Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by cognitive, motor, and social difficulties [1], which affect daily functioning [2] and overall well-being [3]. Reduced participation highlights the importance of active engagement in supporting adaptive functioning and skill acquisition. A lack of active engagement may lead children with ASD to experience further developmental delays [4], increased risk of stereotyped and challenging behaviors [5], and a higher incidence of obesity [6]—all of which are interrelated and pose significant concerns. Given the scope of these difficulties, the continued rise in ASD diagnoses over the past decades underscores the importance of creating accessible interventions to support active participation in everyday contexts [7]. For example, serious games (SGs) have been explored as digital interventions that target functional outcomes beyond entertainment [8,9], including social communication, executive functioning, and motor coordination [10,11]. Recent SGs have broadened interaction design through modalities such as motion-tracking [12-15], interactive stimuli [16-18], extended reality [14,19], and socially assistive robots [20-23] or virtual agents (VAs) [24,25], enabling diverse forms of feedback and goal-directed activities [26-28]. Within this design space, SG interventions for children with ASD have positioned movement-based physical activities as a central design emphasis to support active participation.

The primary objective of this study is to iteratively design serious exergames through a structured co-design process involving professionals from multiple disciplines, including special educators, APE therapists, and human-computer interaction (HCI) researchers. Building on a series of expert interviews and collaboration, the study focuses on exergame development from the initial selection of exercises to the progressive shaping of gameplay. This process spans conceptualization, design, and development, and aims to support active participation and functional development in children with ASD. To explore feasibility and use contexts, we conducted an exploratory multiple-case pilot with 3 children with ASD, focusing on case-based description of engagement and interaction patterns.

This study involving autistic participants was conducted following the ethical guidelines and regulations of the Institutional Review Board at Gwangju Institute of Science and Technology (approval number HR-61-04-04). The approved protocol includes the involvement of professionals from multiple disciplines and the participation of children with ASD and their legal guardians (eg, parents).

Participant recruitment was conducted in accordance with Institutional Review Board guidelines. The study was introduced at the “Dream Tree Children Education Center” by therapists; 4 parents voluntarily enrolled their children for the experiment. Prior to participation, the experimental procedure and the role of the proxy user were explained to the parents, and informed consent was obtained. Additionally, each parent provided written assent, acknowledging their voluntary participation and the use of their data for research. They were informed that the experiment could be discontinued at any time upon request by the child, guardian, or therapist. The experiment commenced upon confirmation of the child’s willingness to participate, and participants received monetary compensation equivalent to approximately US $70.

To maintain confidentiality, all data were deidentified immediately upon collection. Unique identification codes were assigned to each participant, and personally identifiable information was stored separately in a secure, password-protected database, with access restricted to the primary researcher.

Exergames were implemented at a local special education center that supports mental and physical health, as well as basic motor function development, for children with ASD. A 55-inch screen was placed in front of the players to display the exergames (Figures 6C and 6D). A wearable E4 wristband and 2 cameras were used to minimize external stimuli and accommodate hypersensitivity in children with ASD [41]. The E4 wristband was attached to the participant’s wrist and connected to the SG to monitor behavioral and physiological data streams. Two cameras were positioned at the front and back to capture the player’s behavior and record video. Building on this setup, a real-time visualization platform was developed and deployed to facilitate engagement labeling by parents during gameplay. The platform synchronized sensor data with live video and was displayed on two 12.9-inch iPad Pro tablets—one displaying a live video feed of the child, and the other presenting time-series plots visualizing variations in physiological and behavioral signals, including engagement status, accelerometer data, galvanic skin response, heart rate, skin temperature, and motion tracking.

To examine the practical applicability of the iteratively developed exergames, we conducted a human-participant study involving children with ASD. In line with case-oriented approaches used in ASD-related studies—where high interindividual variability necessitates detailed observation at the individual level—3 children with ASD (P1-P3), aged between 10 years and 13 years (mean 11, SD 1.3 years), were finally recruited and evaluated [42]. P4, who was originally scheduled to participate, withdrew consent. The refusal stemmed from P4’s sense of loss and emotional displacement after observing P3 (a peer identified as a rival) establishing a close and successful play interaction with the Jibo robot in a preceding session, effectively feeling that the friend (robot) had been taken. The demographic details of the participants are summarized in Table 2. All planned engagement labeling data and playtimes were collected without missing entries for the included participants.

Before gameplay, participants were introduced to the exergames and informed of their respective roles (Figure 7). A preliminary session was conducted to facilitate rapport building (Figure 6A). During this session, participants interacted with Jibo, a physically embodied robot operated by the experimenter using a Wizard-of-Oz protocol. Children engaged in interactive play with the robot, which included head tilting to navigate a spaceship and responding to movement prompts (eg, dancing or playing guitar), involving audiovisual and physical interaction. The main session began after the therapist confirmed that sufficient engagement and rapport had been established. A practice session followed to assess the child’s readiness to engage with the exergames. Children who were able to follow the game flow—either independently or as determined by the therapist—proceeded to the main game session. The 2 exergames were presented in a counterbalanced order to ensure equitable exposure. The child participated in both games while each child’s parent, located in a separate room, labeled engagement status using a binary scheme (not engaged=0, engaged=1) based on real-time observation (Figure 6B) [44]. Parents were instructed to press the “Enter” key when the child appeared engaged and to refrain from pressing otherwise, following predefined operational definitions and behavioral indicators (Table 3). Before the main session, the experimenter explained the criteria and conducted a calibration using practice examples. We adopted binary labels to reduce boundary ambiguity that can increase subjectivity in finer-grained schemes (eg, ternary or multilevel labels). If the labeler judged that the annotation was incorrect over a given time-series interval, the label assignments for that interval were revised. After the gameplay session, a parent interview was conducted to gather feedback based on proxy observations of the child’s experience [45,46].

We analyzed the temporal trends in engagement status of children with ASD, based on labels provided by their caregivers during gameplay. As each child engaged in the game for a different time duration, we first summarized the total playtime for each child (Table 4). While P1 recorded a session of 3 minutes 13 seconds in the Hazard Avoiding Ride, the average playtime across games generally aligned with expert guidance for children with ASD. To support time-aligned comparisons of engagement across participants, the time axis for each child was normalized by dividing the elapsed time by the total game duration, resulting in a scale from 0 to 1. We then modeled changes in engagement over normalized time using generalized estimating equations (GEE) with a logistic link function to account for repeated measures within participants. Specifically, we fit a GEE logistic model with an autoregressive working correlation of order 1 (AR1), treating each participant as a clustering unit and including participant indicators to control participant-level baseline differences. Engagement status (not engaged=0, engaged=1) served as the outcome variable. Given the small number of participants, we interpret inferential statistics as exploratory and emphasize effect sizes with CIs.

Engagement in both the Fruit Sorting Run and Hazard Avoiding Ride tended to increase over normalized time (Figure 8). Using GEE with a logistic link to account for repeated measures within participants (AR1 working correlation; participant indicators included), normalized time was significantly associated with a higher probability of engagement status =1 in the Fruit Sorting Run (per 0.1 increase: β=0.48; odds ratio 1.62, 95% CI 1.09-2.38; P=.02). The Hazard Avoiding Ride showed the same direction of association (per 0.1 increase: β=0.66; odds ratio 1.93, 95% CI 1.04-3.60; P=.04). The results present participant-specific predicted probabilities and a population-averaged trend estimated from GEE logistic models with an AR1 working correlation (95% CI). Consistent with these estimates, model-estimated P(engagement=1) tended to be higher at later normalized timepoints in both activities.

Along with the temporal trends observed during gameplay, caregiver interviews with all 3 participating children with ASD (P1-P3) supported the observed engagement patterns. The interviews revealed distinct yet converging increases in engagement, along with perceptions of each game’s strengths and suggestions for improvement. Feedback was analyzed by game type to understand how each activity influenced engagement. In the Fruit Sorting Run, caregivers reported that participants remained attentive and engaged throughout the session. P1 appeared focused and responded positively to praise, suggesting a sense of achievement. The caregiver noted that adaptive verbal cues could further support engagement when the child was not running. P1 also showed difficulty differentiating fruits with similar colors and shapes (eg, apple vs tomato; oriental melon vs lemon), which sometimes interrupted task progression. P2 expressed a stronger preference for the Fruit Sorting Run over the Hazard Avoiding Ride. Although he showed initial hesitation, he maintained effort throughout the activity, even though he typically disengages when bored. The caregiver noted that the task aligned with his everyday preference for sorting beads by color and type, which may have supported sustained participation. Because communication was challenging, P2 relied primarily on visual prompts. P3, who usually avoids running, remained physically active without breaks and showed enjoyment. His caregiver observed that he took pride in the fruit-sorting task, which may have contributed to sustained engagement. In the Hazard Avoiding Ride, similar patterns of engagement were reported. All 3 children showed attention and engagement patterns that were consistent with an increasing trend over time, despite differences in prior experience. According to P1’s caregiver, the child focused on the task and followed the VA’s instructions, with interest increasing as the session progressed. P2, unfamiliar with bicycling, found the activity novel and remained highly concentrated; the caregiver noted that repeated exposure could promote learning. P3 was initially anxious but became increasingly engaged. Unlike in typical physical or leisure activities—where he often loses interest—he demonstrated more active and enjoyable participation once familiar with the game.

The primary goal of this study was to design full-body serious exergames for children with ASD through an iterative, expert-informed process that emphasized both playfulness and potential functional benefits. As a result, 2 exergames were developed—Fruit Sorting Run and Hazard Avoiding Ride—each designed to promote engagement through physical activity aligned with functional skill development. To achieve this, rather than relying on a single design iteration or a pre-existing game framework, we adopted a multiphase participatory approach that integrated input from special education experts, caregivers, therapists, and HCI researchers across 4 structured phases (Figure 1). This process yielded not only 2 target exergames but also a systematic methodology that highlights the role of stakeholder involvement in tailoring game content to the cognitive, motor, and daily needs of children with ASD.

Three design considerations emerged from this process: (1) grounding gameplay in goal-oriented exercises aligned with the cognitive and motor capabilities of the target population. The exergame conceptualization was grounded in a 2×2 activity matrix informed by therapist categorization of physical and cognitive demands—a structured method that reflects prior evidence showing that tasks calibrated to users’ cognitive and motor levels support skill development and adaptive functioning in children with ASD [47]. This approach reflects SG design principles that promote aligning game tasks with users’ motor and cognitive abilities through intended in-game goals [38]. Recent studies have further shown that exercises aligned with the cognitive and motor abilities of children with ASD not only improve gross motor skills but also help motivate participation among children with ASD [48,49]. (2) Engagement scaffolding mechanisms—such as prompts and feedback from VAs—were incorporated to support sustained interaction. Existing VA-based systems for children with ASD have demonstrated the utility of contingent feedback in guiding user responses. These systems primarily focused on low-level interactions, responding to gaze [24], nonverbal conversational cues [50], or emotional expressions [51]. While such approaches support reciprocal interaction, they typically emphasize momentary cue–response dynamics rather than task-oriented guidance. Our VA expands on these foundations by delivering gameplay tasks directly, combining verbal instructions and visual cues to deliver gameplay tasks. (3) Prioritizing gameplay elements perceived by stakeholders—especially educators and caregivers—as playful, goal-oriented, and compatible with daily routines, the design approach reflects established principles in SG design for children with ASD [38]. Routine exercises that align with daily activities [52], such as jogging and cycling, can promote active engagement of children with ASD [53] and lead to improvements in skills in planning, inhibition, and cognitive flexibility, along with reductions in repetitive behaviors [54].

The pilot field study suggested that engagement trajectories tended to align with an increasing trend over time during gameplay, despite differences among participants in age, school grade, and rapport levels with the VA and occasional distraction or disengagement reflecting common attentional variability in children with ASD; behavioral labeling and caregiver interviews supported this overall pattern. Caregiver interviews indicated that even children with limited prior experience or low baseline motivation (eg, unfamiliar with bicycling [P2], typically avoidant of running, or prone to early disengagement [P3]) became more involved as they grew accustomed to the game structure. Caregivers noted that children responded positively to praise prompting (P1) and showed behavioral responses to antecedent prompting, following the VA’s instructions as interest and task focus increased during gameplay—a pattern similarly emphasized by Bernardini et al [24], who reported that children with ASD showed increased responsiveness to directive prompts delivered by a VA. Caregivers also noted that children with limited expressive communication (P1 and P2) relied more on intuitive, visual prompting; therefore, follow-up work should present options through explicit and intuitive visual media and avoid ambiguous choice boundaries. This design choice aligns with established intervention practices that commonly use visual supports, structured prompting, and reinforcement to sustain task participation [24]. Observations during gameplay further suggested that adaptive prompting, triggered in moments of hesitation or disengagement, could support sustained interaction and re-engagement [55]. Taken together, these findings highlight the VA’s role in supporting engagement through prompting during gameplay.

This study aimed to develop full-body serious exergames for children with ASD through an iterative, expert-informed design process. As an initial exploration, we tested the games with 3 children to examine engagement patterns during gameplay. Although all participants showed an upward trend in engagement, the small sample size can limit the interpretation of these patterns across the wider population of individuals with ASD. Given the heterogeneity in cognitive, motor, and behavioral characteristics, further studies with larger and more varied samples are required to examine whether similar engagement trends appear across individuals. Examining which child characteristics (eg, communication level, sensory sensitivity, motor proficiency, and baseline interest) are associated with engagement trajectories and responsiveness to prompts may be valuable for informing individualized interaction profiles. This direction may support tailoring prompt timing, modality (visual/verbal), and reinforcement style to the needs of each child rather than applying a single prompting approach. Expanding the dataset would also support future work on adaptive prompting strategies [55].

Physiological signals (eg, heart rate and skin conductance) were streamed during gameplay to assist annotators’ contextual interpretation; however, these data were not incorporated into the present analysis. As the study focused on behavioral engagement labeled by caregivers, physiological measures were treated as supplementary rather than analytic components. Future work may integrate these multimodal data to model engagement dynamics more comprehensively. In particular, combining physiological signals with gameplay context (eg, pauses, slowed pace, and errors) may help detect hesitation or overload and trigger context-sensitive prompts.

The engagement labels used a binary scheme, which reduces engagement to an on/off state and may not capture brief fluctuations or multidimensional aspects of engagement. Future work may consider multilevel or continuous labeling, along with complementary measures, to represent engagement dynamics with higher temporal resolution. Such representations can also enable evaluating how different adaptive prompting strategies affect distinct engagement components (eg, attention vs task persistence) across children.

While prompting was delivered through a VA, it was not dynamically tailored to individual behaviors. To enable context-sensitive and engagement-responsive interaction, future studies will aim to recognize engagement status in real time and apply machine learning to model and deliver adaptive prompting, aligned with O’Guinn et al [56]. Building on this direction, future work will extend the current VA-based prompting scheme from a fixed, prespecified schedule to an engagement-responsive strategy. A feasible next step is to evaluate fixed prompting versus adaptive prompting using a single-case experimental design, such as a multiple-baseline design across participants. Baseline sessions would follow the current fixed prompting logic, whereas intervention phases would introduce adaptive prompting that adjusts timing and modality based on predefined behavioral triggers observable in the existing system (eg, sustained disengagement, repeated errors, or prolonged response latency). Engagement trajectories, recovery from disengagement, and prompt responsiveness would be compared within individuals across phases, enabling rigorous individual-level evaluation under substantial interindividual variability in children with ASD.

Finally, although increased engagement was observed, the study did not assess therapeutic or educational outcomes. Future research should investigate whether these exergames can support specific intervention goals, such as improving emotional regulation, reducing stereotyped behaviors, or enhancing functional skills [57]. If such outcomes are demonstrated, the system may also offer practical value to special educators seeking to promote active participation in school-based or clinical settings. To situate the system within the broader intervention landscape, future studies will test deployment across common contexts (school-based sessions, clinical programs, and home practice) and assess feasibility factors relevant to practitioners (eg, setup time, supervision demands, and alignment with individualized goals).

This study presented the iterative design and initial field evaluation of 2 full-body serious exergames for children with ASD, developed through a multiphase, expert-informed process involving 21 professionals from special education, APE, and HCI. Grounded in goal-directed physical activity and supported by engagement scaffolding mechanisms such as VA-based prompting, the exergames were tailored to promote interaction aligned with developmental needs. The design process was shaped by 3 core considerations: aligning gameplay with children’s cognitive and motor capabilities, incorporating scaffolding mechanisms to support engagement, and ensuring playfulness and perceived developmental value, such as supporting daily living skills.

Beyond prior SG approaches that primarily relied on discrete gestures or in-place interactions, this study advances exertion-intensive, whole-body exergame design by embedding sustained physical activity within structured gameplay, shaped through a series of expert collaborations. As an exploratory multiple-case pilot, we tested the 2 developed exergames with 3 children with ASD to examine feasibility and engagement patterns during gameplay, observing increasing engagement trends over time despite inter-individual differences.

By integrating expert-driven design with case-based field evaluation conducted in a special education setting, this work contributes a grounded design and evaluation framework for developing exertion-intensive exergames for children with ASD. This contribution extends prior work on serious exergames for children with ASD by demonstrating how therapist-informed exercises and structured prompting strategies can be systematically translated into playable, full-body game mechanics. While these observations remain preliminary, they provide initial indications that should be examined and strengthened with larger samples. Taken together, this exploratory multiple-case pilot provides an initial basis for a testable hypothesis that full-body serious exergames, paired with VA-based prompting, may be associated with increasing engagement over time, which should be examined in larger samples.