and competitive designs may emphasize different SDT needs, but these effects depend on their operationalization. Strong cooperation typically includes a shared goal and some degree of task interdependence, whereas parallel pacing with a socially meaningful partner may primarily support relatedness and social presence. Competition more directly emphasizes comparison, competence, and performance striving and may increase intensity, although it can reduce enjoyment when challenge is poorly calibrated (Staiano et al., 2012; Marker et al., 2015; Song et al., 2013; Deci and Ryan, 1985; Ryan and Deci, 2017; Ryan and Deci, 2000). Accordingly, the present replay-based Cooperative-framed condition should be distinguished from a fully interdependent cooperative task.

Prior VR exergaming studies show that self-competition and feedforward-style replay models can increase physical output while supporting intrinsic motivation (Barathi et al., 2018; Guadagnoli and Lee, 2004). Michael and Lutteroth (Michael and Lutteroth, 2020) further demonstrated that racing against past and projected future performances can support performance, motivation, and flow over repeated VR exergame sessions. Palmieri and Deutsch (Palmieri and Deutsch, 2024) found higher exercise intensity and motivation during competitive VR bicycling than during visual-feedback-only cycling, including vigorous intensity during self-competition. These findings reinforce the need to distinguish competitive framing from the pacing and feedback algorithms through which it is implemented.

Friend-modelled social feedforward integrates cooperative and competitive mechanisms by strengthening relatedness and identification through the familiar partner, while simultaneously maintaining a clear competence-based challenge via enhanced pacing. Evidence suggests that this hybrid can yield higher performance and intrinsic motivation compared with self- or stranger-based feedforward at matched speeds, aligning with SDT’s prediction that satisfying relatedness and competence concurrently can support higher-quality motivation (Barathi et al., 2024; Deci and Ryan, 1985; Ryan and Deci, 2017). This perspective also reframes matchmaking: pairing friends of similar baseline output and tuning the challenge upward may reduce social loafing risks and accentuate facilitation, but this claim should be tested rather than assumed (Barathi et al., 2024; Karau and Williams, 1993; Latané et al., 1979).

2.5 Feedforward and adaptive systems

Feedforward refers to a self-modelling approach in which individuals are shown an enhanced representation of their own performance to promote skill development, grounded in motor learning frameworks indicating that self-comparison can support retention and improvement (Guadagnoli and Lee, 2004; Wulf and Shea, 2002). In VR exergaming, interactive feedforward operationalizes this by replaying a player’s prior performance as a live “ghost” and increasing resistance so that keeping pace requires higher actual output (Barathi et al., 2018). Controlled evaluations show that interactive feedforward improves average power while maintaining intrinsic motivation, and is superior to racing a generic virtual competitor, even when players are aware of the manipulation—an important consideration for repeated use (Barathi et al., 2018).

The social extension of feedforward replaces (or supplements) the self-model with an enhanced friend model. This “social interactive feedforward” leverages SDT’s relatedness and social comparison to intensify identification with the target, thereby potentially magnifying both motivational and performance gains compared to self- or stranger-based models at the same nominal enhancement (Barathi et al., 2024; Deci and Ryan, 1985; Ryan and Deci, 2017). Conceptually, this integrates three strands: (1) feedforward provides a concrete future-self or peer target (Barathi et al., 2018; Guadagnoli and Lee, 2004); (2) social facilitation explains why the co-actor’s presence may elevate effort (Guerin, 1993; Zajonc, 1965); and (3) SDT explains why friend-based relatedness can preserve or heighten intrinsic motivation under elevated challenge (Deci and Ryan, 1985; Ryan and Deci, 2017). Empirical results from friend-modelled VR cycling support this synthesis, reporting higher outputs and intrinsic motivation than self-modelled feedforward and articulating actionable design levers (identity binding, challenge calibration, and visibility of progress) (Barathi et al., 2024).

Finally, linking feedforward to affective measurement closes the loop for adaptive systems. Psychophysiological markers used in high-intensity VR exergaming (e.g., skin conductance, pupillometry, and fixations) can serve as online signals for when feedforward or social challenges are motivating versus overwhelming (Cacioppo et al., 2007; Boucsein et al., 2012; Barathi et al., 2020). These signals may inform closed-loop adjustment of resistance, opponent pacing, and social cues to maintain challenge and intrinsic motivation during a session, although such adaptation requires validation over repeated use. Holistic, adaptive VR fitness setups similarly show promise for engagement (Martin-Niedecken et al., 2019).

2.6 Synthesis and study rationale

Taken together, the preceding literature suggests that social exergaming effects cannot be understood only as a general benefit of exercising with others. Instead, different social mechanisms may shape exertion and motivation through distinct pathways. Social facilitation theory explains why the perceived presence of another agent may increase effort and performance; Self-Determination Theory helps differentiate whether this effort is supported through relatedness, competence, or autonomy; social presence research explains why avatars and virtual partners can feel motivationally meaningful; and feedforward/adaptive systems show how challenge and feedback can be calibrated in real time.

However, these strands have rarely been integrated within the same immersive VR exercise study. In particular, it remains unclear whether cooperative and competitive dynamics produce different patterns of physical output, physiological engagement, intrinsic motivation, and subjective experience when tested under the same task and measurement protocol. This motivates the present research questions: first, to compare how social interaction models affect cycling performance; second, to examine their effects on intrinsic motivation; and third, to use participants’ subjective accounts to explain the mechanisms through which cooperation and AI-driven competition shape exertion and engagement.

3 Research questions

• RQ1. How do different social interaction models in VR exergaming influence exercise performance (mean power output)?
• RQ2. How do these social-exercise implementations affect IMI enjoyment as the primary motivational outcome, and what secondary profiles appear across competence, relatedness-relevant experience, effort, value, autonomy, and pressure/tension?

• RQ3. How do participants’ subjective experiences help explain differences in motivation across social modes?
• RQ4. What mechanisms (e.g., perceived support, comparison pressure, challenge calibration) do participants describe that explain why cooperative and competitive models affect motivation and engagement differently?

4 Hypotheses

Based on Self-Determination Theory and Social Facilitation Theory, confirmatory analyses focused on mean power output, IMI enjoyment, and post-session perceived exertion. Other IMI subscales, physiological measures, and open-ended responses were treated as secondary or exploratory evidence. Because the AI-Competitive condition also included adaptive pacing, hypotheses concern differences between complete implementations rather than the isolated causal effect of competition.

H1 Compared with Solo cycling, both Cooperative and AI-Competitive cycling will increase mean power output.

H2 The Cooperative-framed and AI-Competitive implementations will show different motivational profiles: Cooperative-framed cycling is expected to show higher enjoyment and more relatedness-oriented descriptions, whereas AI-Competitive cycling is expected to show more competence- and challenge-oriented descriptions. IMI subscales other than enjoyment are treated as secondary outcomes.

H3 Compared with Solo cycling, both Cooperative and AI-Competitive cycling will reduce perceived exertion despite higher objective workload.

In addition to these hypotheses, we used physiological measures and qualitative responses to explore how participants experienced the two social conditions. Specifically, we examined whether Cooperative cycling was described in terms of relatedness, support, and shared pacing, while AI-Competitive cycling was described in terms of challenge, comparison, and goal striving.

Each participant completed three within-subject sessions; the Solo condition was administered first to obtain baseline data, and the order of the two social conditions was counterbalanced.

5 Methodology

5.1 Study overview

This study employed a within-subject experimental design to examine how different social interaction models—solo, friend-modelled cooperative, and AI-driven competitive—affect exercise performance, intrinsic motivation, and engagement during virtual reality (VR) cycling; the Solo baseline was always first, followed by counterbalanced Cooperative and AI-Competitive sessions. Each participant completed three partially counterbalanced sessions under these social conditions, allowing direct comparison of physiological and psychological responses while controlling for inter-individual differences.

The experimental design was approved by the institutional research ethics committee, and all participants provided written informed consent prior to participation.

5.2 Participants

We enrolled 30 healthy adults (M age = 23.2 ± 3.7 years). Participants reported regular engagement in light-to-moderate physical activity but no professional athletic training. Exclusion criteria were assessed through a self-reported pre-screening questionnaire and included any self-reported cardiovascular disorder, high susceptibility to motion sickness, or prior adverse experience with VR. Prior adverse VR experience was assessed using a binary yes/no item asking whether participants had previously experienced significant discomfort, sickness, panic, or other adverse reactions when using VR. All participants completed the full protocol. In addition to statistical significance testing, effect sizes and 95% confidence intervals are reported to support interpretation of the magnitude and practical relevance of observed differences.

Participants completed a brief pre-session questionnaire assessing their relationship to the cooperative pacing avatar. Prior to the VR sessions, participants indicated whether they were participating with a known peer (e.g., friend, family member, partner, or classmate). If applicable, they selected the relationship category and rated perceived closeness on a 1–7 scale. These self-reports were used to classify whether the cooperative session used a self-modelled avatar (self-model cooperative) or a known-peer avatar (friend-model cooperative), and to support the interpretation of relatedness and social presence outcomes.

5.3 Apparatus and materials

The VR exergaming system was developed in Unity 2022.3 and executed on a Windows PC equipped with a Meta Quest3 headset, motion sensors, and a Wattbike providing real-time power and cadence data. Heart rate (HR) was continuously measured using a Polar H10 chest strap, integrated through a Python-to-Unity communication bridge; prior validation work supports the Polar H10 as a source of RR-interval data during rest and exercise (Gilgen-Ammann et al., 2019). Electrodermal activity (EDA) and skin temperature were measured using a Shimmer3 GSR + sensor equipped with Ag/AgCl electrodes (Shimmer Sensing, Dublin, Ireland), following standard electrodermal recording principles (Boucsein et al., 2012). The device records skin conductance and temperature continuously at 51.2 Hz and includes an optical PPG channel for estimating heart rate, providing synchronized psychophysiological data during VR exercise.

The Wattbike supplied cycling power and cadence to the VR application; the Polar H10 supplied heart-rate and RR-interval data through the Python-to-Unity bridge; and the Shimmer3 recorded electrodermal activity and skin temperature. Device streams were time-aligned with VR condition and event records for subsequent multimodal analysis (Figure 1).

Participants viewed a virtual cycling track rendered as a continuous loop with moderate visual complexity to minimize visual fatigue. Consistent soundscapes, including ambient wind and soft instrumental background music, were presented across all conditions. Environmental factors such as lighting, temperature (maintained at 21 C ± 1 °C), and ventilation were standardized throughout all sessions to ensure experimental control.

Three conditions were implemented in the virtual environment. In Solo, participants cycled without a virtual counterpart. In the Cooperative-framed condition, participants rode beside a replay-based virtual pacing partner generated from their own or a known peer’s previously recorded performance. In AI-Competitive, participants rode against a computer-controlled virtual agent whose pace changed in real time according to the participant’s heart rate. Replay-based virtual partners and computer-controlled virtual agents are therefore distinguished from avatars controlled live by another person (Figure 2).

5.4 Experimental conditions

Each participant completed three 20 min VR cycling sessions: (1) Solo baseline, (2) Cooperative-framed pacing, and (3) AI-Competitive pacing. Solo recorded the participant’s initial self-paced performance and generated the self-replay profile. In Cooperative-framed pacing, participants selected either a self-modelled replay of their Solo performance or a friend-modelled replay from a known peer. The two riders did not contribute to a shared score or a mutually dependent team outcome; the manipulation emphasized companionship, shared pacing, and cooperative social framing. In AI-Competitive pacing, participants rode against an adaptive virtual agent whose pace was adjusted according to heart rate relative to an individually calibrated target zone. Because only AI-Competitive included this adaptive regulation, condition effects represent differences between bundled implementations.where H R l o w and H R h i g h denote the lower and upper bounds of the participant-specific target zone. During the AI-Competitive session, the opponent’s pacing was adjusted according to the following rule (Figure 3):where v A I t is the virtual opponent’s pacing speed and Δ v is the incremental adjustment applied at each update interval. This rule increased the opponent’s pace when participants were below the target zone, maintained pace when they were within the zone, and reduced pace when they exceeded the zone. The purpose of this adaptation was to maintain a competitive challenge while reducing the risk of excessive exertion.

H R l o w ≤ H R t ≤ H R h i g h

v A I t + 1 = v A I t + Δ v , i f H R t < H R l o w v A I t , i f H R l o w ≤ H R t ≤ H R h i g h v A I t − Δ v , i f H R t > H R h i g h

The Solo session was always administered first to capture each participant’s baseline cycling profile and to generate the self-model used in the cooperative condition. The order of the remaining two sessions (Cooperative and AI-Competitive) was counterbalanced across participants to mitigate potential order and fatigue effects.

Participant instructions. Participants were asked to cycle continuously for the full trial at a sustainable effort and were informed that they could stop if they experienced discomfort. In Solo, they selected their own pace without a virtual counterpart. In Cooperative-framed pacing, they were asked to ride alongside the replay-based partner and use it as a shared pacing reference rather than compete for an individual win. In AI-Competitive pacing, they were asked to use the adaptive agent as a performance target while remaining within a sustainable effort. The wording reported here should be read as the functional content of the instructions; no claim is made that the Cooperative-framed condition created a strongly interdependent team task.

5.5 Procedure

Participants first attended an introductory session that included informed consent, familiarization with the VR system, and sensor calibration. During each experimental session, participants were fitted with a heart rate monitor (Polar H10), a galvanic skin response (GSR) sensor, and a VR headset. They began with a 3 min self-paced warm-up to stabilize heart rate, followed by a 20 min cycling trial conducted in one of the three experimental modes. After each session, participants completed standardized self-report questionnaires and performed a five-minute cool-down.

Post-session open-ended written responses. After completing the questionnaires, participants answered a short set of open-ended questions about (i) what motivated them in the session, (ii) how the social counterpart affected effort and enjoyment, and (iii) any usability or discomfort issues. These responses were used for the qualitative analysis reported in Section 3.10.

Cycling performance data—including speed, cadence, and power output—and heart rate were automatically recorded at 1 Hz and exported as CSV files for later analysis. Electrodermal activity (EDA) and skin temperature were simultaneously measured using the Shimmer3 GSR + sensor (sampling rate: 51.2 Hz) and time-synchronized with the cycling and VR event data to enable multimodal analysis.

5.6 Measures

Performance Metrics: Cycling performance was continuously recorded by the VR system and smart trainer to quantify physical output and pacing stability. The key metrics included mean power output (W), representing average mechanical intensity; mean cadence (rpm), reflecting pedalling rhythm and motor engagement; and total cycling duration (20 min fixed across all conditions).

Physiological Measures: Heart rate (HR, bpm) was recorded using a Polar H10 chest strap, while heart-rate variability (HRV, RMSSD) was derived from inter-beat intervals as an exploratory index of autonomic variation (Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology, 1996). Electrodermal activity (EDA, µS) was collected using Ag/AgCl finger electrodes. Mean skin conductance level and phasic response frequency were used as indices of sympathetic arousal (Boucsein et al., 2012); higher values indicate greater arousal but do not identify whether affect is positive or negative. Skin temperature (°C) was treated as a complementary peripheral measure because it may vary with thermoregulation, vascular response, movement, ambient conditions, and sensor contact. It was not interpreted as a direct measure of emotional valence.

Psychological Measures: After each session, participants completed the Intrinsic Motivation Inventory (IMI), covering enjoyment, effort, value, competence, autonomy, and pressure/tension (McAuley et al., 1989). Enjoyment was the primary IMI outcome; the remaining subscales were secondary. The Social Presence Scale for VR assessed perceived co-presence and connection (Makransky et al., 2017), and PACES measured exercise enjoyment (Kendzierski and DeCarlo, 1991). Perceived exertion was recorded once immediately after each trial using the Borg category-ratio scale (CR10; 0 = nothing at all, 10 = maximal) (Borg, 1982). Because no resting RPE was recorded, analyses compare post-session ratings rather than within-session change from baseline.

Qualitative Measures: Participants also provided open-ended responses after each condition. Responses were analysed using inductive thematic analysis (two-pass coding; disagreements resolved through discussion) to identify recurring motivational and experiential themes aligned with SDT and social facilitation constructs (Braun and Clarke, 2006).

In addition, participants rated perceived usability and immersion after each session using two single-item 7-point Likert scales (1 = strongly disagree; 7 = strongly agree). These items were included as post-session control measures to assess interface clarity and perceived experiential absorption.

5.7 Stimulus and environmental control

To isolate the effects of social dynamics, all sessions were conducted within the same virtual environment, featuring identical track design, lighting, and audio conditions. The smart trainer resistance was kept constant within participants and calibrated during familiarisation to elicit a moderate, sustainable intensity indexed by heart rate and perceived exertion (we did not collect metabolic data and therefore did not estimate VO₂max). In the AI condition, the opponent’s pacing was adaptively modulated based on the participant’s real-time heart rate relative to an individually calibrated target zone. The Solo condition was always administered first; the order of the two subsequent social conditions was counterbalanced across participants to mitigate order and fatigue effects. All participants were instructed to maintain adequate hydration and refrain from strenuous physical activity for 24 h prior to testing.

5.8 Ethical considerations

The study was approved by the institutional research ethics committee and conducted in accordance with the Declaration of Helsinki. All participants received detailed information sheets and provided written informed consent before participation, with the right to withdraw at any time without penalty.

A qualified first-aid researcher was present throughout all sessions. The VR cycling system underwent safety checks prior to each trial, and breaks were provided upon request to minimize cybersickness or overexertion. No adverse events were reported.

5.9 Statistical analysis

All statistical analyses were performed using R (version 4.3) and SPSS (version 29). One-way repeated-measures ANOVAs compared the three conditions for the prespecified outcomes of mean power output, IMI enjoyment, and post-session RPE. Other IMI subscales and physiological variables were analysed as secondary or exploratory outcomes. Greenhouse-Geisser corrections were applied when sphericity was violated, and post-hoc pairwise comparisons used Bonferroni adjustment. Effect sizes were reported as partial eta squared for omnibus tests and Cohen’s d for pairwise contrasts, with 95% confidence intervals where available. Pearson correlations were exploratory. Statistical significance was set at alpha = 0.05. For Figure 10, within-subject 95% confidence intervals were calculated separately for each IMI subscale using participant-level normalization and the Morey correction (Morey, 2008).

5.10 Qualitative analysis

Open-ended written responses collected after each session were analysed using inductive thematic analysis (Braun and Clarke, 2006). The analysis focused on participants’ descriptions of motivation, perceived effort, social interaction, challenge, and usability across the Solo, Cooperative, and AI-Competitive conditions.

The coding process followed a two-pass procedure. In the first pass, the first author reviewed all responses to become familiar with the dataset and generated initial descriptive codes by grouping semantically similar statements. These initial codes included, for example, challenge and goal pursuit, shared pacing, social accountability, reduced isolation, effort distraction, performance comparison, immersion and feedback clarity, and discomfort or usability issues. In the second pass, the codes were refined through constant comparison and organized into higher-level themes that captured recurring motivational mechanisms across conditions.

Coding was conducted by the first author, and ambiguous coding decisions were discussed with the supervisory team during manuscript development.

6 Results

All statistical analyses were performed using repeated-measures ANOVA with Bonferroni-corrected post hoc comparisons. Across performance, physiological activation, and psychophysiological arousal variables, significant main effects of social condition were observed, indicating that social dynamics in VR exergaming were associated with differences in exertion and engagement.

6.1 Performance metrics

Mean cycling power output increased progressively across conditions, from Solo (M = 127.1 W, SD = 18.6) to Cooperative (M = 140.2 W, SD = 17.4) and AI-Competitive (M = 147.0 W, SD = 19.2). A repeated-measures ANOVA revealed a significant main effect of condition, F(2, 58) = 34.81, p < 0.001, η_p ² = 0.55. Post hoc comparisons indicated that both social conditions yielded significantly higher power than Solo (Cooperative vs. Solo: t(29) = 8.53, p < 0.001, d = 1.88; AI vs. Solo: t(29) = 7.12, p < 0.001, d = 1.39). The AI condition also exceeded the Cooperative condition (t(29) = 3.55, p = 0.001, d = 0.69).

Power output was therefore higher in both social conditions than in Solo and was highest in AI-Competitive. Power trajectory analysis showed a positive linear slope in Cooperative-framed pacing (M = 0.046, p = 0.02), whereas Solo trials were stable or slightly declining (Figures 4, 5).

6.2 Physiological responses

Heart rate (HR) increased significantly across conditions—Solo (M = 122 bpm), Cooperative (M = 128 bpm), and AI (M = 133 bpm)—with a main effect of condition, F(2, 58) = 41.27, p < 0.001, η_p ² = 0.59. Post hoc tests showed all pairwise comparisons were significant (p < 0.001), indicating higher cardiovascular activation in social contexts.

Heart rate variability (HRV, RMSSD) also differed significantly across conditions, F(2, 58) = 19.32, p < 0.001, η_p ² = 0.40. Both Cooperative and AI conditions showed higher HRV than Solo (Cooperative–Solo: t(29) = 4.15, p < 0.001, d = 0.84; AI–Solo: t(29) = 4.57, p < 0.001, d = 1.03). Together, elevated HR with higher HRV may indicate increased physiological activation alongside preserved autonomic variability; stress-related interpretations cannot be ruled out from these measures alone (see Figure 6).

6.3 Electrodermal activity (EDA)

Skin conductance levels (SCL) were higher in both social conditions than in Solo (Cooperative-framed-Solo: t(29) = 6.85, p < 0.001, d = 1.95; AI-Solo: t(29) = 8.01, p < 0.001, d = 3.62). Peak EDA frequency increased from Solo to Cooperative-framed (t(29) = 4.22, p < 0.001) and was lower in AI-Competitive than in Cooperative-framed (t(29) = −5.31, p < 0.001). These measures indicate differences in sympathetic arousal but do not distinguish positive from negative affect (Figure 7).

6.4 Exploratory subgroup analysis: self-model vs. friend-model cooperative

Within the cooperative session, participants could choose to ride alongside either a self-modelled avatar derived from their own Solo baseline performance or a friend-modelled avatar derived from a known peer’s baseline performance. Because this choice was participant-selected rather than randomised, the following comparison is reported as an exploratory, descriptive subgroup analysis rather than a causal test.

Participants who selected the self-model (n = 9) showed marginally greater mean power gains (+14.4 W) than those who selected the friend-model (+12.5 W), although this difference was not statistically significant, t(18) = 0.65, p = 0.52, g = 0.26 (Figure 8). The self-model subgroup exhibited smoother pacing and higher HRV, whereas the friend-model subgroup showed more frequent EDA peaks; because the subgroup comparison was participant-selected and descriptive, this pattern should be interpreted cautiously.

6.5 Correlations between arousal and regulation

In the AI condition, HRV was positively correlated with mean EDA, r(28) = 0.45, p = 0.018, indicating that participants who maintained stronger autonomic flexibility also showed greater emotional activation. This association is consistent with co-variation between arousal and autonomic regulation during adaptive challenge, but it does not establish a co-regulation mechanism (Figure 9).

6.6 Intrinsic motivation inventory (IMI)

Repeated-measures ANOVAs showed significant main effects of condition for all IMI subscales: Enjoyment, F(2, 58) = 84.40, p < 0.001; Effort, F(2, 58) = 15.15, p < 0.001; Value, F(2, 58) = 43.10, p < 0.001; Competence, F(2, 58) = 13.15, p < 0.001; Autonomy, F(2, 58) = 12.31, p < 0.001; and Pressure, F(2, 58) = 49.61, p < 0.001.

The condition means indicated that enjoyment, effort, and value were highest in the Cooperative-framed condition, followed by AI-Competitive and Solo. Pressure/tension ratings were lowest in the Cooperative-framed condition. Post hoc comparisons indicated that Cooperative-framed > AI-Competitive > Solo for enjoyment (all p < 0.001) and that both social conditions exceeded Solo for value, competence, and autonomy. Pressure/tension ratings were lower in both social conditions than in Solo, with the lowest ratings in the Cooperative-framed condition. These results indicate that the Cooperative-framed implementation was associated with the broadest gains across the IMI subscales in this sample (see Figure 10).

6.7 Perceived exertion (Borg RPE)

Post-session perceived exertion differed significantly across conditions, F(2, 58) = 22.07, p < 0.001, with mean CR10 ratings of 4.56 ± 2.71 for Solo, 1.66 ± 1.87 for Cooperative-framed, and 1.88 ± 1.61 for AI-Competitive (Figure 11). Both social conditions had lower ratings than Solo (p < 0.001), while Cooperative-framed and AI-Competitive did not differ (p = 0.64). Because a resting or pre-session RPE value was not collected, these results describe differences in post-session ratings rather than change from baseline.

6.8 Post-scenario ratings

Closed-ended items showed high motivation, usability, and subjective absorption ratings across social conditions. Participants agreed that both social conditions encouraged greater effort (means > 4.6/7) and made them work harder than in a standard gym workout (means > 5.0/7). Usability and subjective absorption ratings exceeded 5.5, while motion-sickness ratings remained low (approximately 1.2).

6.9 Qualitative insights

Inductive thematic analysis identified four primary themes reflecting how participants interpreted motivation, effort, social presence, and challenge across the three VR cycling scenarios. These themes were used to contextualize the quantitative findings rather than as independent confirmatory outcomes.

6.9.1 Theme 1: competitive drive and performance challenge

Participants frequently described the AI-Competitive condition as motivating because it provided a clear performance target and encouraged comparison with the virtual opponent. For example, one participant stated, “The ghost of my previous best kept me pushing” (P12, AI-Competitive; mean power = 147.33 W; RPE = 2.00; IMI enjoyment = 4.67/7). This response illustrates how the competitive virtual agent redirected attention toward goal pursuit and helped sustain effort. Several participants similarly reported that the AI virtual agent increased focus and intensity, although a minority described the competitive target as pressuring when it felt difficult to catch or maintain pace.

6.9.2 Theme 2: social presence and shared accountability

Cooperative sessions were often described in terms of companionship, shared pacing, and reduced isolation. For example, one participant wrote, “Riding with a teammate made me feel less alone and more committed” (P24, Cooperative; mean power = 180.03 W; RPE = 0.50; IMI enjoyment = 5.00/7). This theme helps explain why Cooperative cycling produced the strongest enjoyment and relatedness-oriented motivation: participants did not simply respond to the avatar as a pacing cue, but as a socially meaningful partner that made exertion feel shared and supported.

6.9.3 Theme 3: usability and immersive feedback

Participants also highlighted the importance of clear visual feedback, responsive pacing, and intuitive system interaction. These comments suggest that the motivational effects of cooperation and competition were supported by the usability of the VR cycling interface. Clear feedback helped participants interpret their current performance and adjust pacing during the session, reducing ambiguity and supporting flow-like engagement.

6.9.4 Theme 4: balanced engagement under social modulation

Several responses reflected a dynamic relationship between effort and enjoyment. Cooperative cycling was often described as making the session feel more socially comfortable or making “time pass faster,” whereas AI competition was associated with alertness, challenge, and sustained focus. Across themes, qualitative findings were consistent with the exploratory qualitative expectation: Cooperative cycling was described in terms of relatedness, shared accountability, and reduced isolation, whereas AI-Competitive cycling was described in terms of challenge, comparison, focus, and performance striving.

Across themes, qualitative findings converged with quantitative results: cooperation was linked to relatedness and accountability, whereas AI competition emphasized competence and performance striving.

7 Discussion

7.1 Overview

This study compared three bundled VR cycling implementations. Both social conditions were associated with higher physical output than Solo, Cooperative-framed pacing showed the highest enjoyment, and AI-Competitive pacing showed the highest power output. These results are consistent with complementary relatedness- and competence-oriented experiences, but they do not isolate cooperation or competition as single causal factors. In particular, AI-Competitive combined competitive framing with heart-rate-responsive pacing, while Cooperative-framed used a replay-based pacing partner without a shared team outcome.

7.2 Social facilitation and physiological engagement

Consistent with Social Facilitation Theory, participants showed greater physical output when accompanied by cooperative or competitive agents compared with solo cycling (Park and Catrambone, 2007; Guerin, 1993; Zajonc, 1965; Blascovich et al., 2002). The observed increases in power output, HR, and EDA suggest that the presence of others—whether represented by a friend or an adaptive AI—was associated with higher exertion and emotional engagement (Hibbs et al., 2024; Tarr et al., 2018).

The concurrent elevation of HR and HRV may indicate increased arousal alongside preserved autonomic variability rather than simple overload, but this interpretation should be treated cautiously (Cacioppo et al., 2007; Task Force of the European Society of Cardiology and the North American Society of Pacing and Electrophysiology, 1996). From a design perspective, this suggests that socially contingent agents may enhance performance intensity when their behaviour is calibrated to the user’s exercise state rather than simply increasing pressure. In the present system, this was implemented through pacing-based adaptation: the AI-Competitive opponent adjusted speed relative to the participant’s heart-rate target zone, while the Cooperative avatar provided a socially meaningful pacing reference. Future systems could extend this approach by adapting opponent distance, pacing, challenge intensity, and supportive feedback using physiological and self-report signals such as HR, HRV, EDA, RPE, and discomfort reports.

However, this implication should not be read as evidence that any AI agent will produce similar motivational or physiological effects. More anthropomorphic or conversational AI companions may introduce additional social variables, including perceived agency, trust, realism, social presence, voice, responsiveness, and users’ willingness to relate to the agent as a social partner. Prior work on social responses to computers and virtual agents shows that users may respond socially to computational systems, but that trust and social presence are shaped by perceived anthropomorphism and agent behaviour (Nass and Moon, 2000; Munnukka et al., 2022; Fakhimi et al., 2023). Relational agents have also been used to support physical activity, but their effectiveness depends on repeated interaction, perceived rapport, and users’ acceptance of the agent as a health-supportive partner (Bickmore et al., 2005; Sillice et al., 2018). Future VR exergaming research should therefore examine how agent-level attributes, such as embodiment, voice, anthropomorphic cues, relational framing, and adaptive responsiveness, moderate the motivational effects observed in the present pacing-based AI-Competitive condition.

Post-scenario ratings indicated that both social conditions received high subjective absorption and usability ratings. Social presence refers here to the subjective experience of being with a virtual counterpart, whereas immersion refers to properties of the VR system; the single-item rating is therefore interpreted as subjective absorption rather than as a technical measure of immersion.

7.3 Self-determination and motivational dynamics

Interpreting Table 1 through SDT, Social Facilitation Theory, social presence, and adaptive-challenge frameworks (Tarr et al., 2018; Deci and Ryan, 1985; Ryan and Deci, 2017; Zajonc, 1965; Guadagnoli and Lee, 2004), the findings suggest that the bundled implementations were associated with different psychological profiles. Cooperative-framed pacing was associated with enjoyment and relatedness-oriented descriptions, whereas AI-Competitive pacing combined competence-oriented comparison with adaptive challenge. Because feedback and pacing were not matched across conditions, these differences should not be attributed solely to social goal structure.

TABLE 1. Empirically observed social mechanisms and motivational patterns in immersive VR cycling.

Importantly, these need-based effects were not mutually exclusive. The friend model combined relatedness with gradual performance gains, suggesting that affective connection may scaffold mastery-oriented engagement during the session. In contrast, the AI model elicited peak performance while maintaining smoother physiological arousal patterns, indicating a high-engagement state rather than isolated competitive spikes. This pattern suggests that competence activation in immersive VR can occur without necessarily undermining autonomy or inducing excessive stress, although longer-term evidence is needed.

The qualitative themes further illuminate this dynamic interplay. Social Presence and Team Support narratives reinforced the role of relatedness in buffering perceived strain, whereas Competitive Drive and Challenge narratives reflected competence-oriented regulation of effort. Together, these findings extend SDT applications in VR exergaming by suggesting that social mechanism design—not merely task intensity—structures how competence and relatedness are experienced during high-intensity exercise.

7.4 Complementary mechanisms and design balance

The results highlight that Cooperative and AI-Competitive modes should not be interpreted as one uniformly superior social design. Instead, they produced partially overlapping but motivationally distinct effects. Both social modes reduced perceived exertion relative to Solo, and they did not differ significantly from each other on RPE, suggesting that socially mediated engagement may generally buffer subjective effort during VR cycling. However, they differed in other outcomes: Cooperative cycling produced the strongest enjoyment, relatedness-oriented motivation, and perceived social presence, whereas AI-Competitive cycling produced the highest mean power output and a more challenge-oriented performance profile.

Therefore, the design implication is not that systems should automatically switch between cooperation and competition. Rather, VR exergames should align social mode selection with the intended exercise goal and user state. Cooperative pacing may be more appropriate when the design goal is enjoyment, comfort, relatedness, or repeated-use motivation, while adaptive competition may be more appropriate when the goal is short-term intensity, challenge, or performance improvement. Psychophysiological feedback may help identify when a selected mode is becoming under-stimulating or overly demanding, but the present study does not directly test whether dynamic switching improves adherence or motivation.

Future work should therefore evaluate adaptive mode-selection systems explicitly, for example, by comparing fixed cooperative, fixed competitive, user-selected, and system-adaptive conditions over repeated sessions. Such work should also consider user preference, competitiveness, cooperation preference, comfort, and prior relationship with virtual partners before assuming that switching between social modes will be beneficial. Longer-term adherence likely depends not only on adaptive challenge, but also on holistic and multimodal design factors, including whole-body interaction, family-friendly contexts, auditory augmentation, and personalized avatars (Martin-Niedecken et al., 2019; Ijaz et al., 2017; Ma et al., 2018; Sinclair et al., 2007).

7.5 Design implications

The present findings suggest several design implications for socially adaptive VR exergaming systems, grounded in the observed differential effects of cooperative and competitive social models.

• Strategically Deploy Cooperative and Competitive Modes for Distinct Motivational Goals.

Our results indicate that cooperative configurations primarily enhanced intrinsic enjoyment, relatedness, and perceived effort buffering, whereas AI-driven competition was associated with the highest performance intensity and competence-driven striving. Rather than treating cooperation and competition as interchangeable social features, designers should align these mechanisms with intended outcomes—using cooperative pacing to support enjoyment and intrinsic motivation, and adaptive competition to amplify short-term performance intensity.

• 2. Leverage Social Presence to Modulate Perceived Exertion.

The dissociation between increased objective workload and reduced perceived effort in social conditions suggests that immersive social presence may attenuate subjective strain during high-intensity exercise. Designers may therefore incorporate co-action cues, shared pacing visualization, or comparative ghost agents to support effort without increasing perceived fatigue. Importantly, such mechanisms should preserve autonomy and avoid excessive evaluative pressure.

• 3. Integrate Physiological Feedback for Adaptive Social Calibration.

The concurrent elevation of HR and HRV may indicate increased arousal alongside preserved autonomic variability under moderate social pressure. This supports further investigation of psychophysiological signals (e.g., HR, EDA) for dynamically adjusting pacing difficulty or social intensity. Systems could modulate challenge levels or shift between cooperative and competitive framing based on real-time arousal patterns, but this adaptive mode-selection approach remains a design hypothesis.

• 4. Balance Familiarity and Challenge in Social Agent Design.

Friend-modelled cooperation was associated with affective engagement, while adaptive AI competition was associated with a stable high-performance profile in this study. Designers should therefore balance relational familiarity with structured challenge, avoiding overstimulation while preserving motivational salience.

Together, these implications highlight that social mechanism design in immersive VR should not merely increase engagement, but deliberately structure how competence and relatedness are experienced across exercise contexts.

7.6 Limitations and future work

Although the controlled laboratory design ensured experimental rigor, it does not fully capture the complexity of at-home or online multiplayer VR exercise contexts. In the laboratory, the physical environment, equipment setup, session duration, safety monitoring, and technical support were standardized. This strengthens internal validity, but it may overestimate how smoothly users would engage with the system in everyday use. At home, users may experience setup burden, limited space, inconsistent internet connectivity, distractions, safety concerns, and reduced researcher support. These factors could affect not only adherence, but also the perceived value of cooperative or competitive social modes. Future studies should therefore examine whether the motivational benefits observed here persist when users must independently set up the equipment, manage discomfort, and decide whether to return to the system over multiple weeks.

The principal design limitation is that the Cooperative-framed condition did not implement a strongly interdependent cooperative task. Participants rode beside a replay-based self- or friend-modelled partner, but there was no shared score, reciprocal control, or team outcome. The condition therefore manipulated cooperative framing, companionship, and shared pacing more directly than formal cooperation. In addition, participants selected the self- or friend-model rather than being randomly assigned, creating heterogeneity within the condition. Future work should compare this design with cooperative tasks based on a shared power target, team score, relay structure, or other mutually dependent outcome.

A second design limitation is that AI-Competitive combined competitive framing with heart-rate-responsive pacing, while Solo and Cooperative-framed pacing did not receive equivalent adaptive regulation. The higher power output in AI-Competitive may therefore reflect competition, algorithmic challenge, or their combination. Solo was also always administered first because it generated the replay profile, so practice, familiarization, or order effects cannot be fully separated from condition. Future studies should use matched or yoked pacing algorithms and, where technically possible, fully counterbalance condition order.

A second limitation is that we did not collect metabolic data, such as oxygen consumption, and therefore could not quantify VO₂-based exercise intensity. Instead, exercise intensity was indexed using power output and cardiovascular/psychophysiological signals, including HR, HRV, and EDA. This limits interpretation in two ways. First, we cannot determine whether the higher power output observed in the AI-Competitive condition reflected a proportional increase in metabolic cost or whether participants achieved greater mechanical output with different efficiency. Second, without VO₂ or energy-expenditure data, we cannot directly classify the physiological intensity of each condition using standard exercise physiology thresholds. As a result, our interpretation should be limited to relative differences in mechanical performance, cardiovascular activation, and subjective experience, rather than claims about metabolic workload or training dose. Future work should integrate VO₂, respiratory exchange ratio, or indirect calorimetry to determine whether socially mediated increases in power output translate into meaningful changes in energy expenditure and exercise intensity.

The sample was also relatively homogeneous, consisting primarily of young adults with moderate fitness. This limits demographic generalization because different user groups may respond differently to cooperative and competitive social mechanisms. Older adults, for example, may place greater value on safety, comfort, social support, and reduced intimidation, and may therefore respond more positively to cooperative pacing than competitive pressure. Rehabilitation users may require more conservative intensity adaptation, clearer safety feedback, and clinician-supervised progression; for these users, AI competition could be motivating only if carefully calibrated to avoid overexertion or frustration. Conversely, trained athletes or highly competitive users may respond more strongly to adaptive competition and performance benchmarks than to cooperative support. Future studies should therefore test whether age, baseline fitness, clinical status, competitiveness, cooperation preference, and prior VR experience moderate the effects of social mode on power output, perceived exertion, enjoyment, and adherence. These population-specific considerations are consistent with prior reviews of the social effects of exergames in older adults (Li et al., 2018).

Another limitation is that the study did not include a pre-session validated measure of competitive orientation. In particular, we did not administer the 12-item Multidimensional Competitive Orientation Inventory (MCOI), which distinguishes hypercompetitive orientation, self-developmental competitive orientation, anxiety-driven competition avoidance, and lack of interest in competition (Orosz et al., 2018). Consequently, we could not test whether competitive orientation moderated participants' performance, enjoyment, pressure/tension, or perceived exertion in the AI-Competitive condition. Individual orientation may also have influenced responses to the self-model replay in the Cooperative-framed condition. Future studies should administer the MCOI before experimental exposure and test orientation-by-condition interactions rather than assuming that all users respond similarly to competition.

The study also recorded RPE only after each trial and did not obtain a resting or pre-session rating. Accordingly, the analysis cannot estimate change in perceived exertion from baseline. Post-session CR10 values should be interpreted as condition-specific ratings, and future studies should collect repeated RPE measurements alongside standardized verbal anchors.

Future research should move beyond the general question of whether social VR exergaming is effective and address more specific design and evaluation questions. First, longitudinal studies should examine whether cooperative and competitive benefits persist after novelty effects decline, and whether users continue returning when sessions are self-initiated rather than researcher-scheduled. Second, studies should compare fixed cooperative, fixed competitive, user-selected, and system-adaptive social modes to determine whether adaptive mode selection provides benefits beyond simply offering users a preferred mode. Third, networked multiplayer studies should examine whether real-time human partners produce different motivational and physiological effects from friend-modelled avatars or AI-controlled opponents. Fourth, VSA or AI-companion studies should test which agent attributes—such as voice, embodiment, emotional responsiveness, anthropomorphism, trustworthiness, and relational framing—are necessary for users to experience AI support as socially meaningful rather than merely instructional. Finally, home-based and telerehabilitation studies should examine safety, comfort, adherence, and clinical feasibility over repeated sessions, especially for older adults and rehabilitation users (Tarr et al., 2018; Deci and Ryan, 2008; Fernandes et al., 2022; Meulenberg et al., 2022; Rüth and Kaspar, 2021).

Moreover, integrating home-based and telerehabilitation applications could further validate these benefits beyond controlled laboratory environments (Fernandes et al., 2022; Meulenberg et al., 2022; Rüth and Kaspar, 2021).

7.7 Contributions

This research advances the understanding and design of socially adaptive VR exergaming systems through three primary contributions: A modular VR exergaming platform integrating real-time physiological and social feedback. We developed a Unity-based VR cycling system that synchronizes biometric inputs (HR, HRV, EDA) with social interaction models—solo, friend-modelled cooperative, and AI-driven competitive—enabling systematic investigation of motivation and physiology in immersive exercise. Empirical evidence for differential motivational pathways in social exergaming.

Through a within-subject experiment (N = 30), we observed distinct profiles across the bundled implementations. Cooperative-framed pacing was associated with enjoyment and relatedness-oriented accounts, while the adaptive AI-Competitive implementation produced the highest power output and competence-oriented accounts. These associations do not demonstrate that cooperation or competition alone produced the observed differences.

Our findings integrate these perspectives into a model explaining how social presence, competence, and adaptive feedback may jointly support engagement in VR fitness. This framework provides design guidance for developing socially intelligent exergames that personalize challenge, emotional tone, and social interaction; whether these features improve longer-term exercise adherence remains a question for longitudinal research.

8 Conclusion

This study suggests that different VR social-exercise implementations are associated with distinct short-term patterns of motivation, exertion, and enjoyment. Cooperative-framed replay pacing was associated with enjoyment, companionship, and shared pacing, while an adaptive AI-Competitive implementation was associated with higher power output and challenge-oriented experience. Because the conditions differed in both social framing and pacing control, the results should be interpreted as a comparison of complete implementations rather than isolated causal effects of cooperation and competition. Future matched-control and longitudinal studies are required to determine whether these patterns support sustained exercise engagement.

Beyond performance gains, socially intelligent exergaming may help make exercise more emotionally engaging, socially connected, and health-oriented, bridging the gap between fitness and meaningful human interaction.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. Generative AI tools were used to support language refinement, structural editing, and manuscript preparation. All intellectual content, study design, data analysis, interpretation, and conclusions were developed, verified, and approved by the authors. The authors take full responsibility for the final content of the manuscript.

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

Publisher’s note

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

Footnotes

Edited by: Beatrice Aruanno, University of Pisa, Italy

Reviewed by: Yogi Udjaja, National Taiwan Normal University, Taiwan

John Luca Palmieri, Icahn School of Medicine at Mount Sinai, United States