Immersive simulation interventions for children and adolescents with autism spectrum disorder: technologies, cognitive and socio-emotional development, and AI-driven emotion estimation: a systematic review

1 Introduction

Autism spectrum disorder (ASD) is a neurodevelopmental condition affecting an estimated 1 in 127 persons worldwide in 2021 (2021 Global Burden of Disease (GBD), 2024). Children with ASD show impairments across multiple domains, including social communication and interaction, motor abilities, attention, and emotion recognition (American Psychiatric Association, 2013). These difficulties can affect their self-concept, interpersonal relationships, and capacity for independent living (Martí-Vilar et al., 2023). This unique combination of characteristics makes each autistic individual distinct and sometimes challenging to understand and support. These differences significantly influence their social participation (Taheri et al., 2016) and affect the daily lives of their families (Karst and Van Hecke, 2012; Tonkin et al., 2014). We believe that in an inclusive society, all children and their families should have access to the same activities and opportunities to participate and interact with their community members. Although social participation and leisure activities are recognized as fundamental rights, they also contribute significantly to the well-being of children and their families (Abbott, 2006; Coyl-Shepherd and Hanlon, 2013; Menear and Neumeier, 2015). Conversely, limited participation increases the risk of additional social and psychological challenges for autistic individuals (Kawachi and Berkman, 2001). To promote an inclusive society, community activities should engage all members, including families, while supporting the development of individual competencies that enable autistic children to participate actively, meet their sensory, emotional, and social needs, and enhance communication, emotional regulation, and social skills.

According to Mosher and Carreon (2021), immersive technology is an effective tool for encouraging autistic students to learn and develop new skills. A variety of immersive technologies, such as augmented reality (AR), virtual reality (VR), mixed reality (MR), and extended reality (XR), can promote skill development and social engagement. They are effective mainly because they can create immersion. Immersion refers to how real the experience feels, which leads to a strong sense of presence. This presence is a mental and emotional state that makes the user feel as if they are truly inside the virtual environment (Slater et al., 1996). Immersive technologies encompass several distinct modalities that can be positioned along Milgram and Kishino’s (1994) reality–virtuality continuum:

• Virtual reality (VR): Fully or predominantly computer-generated environments that largely replace the user’s perception of the physical world, providing a high level of immersion and presence.
• Augmented reality (AR): Digital information or virtual objects are overlaid onto the real-world environment while maintaining the user’s awareness of and interaction with the surrounding physical space (Van Krevelen and Poelman, 2010).
• Mixed reality (MR): Physical and virtual objects coexist and interact dynamically in real time, enabling users to engage simultaneously with both real and digital elements within a shared environment (Rebbani et al., 2021).
• Extended reality (XR) or xReality: An umbrella term encompassing VR, AR, MR, and related immersive technologies that extend human perception and interaction beyond conventional physical environments (Khan, 2023). XR represents a broad conceptual framework rather than a distinct technology itself. Some authors further extend this concept through terms such as *R or All Reality to describe a wider spectrum of present and future reality-mediating technologies (Khan, 2023).

VR technologies vary in how deeply they immerse the user, supported by work showing that immersion ranges from basic screen-based systems to fully surrounding virtual environments (Beck et al., 2019). At the lowest level, non-immersive VR relies on standard computer screens with limited sensory engagement. Semi-immersive setups, such as large projections or multi-screen displays, enhance presence through wider visual fields (Castronovo et al., 2013). Fully immersive VR (IVR), most notably head-mounted displays, delivers synchronized stereoscopic images that respond to head movement, creating a compelling 3D experience (Wenk et al., 2023). These systems require real-time, fluid rendering to maintain comfort and prevent issues such as motion sickness, underscoring the balance between visual realism and performance.

In recent years, IVR has shown genuinely promising potential for supporting children and adolescents with ASD, especially in the area of social communication. A growing body of systematic reviews and meta-analyses reinforces this trend. For example, Mittal et al. (2024) reported notable gains in social skills (SMD = 1.43; 95% CI: 0.01–2.84), emotional understanding (SMD = 2.45; 95% CI: 0.21–4.18), and broader cognitive abilities following VR-based interventions. Similarly, Yang et al. (2025), who reviewed 14 studies, found consistently positive effects, with high-functioning autistic individuals appearing to benefit the most from immersive environments. Not only social skills (Cheng et al., 2015; Ip et al., 2018; Lorenzo et al., 2013; Matsentidou and Poullis, 2014) but also related studies show the potential of IVR interventions for ASD in various areas, such as emotion recognition (Ip et al., 2018; Lorenzo et al., 2013), daily living skills (Adjorlu et al., 2017), vocational skills (Bozgeyikli et al., 2017), and phobias (Maskey et al., 2014; Maskey et al., 2019). Overall, these findings emphasize that VR can serve as a valuable and evidence-based tool for improving essential developmental skills in people with ASD.

Several systematic reviews have examined VR-based interventions for children and adolescents with ASD over the past decade. Mesa-Gresa et al. (2018) provided one of the earliest evidence-based reviews, integrating clinical and technical databases and reporting moderate support for VR treatments, although they did not differentiate immersive from non-immersive technologies or focus specifically on children and adolescents. They combined clinical and some technical databases, but they did not include Scopus (the largest multidisciplinary database for tracking citations) or IEEE Xplore (the primary source for VR/AR/MR engineering-focused studies). Pivotto et al. (2024) explored cave automatic virtual environment (CAVE)-based immersive environments, highlighting benefits such as multi-participant engagement, sensory comfort, and social skills development, but their review was limited to CAVE systems and recreational applications, without assessing other immersive technologies or broader outcomes. They conducted a broad search across seven databases, including APA PsycINFO, ERIC, ProQuest Advanced Technologies, ProQuest Dissertations and Theses Global, PubMed, Scopus, and Web of Science; however, they did not specifically include IEEE Xplore.

Li et al. (2023) conducted a systematic review and meta-analysis of social and affective skill training, summarizing intervention effectiveness while noting limited theoretical adaptation for VR, but they did not examine the design features, technological affordances, or alignment with participant characteristics. They focused on clinical and biomedical databases such as PubMed, Scopus, Cochrane, and Embase; however, they did not include IEEE Xplore or PsycINFO (psychology/behavior-focused research relevant to ASD interventions). Gandall and Sanchez Peral (2025) focused on self-regulation and interoceptive skills in children with ASD, emphasizing human factors in VR design and did not consider broader skill domains, strict inclusion criteria, and comparator groups. Their reviews searched PubMed, PsycINFO, and Google Scholar to capture clinical and psychological studies; they did not include Scopus, Web of Science, or IEEE Xplore, limiting the coverage of highly cited, interdisciplinary, and technology-focused research.

Han et al. (2025) synthesized VR applications for social communication disorders, including ASD, examining usability, gaze tracking, and inclusive education. Immersive technology types and intervention design features were not systematically investigated. They searched through PubMed, Scopus, ERIC, CINAHL, and PsycINFO to cover biomedical, educational, and psychological studies but did not include IEEE Xplore or Web of Science. Finally, the most recent study by Azman et al. (2025) reviewed VR-based social skills training for high-functioning ASD, reporting gains in social interaction, cognition, and nonverbal communication. The review was limited by the small number of studies and lack of analysis of technological features or comparator conditions. Their systematic review searched PubMed, Cochrane, Scopus, and Embase to capture clinical and biomedical studies but did not include IEEE Xplore or PsycINFO.

Collectively, these reviews show the potential of VR interventions for children and adolescents with ASD. However, important gaps remain. These include the need for a systematic inclusion of various immersive technologies, a detailed examination of technological and sensory features, complete reporting of intervention design, careful characterization of participant profiles, and consideration of suitable comparison groups. To address this gap, this study plans to conduct a systematic review using multiple high-quality databases. These include PubMed, Scopus, Web of Science, PsycINFO, and IEEE Xplore. The goal is to ensure thorough coverage of both health-related outcomes and technological advancements in VR, AR, MR, and XR interventions for children and adolescents with ASD.

2 Methods

2.1 Registration and reporting

This systematic review was registered in PROSPERO (CRD420251147405) on 06 October 2025. This step aimed to ensure transparency, avoid duplication, and outline the review methods. The review followed the PRISMA 2020 guidelines for systematic reviews (Page et al., 2021). Because the outcomes varied across studies, this review uses a narrative synthesis instead of a meta-analysis. Only studies that evaluated immersive technology-based interventions for children and adolescents with ASD were included, which guided the selection, extraction, and synthesis processes. For this review, an immersive technology intervention was defined as any digital system that creates, augments, or blends virtual content with the user’s environment and enables interactive engagement through visual, auditory, haptic, motion tracking, or multisensory interfaces. Accordingly, eligible interventions included VR, AR, MR, XR, CAVE/projection-based systems, haptic/sensor-enhanced environments, and hybrid immersive platforms. Conventional computer-based programs lacking immersive or interactive environmental features were excluded.

2.2 Defining the research questions

Defining clear research questions is essential for guiding study selection and ensuring that relevant evidence is captured. In this review, we used the PICOT framework to organize the research questions (Riva et al., 2012). This framework focuses on population (P), intervention (I), comparison (C), outcome (O), and time/type (T). Because immersive technologies in autism interventions are still developing, we placed less emphasis on outcomes. Many studies are pilot or short-term exposure trials where significant clinical effects may not be expected. The PICOT components for this review were defined as follows:

• P (Population): Children and adolescents with autism spectrum disorder (ASD), aged ≤19 years.
• I (Intervention): Immersive technology-based interventions, including VR, AR, MR, or XR.
• C (Comparator): Pre–post designs, or alternative therapies. Comparisons were included when reported.
• O (Outcome): Studies were eligible if they reported any intervention-related outcome relevant to social, emotional, cognitive, behavioral, motor, adaptive, participation, usability, engagement, safety, or feasibility domains. Eligibility was not restricted to studies reporting statistically significant or positive effects; null, mixed, or negative findings were retained when they met the population, intervention, and study-type criteria.
• T (time/type): Studies were considered irrespective of intervention duration, reflecting the exploratory and pilot nature of most immersive technology research.

Table 1 presents the alignment of each research question with the corresponding PICOT component that guided the focus and scope of the review.

TABLE 1. Mapping research questions to PICOT components.

2.3 Literature search strategy and information source

A comprehensive search strategy was developed and refined to identify studies. The search encompassed five major electronic databases: Scopus (Elsevier), Web of Science (Clarivate), PubMed, IEEE Xplore Digital Library, and PsycINFO (via ProQuest), covered the period from 2015 to 2025, and was restricted to English language publications. The search strategy combined controlled vocabulary terms and free-text keywords related to autism, immersive technologies, and participant characteristics. The core search string, adapted for each database, is presented in Table 2. The complete search strategy, including all keywords, Boolean operators, and database-specific adaptations, is provided in Supplementary Appendix A. The database search initially retrieved 3,332 records: Scopus (n = 1,880), Web of Science (n = 1,109), IEEE Xplore Digital Library (n = 248), PubMed (n = 76), and PsycINFO (n = 19). After removing duplicates, early access articles, book chapters, retracted publications, and review articles, 1,984 records remained for title and abstract screening.

TABLE 2. Core search string for literature search.

2.4 Inclusion and exclusion criteria

The selection of studies for this review followed a structured pre-screening and full-text screening process, guided by a predefined framework (see Supplementary Appendix B). Three independent reviewers assessed all records retrieved from the database search. A study was included in the review if all three reviewers agreed that it met the inclusion criteria; otherwise, it was excluded. Anonymous decisions were coded as Include (1) or Exclude (0). If there was any uncertainty, the study was advanced to the full-text screening phase for further evaluation. Pairwise Cohen’s κ values indicated substantial agreement between R1 and R2 (κ = 0.78), moderate agreement between R2 and R3 (κ = 0.54), and fair agreement between R1 and R3 (κ = 0.39). After removal of duplicate records and other ineligible publication types, the 1,984 remaining records underwent title and abstract screening according to the predefined inclusion and exclusion criteria. This process yielded 581 potentially relevant records for further evaluation. Among these, 354 records were excluded because they did not meet the eligibility criteria, including studies that did not involve immersive technologies (n = 146), had sample sizes smaller than five or missing sample information (n = 92), involved adult-only samples or did not report age (n = 54), did not include participants with ASD (n = 17), had ineligible publication types (n = 24), or were excluded for other reasons such as non-English language or absence of participants (n = 21). Consequently, 227 records were considered potentially eligible and full texts were sought for retrieval. Of these, 86 reports could not be retrieved, leaving 141 articles available for full-text assessment. All retrieved full-text articles satisfied the predefined eligibility criteria and were therefore included in the qualitative synthesis. A full visual representation of the study selection process is shown in the PRISMA 2020 flow diagram (Figure 1).

Disagreements during full-text assessment were resolved through discussion and consensus among the three reviewers using the predefined eligibility criteria. When eligibility was unclear during title and abstract screening, studies were retained for full-text review to minimize the risk of erroneous exclusion.

Studies including fewer than five participants with ASD were excluded to enhance methodological comparability across studies and to support coherent synthesis of group-level intervention outcomes. Very small studies and single-case experimental designs (SCEDs) often employ distinct analytical frameworks that are not directly comparable with group-based intervention studies. While this criterion improved consistency in the synthesis process, it may have reduced the inclusion of early-stage feasibility studies and SCED evidence, which are common in emerging immersive technology research for ASD.

2.5 Data extraction

A structured data extraction framework was created to ensure consistency, completeness, and a careful approach to capturing information from all eligible studies (see Supplementary Appendix C for the full template). The extraction sheet was revised several times and tested on a small group of included articles to check for clarity and fit with the review objectives. Two reviewers extracted the data independently, and any discrepancies were resolved through discussion. Cases needing further review were examined together to reach an agreement.

The data extraction template included study-level and intervention-level variables. These variables covered bibliographic information, such as source, title, country, and publication year. They also included participant characteristics, such as sample size, aggregated ASD sample, age range, diagnostic method, and comorbidities. The template included intervention features, detailing the type of immersive technology, devices used, software or simulation environment, sensory modalities, and domain or scenario. For classification purposes, VR referred to fully or partially computer-generated environments experienced through immersive or non-immersive interfaces; AR referred to digital elements overlaid onto the real world; MR referred to environments in which physical and virtual objects coexist and interact in real time; and XR was treated as an umbrella term encompassing VR, AR, MR, and related immersive systems.

Additional fields captured study design, comparator conditions when applicable, targeted skills, outcome measures, and reported effects or significance values. We recorded indicators of technological innovation, such as AI-driven features, adaptive feedback systems, or eye tracking. For the purposes of this review, AI-enabled components were defined broadly as computational approaches that supported automated inference, adaptation, personalization, or behavioral analysis within immersive environments. These components included machine learning and emotion-recognition algorithms, multimodal data fusion, predictive models, and adaptive feedback mechanisms informed by data obtained from eye tracking, motion capture, depth sensing, or related technologies. Accordingly, AI-enabled interventions encompassed both explicit machine learning applications and intelligent adaptive systems that dynamically modified intervention delivery based on user behavior and performance.

To support understanding of methodological quality and practical implications, the extraction framework included fields for reporting impacts or benefits, challenges, limitations, safety considerations (such as cybersickness and overstimulation), and the authors’ suggested future research directions. Lastly, each study was evaluated for methodological quality and risk of bias (RoB) using the Mixed Methods Appraisal Tool (MMAT), and the corresponding rating was noted in the extraction sheet.

2.6 Quality assessment

2.6.1 Risk of bias assessment

The methodological quality of included studies was systematically evaluated using a quantitative risk of bias (RoB) assessment adapted from Wiebe et al. (2022). Each study was assessed across five key domains: (1) presence of a control or comparator group, (2) randomization or allocation procedures, (3) inclusion of pre/post measurements or follow-up assessments, (4) blinding where applicable, and (5) study registration. The overall risk-of-bias score was calculated as the percentage of unfulfilled criteria. Studies with ≤25% unfulfilled criteria (0–1 unmet items) were classified as having a low risk of bias, those with 26%–50% unfulfilled criteria (2 unmet items) as having a moderate risk of bias, and those with >50% unfulfilled criteria (3–5 unmet items) as having a high risk of bias. A summary of the overall RoB across studies is presented in Figure 8, while detailed study-level assessments, including the five assessment criteria, score calculation procedure, calculated percentages, and thresholds for low-risk, moderate-risk, and high-risk classifications, are provided in Supplementary Appendix D.

2.6.2 Mixed methods appraisal tool

The studies were assessed for their methodological quality using the Mixed Methods Appraisal Tool (MMAT), version 2018 (Hong et al., 2018). This tool is validated for evaluating the quality of qualitative, quantitative, and mixed-methods studies. Each study was classified based on its design type: qualitative, quantitative randomized controlled trial (RCT), quantitative non-randomized, quantitative descriptive, or mixed methods. They were then evaluated using the five relevant criteria. Individual study MMAT scores are summarized in Figure 9, and the complete list of MMAT questions, organized by study design, is provided in Supplementary Appendix E. This assessment allowed for systematic, comparable evaluation of methodological quality across diverse study designs and informed the interpretation of results regarding the robustness of the findings.

3 Results

3.1 Temporal trends and geographic distribution of studies

Figure 2 illustrates the temporal distribution of publications, showing a steady increase in research activity over the past decade. Notably, the publication peaks occurred in 2022 and 2024 (n = 20 each), followed by 2020 (n = 19) and 2019 and 2023 (n = 13 each), reflecting growing academic interest in this area over time.

In terms of geographic distribution, the included studies represent diverse regions globally (Figure 3). North America accounted for 31.9% of studies, with the United States contributing the largest share (n = 37), followed by Canada (n = 5). Asia contributed 33.3% of studies, with China (n = 11), India (n = 8), and Japan (n = 2) as prominent contributors. Europe represented 29.7% of studies, with Spain (n = 15) leading, followed by France (n = 3) and Germany (n = 1). Other regions were less represented, with South America at 2.2%, Africa at 1.4%, and Oceania at 1.4%. The country-level distribution and a world map illustrating research contributions are presented in Figure 3, highlighting global engagement while also indicating underrepresented regions.

3.2 Participant characteristics (RQ1)

The studies included in this review varied widely in sample size. Most were small-scale studies: Of the 141 included studies, 48 studies (34.0%) included fewer than 10 participants, and 66 studies (46.8%) enrolled between 10 and 29 participants. Medium-sized studies (30–59 participants) accounted for 14 studies (9.9%), while 9 studies (6.4%) were large (60–99 participants). Only 4 studies (2.8%) included very large samples of 100 participants or more, as shown in Figure 4.

The 141 studies included in this review involved approximately 3,052 participants with ASD and approximately 323 typically developing (TD) participants for comparison in selected studies. Participants ranged in age from 2.5 years to 19 years, with most studies focusing on children and adolescents aged 6–18 years. Sex distribution was inconsistently reported. Among the 16 studies that provided sufficient sex-specific participant data, 188 participants were male, and 31 were female.

Participants were primarily diagnosed with standardized clinical tools, including Diagnostic and Statistical Manual of Mental Disorders (DSM- 5 or DSM-IV) criteria (Frolli et al., 2022; Herrero and Lorenzo, 2020; Jeyarani et al., 2024; Maskey et al., 2019; Samantaray et al., 2024; Zhang et al., 2020), the Autism Diagnostic Observation Schedule (ADOS) (Fan et al., 2015; Koirala et al., 2019; Koirala 2021; Wade et al., 2016; Zhang et al., 2020), and its updated version ADOS-2 (Boo et al., 2022; Frolli et al., 2021; Frolli et al., 2022; Zhao H. et al., 2021; Zhao et al., 2018a), the Social Communication Questionnaire (SCQ) (Babu et al., 2022; Porayska-Pomsta et al., 2018; Samantaray et al., 2023; Samantaray et al., 2024; Zhao et al., 2016), the Social Responsiveness Scale Second Edition (SRS-2) (Abd El-Sattar, 2025; Keshav et al., 2019; Samantaray et al., 2024; Zhang et al., 2017; Zhao et al., 2016; Zhao J. Q. et al., 2021), and the Childhood Autism Rating Scale (CARS) (Kouhbanani et al., 2021; Krishnappa Babu and Lahiri, 2020; Narayanan, 2019; Samantaray et al., 2024; Zhao J. Q. et al., 2021), with most classified as high-functioning. Some studies included participants with co-occurring conditions such as Attention-Deficit/Hyperactivity Disorder (ADHD) (Ghanouni et al., 2021; Ide-Okochi et al., 2024; López-Bouzas et al., 2024; Maskey et al., 2019), intellectual disability (Bauer et al., 2023; Oki et al., 2025), or language impairments (Miller et al., 2020). In studies with TD comparison groups, ASD:TD ratios were approximately 1:1.2–1.5. In studies including TD comparison groups, ASD:TD ratios were approximately 1:1.2–1.5.

Reporting of comorbidities across the 141 studies was highly inconsistent. Many studies did not explicitly report co-occurring conditions or noted them only generally under neurodevelopmental disorders (NDD). When specified, common comorbidities include ADHD, intellectual disability, anxiety disorders (Ip et al., 2016), learning difficulties (Porayska-Pomsta et al., 2018), dyspraxia (Maskey et al., 2019), dyslexia (Maskey et al., 2019), tic disorder (Intawong et al., 2025), cerebral palsy (Intawong et al., 2025), obsessive-compulsive disorder (OCD) (Carreon et al., 2024b; Nazari et al., 2024), and sensory processing disorders. Several studies excluded participants with specific comorbidities such as seizures, genetic syndromes, cardiovascular conditions, or severe sensory impairments. The reported prevalence of comorbidities varied, with ADHD present in approximately 20%–35% of participants in studies that assessed it, and intellectual disability noted in 30%–70% of participants in certain samples.

3.3 Technologies and simulation types (RQ2–RQ3)

The included studies employed a wide range of immersive technologies, varying in both hardware and software configurations. Table 3 provides a detailed overview of the technology groups, typical devices, software or simulation environments, and the distribution of studies across these categories. Most studies (78 studies; 55.3%) utilized VR, including IVR, head-mounted displays (HMDs), mobile VR, and driving simulators. Common devices include Oculus Quest 2, HTC Vive, Google Cardboard, and custom HMD setups, while software platforms often involve Unity3D, Unreal Engine, Steam VR, Vizard, or custom VR engines.

TABLE 3. Summary of technology groups, devices, and simulation environments used in the included studies.

AR technologies were used in 18 studies (12.8%), typically employing HoloLens, Google Glass, tablets, or smartphones, with software frameworks such as Unity3D, Vuforia, and custom mobile AR applications. MR, reported in 12 studies (8.5%), involved full-body interactions or hybrid AR/VR systems, often leveraging HoloLens 2, Vive Pro Eye, and motion capture trackers. MR applications were developed using Unity3D, Unreal Engine, Mixed Reality Toolkit (MRTK), or custom MR platforms.

Collaborative virtual environments (CVEs) or collaborative VR setups were used in 15 studies (10.6%), facilitating multi-user interactions and often incorporating haptic devices and networked VR platforms. Unity3D with Photon networking and custom collaborative applications were commonly employed. Haptic or sensor-enhanced VR systems were reported in nine studies (6.4%), integrating haptic interfaces, instrumented treadmills, or eye tracking devices, with specialized software such as Unity3D haptic plug-ins or MATLAB-based analyses. Projection-based and CAVE systems, including immersive projection rooms or “blue room” setups, were implemented in six studies (4.3%), with multi-wall or floor projections and motion capture integration, typically supported by Vizard or custom CAVE software. A small subset of studies (three studies; 2.1%) used other or hybrid technologies, including cross-reality platforms, serious games, metaverse applications, or virtual learning environments, combining multiple devices and custom software solutions.

3.4 AI-enabled immersive system components (RQ4)

A growing number of immersive technology interventions integrated AI-driven software modules, particularly for emotion recognition and automated behavioral analysis. These systems used facial-expression classification (Lorenzo et al., 2016), voice-based affect detection (Li et al., 2015), or multimodal fusion (Bian et al., 2015; Krishnappa Babu and Lahiri, 2020; Meng and Yeh, 2022; Moon et al., 2024; Nalepka et al., 2022; Zhang et al., 2017) to infer a participant’s emotional state. Other interventions incorporated adaptive algorithms, which dynamically modified task difficulty, pacing, or instruction style based on real-time performance metrics (Amat et al., 2021; Hocking et al., 2022; Simões et al., 2018; Zhao et al., 2017). In some cases, machine learning models were embedded within the VR application to personalize training pathways, increase responsiveness, and enhance the ecological validity of social skills scenarios (Alcañiz et al., 2022; Alcaniz Raya et al., 2020; Dhamodharan et al., 2020; Ip et al., 2016; Jeyarani et al., 2024; Lu et al., 2018; Meng and Yeh, 2022; Moon et al., 2024; Nalepka et al., 2022; Schmidt et al., 2023; Tran, 2024; Zhang et al., 2017; Zhao et al., 2018a).

Several interventions employed specialized hardware to support intelligent, AI-enabled interaction. Eye tracking sensors (e.g., Tobii eye trackers) were used to capture gaze behavior, joint attention, and visual focus patterns during VR tasks, providing real-time data for both assessment and adaptive feedback (Amat et al., 2021; Babu et al., 2022; Bozgeyikli et al., 2018; Carreon et al., 2024a; Cecil et al., 2017; Elkin et al., 2022; Garzotto et al., 2024; Ide-Okochi et al., 2024; Li et al., 2024; Lu et al., 2018; Malihi et al., 2020a; Minissi et al., 2023; Tokarskaya et al., 2020; Wade et al., 2015; Wade et al., 2016; Zhang et al., 2015). In addition, motion-capture cameras, depth sensors (e.g., Kinect-style systems), and inertial tracking units enabled fine-grained monitoring of head, hand, and body movements (Gali et al., 2024; Liu et al., 2024; Lu et al., 2018; Meng and Yeh, 2022; Minissi et al., 2023; Oki et al., 2025; Romadlon Junaidi et al., 2022; Sayis et al., 2019; Sayis et al., 2020; Tentori et al., 2015). These tracking modalities allowed the systems to analyze user behavior continuously and feed data into AI models for emotion detection, engagement estimation, and performance prediction.

Advanced immersive simulations incorporated features that enabled more naturalistic, socially meaningful engagement. Avatar-based emotion mirroring allowed virtual characters to change facial expressions or gestures in response to the participant’s interpreted emotional state, powered by underlying AI recognition modules (Bauer et al., 2023; Liu et al., 2024; Tentori et al., 2015; Yeh and Meng, 2025). Some systems also embedded behavioral cueing, such as gaze direction changes, proximity adjustments, and real-time social prompts, to guide users toward appropriate responses (Adjorlu et al., 2017; Al-Saddi et al., 2023; Boyd et al., 2018; Cao et al., 2025; Herrero and Lorenzo, 2020; Intawong et al., 2025; Shahab et al., 2017; Zhang et al., 2020; Zhao et al., 2022). These interaction elements enhanced realism and created opportunities for practicing complex social behaviors within controlled, safe immersive environments.

Several studies used multisensory design to support personalized therapy. Visual, auditory, and sometimes tactile cues were delivered in coordination with AI signals. For instance, they provided corrective prompts when attempts to recognize emotions failed. They also increased supportive cues when disengagement was detected through gaze or movement patterns (Boyd et al., 2018; Carreon et al., 2024a; Gelsomini, 2016; Ip et al., 2020; Peng et al., 2019; Sanku et al., 2023; Shahab et al., 2017; Xu et al., 2026; Yan et al., 2025; Zhao et al., 2017). These feedback systems, guided by AI analysis, allow interventions to adjust the stimulus strength, interaction timing, or reinforcement methods to fit each person’s needs.

3.5 Types of control groups and comparator conditions (RQ5)

We examined the types of control groups and comparator conditions used across the studies in our review to answer RQ5. The studies demonstrated considerable variation in design, ranging from no-control single-arm studies to active comparators and typically developing peer comparisons. Figure 5 illustrates the distribution of control groups and comparator conditions across the studies, highlighting the predominance of non-control/pre–post designs and the relative frequency of other comparator types.

Of the 141 included studies, the largest proportion (32.6%; n = 46) did not include a formal control group. These studies typically employed single-arm pre–post or feasibility designs (Thomsen and Adjorlu, 2021; Jeyarani et al., 2024; Krishnan et al., 2024; Fan et al., 2015; Ip et al., 2020; Schwarze et al., 2019; Gelsomini, 2016; Intawong et al., 2025; Ghanouni et al., 2021; Yang and Lee, 2022; Moon et al., 2024; Ide-Okochi et al., 2024; Krishnappa Babu and Lahiri, 2020; Ju et al., 2024; Boyd et al., 2022; Zhang et al., 2017; Dhamodharan et al., 2020; Zhang et al., 2015; López-Bouzas et al., 2024; Akhtar and Guha, 2019; Boo et al., 2022; Zhao et al., 2017; Amat et al., 2021; Liu et al., 2021; Keshav et al., 2019; Zhao J. Q. et al., 2021; Tzanavari et al., 2015; Garzotto et al., 2024; Sanku et al., 2023; Bauer et al., 2022; Meng and Yeh, 2022; Alcañiz et al., 2022; Abd El-Sattar, 2025; Bauer et al., 2023; De Luca et al., 2021; Lu et al., 2018; Johnston et al., 2019; Crowell et al., 2020; Nalepka et al., 2022; Yeh and Meng, 2020; McCleery et al., 2020; Wang et al., 2021; Weiss et al., 2015; Parkinson et al., 2023; Junaidi et al., 2021; Minissi et al., 2023; Schmidt et al., 2023; Romadlon Junaidi et al., 2022; Newbutt et al., 2020; Miller et al., 2020; Boyd et al., 2018; Garzotto et al., 2016). Outcomes generally indicate improvements relative to baseline, although causal conclusions are limited due to the lack of comparison conditions.

Several studies (7.8%) (Tokarskaya et al., 2020; Zhang et al., 2020; Wedyan et al., 2021; Frolli et al., 2022; Malihi et al., 2020a; Carreon et al., 2024a; Junaidi et al., 2020; Nazari et al., 2024; Al-Saddi et al., 2023; Naranjo et al., 2017; Falivene et al., 2025) employed within-subject comparisons, including pre–post tests, historical baseline comparisons, or crossover designs. These designs allowed partial control for individual differences and consistently demonstrated gains in targeted skills when participants were exposed to the interventions.

Active comparator designs were used in 9.2% of the studies (compared two or more intervention methods) (Wade et al., 2016; Wade et al., 2015; Frolli et al., 2021; Alabood et al., 2023; Kouhbanani et al., 2021; Cao et al., 2025; Gali et al., 2024; Bozgeyikli et al., 2018; Hocking et al., 2022; Li et al., 2025; Liu et al., 2024; Oki et al., 2025; Moya-Albor et al., 2018) that included active comparators, such as LEGO therapy, therapist-mediated interventions, immersive versus non-immersive VR, or performance-based versus gaze-contingent tasks.

Typically developing (TD) or neurotypical (NT) control groups were included in 22.0% of the studies (Koirala et al., 2021; Koirala et al., 2019; Samantaray et al., 2024; Porayska-Pomsta et al., 2018; Sayis et al., 2020; Narayanan, 2019; Carreon et al., 2024b; Sayis et al., 2019; Lorenzo et al., 2016; Wongwan et al., 2025; Samantaray et al., 2023; Gali-Perez et al., 2021; Yan et al., 2025; Lee and Yang, 2025; Liu et al., 2022; Cecil et al., 2017; Elkin et al., 2022; Al Kheraif et al., 2024; Zhao et al., 2022; Alcaniz Raya et al., 2020; Babu et al., 2022; Bian et al., 2015; El-Gohary et al., 2022; Shahab et al., 2017; López-Belmonte et al., 2022; Dow et al., 2024; Li et al., 2015; Shahab et al., 2022; Silva et al., 2025; Fallea et al., 2025; Simões et al., 2018). In these comparisons, participants with ASD consistently showed lower baseline performance than TD peers. After intervention, most studies reported measurable improvements in the ASD groups.

Waitlist, treatment-as-usual, or other passive control conditions were employed in 3.5% of the reviewed studies (Maskey et al., 2019; Zhao et al., 2018a; Biffi et al., 2018; Malihi et al., 2020b; Ip et al., 2016). Across these comparisons, intervention groups consistently demonstrated superior outcomes relative to passive controls, suggesting added benefits beyond routine care or no-intervention conditions.

Technical or device-based comparator conditions were also used in 3.5% of the studies (Boo et al., 2022; Ip et al., 2018; Johnston et al., 2022; Zhao et al., 2018a; Almazaydeh et al., 2022), focusing on differences in hardware or software configurations, such as head-mounted display types or VR rendering mechanisms. Reported outcomes primarily reflected variations in engagement, usability, or system performance, rather than substantive differences in core skill acquisition.

3.6 Objectives and outcomes (RQ6–RQ7)

The therapeutic objectives of the reviewed immersive technology interventions were diverse, addressing a broad range of developmental, functional, and clinical targets. As illustrated in Figure 6, the primary focus of the reviewed studies was social and emotional skills, which represented 31.9% of the total sample. These were followed by interventions targeting cognitive and educational skills (19.9%), daily living and vocational skills (15.6%), motor and physical skills (14.2%), driving and navigation skills (9.9%), and sensory, phobia-related, and therapeutic regulation outcomes (8.5%). Several studies addressed more than one therapeutic target.

When we consider each objective domain in more detail, a few consistent patterns emerge, which are described in the following subsections.

3.6.1 Social and emotional skills

The largest proportion of studies focused on social and emotional skill development, including emotion recognition, emotional regulation, joint attention, and social interaction. These interventions commonly employed virtual classrooms to practice socially appropriate responses, turn-taking, and interaction with teachers and peers (Thomsen and Adjorlu, 2021). Other studies used playground environments to support cooperative play, social initiation, and rule-based interaction in semi-structured social contexts (Koirala et al., 2021). Peer interaction scenarios, such as birthday parties and group activities, were frequently used to train conversational reciprocity, eye contact, and perspective-taking (Schwarze et al., 2019). Reported outcomes included improvements in emotion identification, increased social engagement, and greater confidence in social situations, with some evidence of generalization to real-world settings (Ju et al., 2024).

3.6.2 Cognitive and educational skills

A substantial proportion of the reviewed literature targeted cognitive and educational outcomes, focusing on attention, executive functioning, and academic learning. Many studies implemented virtual classroom settings designed to enhance sustained attention, task persistence, and response inhibition during instructional activities (Boo et al., 2022). Educational immersive technology applications were also used to teach literacy, numeracy, STEM-related concepts, and spatial cognition through interactive and adaptive learning tasks (Amat et al., 2021). Outcomes frequently reported improvements in attentional control, task accuracy, and learning motivation, suggesting that immersive environments may support cognitive engagement more effectively than traditional instructional approaches (Gali-Perez et al., 2021).

3.6.3 Daily living and vocational skills

Daily living and vocational skills constituted 15.6% of the reviewed studies. These interventions primarily aimed to promote independence in everyday activities by simulating realistic environments such as supermarkets, kitchens, and public transport systems (Sanku et al., 2023). Shopping scenarios were widely used to train item selection, money management, and sequential task execution (Cecil et al., 2017). Vocational-focused studies addressed workplace routines, task completion, and functional communication within employment-related contexts (Al-Saddi et al., 2023). Reported outcomes included increased task autonomy, improved procedural accuracy, and reduced reliance on caregiver support (Lu et al., 2018).

3.6.4 Motor and physical skills

Interventions targeting motor and physical skills accounted for 14.2% of the reviewed studies. These studies focused on improving gross motor coordination, balance, gait, and fine motor control through motion-tracked VR activities (Babu et al., 2022). Gross motor training commonly involved whole-body movements, balance tasks, and rhythmic exercises (Bian et al., 2015), while fine motor interventions targeted precision, grip strength, and hand–eye coordination using object-manipulation tasks (Wang et al., 2021). Outcomes consistently indicated improvements in motor accuracy, postural stability, and engagement in physical activity (Shahab et al., 2017).

3.6.5 Driving and navigation skills

Driving and navigation skills were addressed in 9.9% of the studies, primarily involving adolescents and young adults. These interventions aimed to improve spatial awareness, route planning, hazard perception, and adherence to traffic rules (Junaidi et al., 2021). Virtual driving simulators were frequently used to train lane maintenance, speed regulation, and decision-making under varying traffic conditions (Moya-Albor et al., 2018). Navigation-focused studies addressed pedestrian safety and public transport use within urban environments (Romadlon Junaidi et al., 2022). Outcomes suggested improvements in navigation accuracy, situational awareness, and preparedness for real-world travel (Yuan and Ip, 2018).

3.6.6 Sensory, phobia, and therapeutic regulation

A smaller subset of studies (8.5%) focused on sensory processing, phobia reduction, and emotional regulation. These interventions commonly employed graded exposure within controlled virtual environments to reduce anxiety and sensory hypersensitivity (Ip et al., 2016). Examples included exposure to feared sounds, crowded spaces, or medical procedures (Almazaydeh et al., 2022). Outcomes generally indicated reduced anxiety responses, improved sensory tolerance, and increased participation in previously avoided situations, highlighting VR’s potential as a therapeutic regulation tool (Garzotto et al., 2016).

3.7 Study designs and research gaps (RQ8)

As illustrated in Figure 7, most reviewed studies employed quantitative non-randomized designs (41.1%), followed by quantitative descriptive approaches (33.3%). Randomized controlled trials accounted for 15.6% of the studies, while mixed methods (6.4%) and purely qualitative designs (3.5%) were comparatively less common. This distribution highlights a predominant reliance on non-experimental and observational methodologies, with rigorous experimental designs remaining relatively limited across the field.

The reliance on small-scale and non-randomized designs presents a critical research gap, as many studies are underpowered and involve feasibility or single-group evaluations. Such methodological constraints limit the generalizability of findings and reduce confidence in the reported effectiveness of VR interventions across the autism spectrum, particularly for lower-functioning individuals (Koirala et al., 2021; Schwarze et al., 2019; Tran, 2024), female participants (Frolli et al., 2021; López-Bouzas et al., 2024; Tokarskaya et al., 2020), and those with co-occurring conditions (Frolli et al., 2022; Salimi et al., 2024; Tentori et al., 2015).

Another significant gap concerns long-term outcomes (Fan et al., 2015; Moon et al., 2024; Tokarskaya et al., 2020; Wedyan et al., 2021; Zhao et al., 2016) and real-world generalization (Koirala et al., 2019; Maskey et al., 2019; Porayska-Pomsta et al., 2018; Shahidi et al., 2023; Wade et al., 2015). Although VR interventions often demonstrate short-term skill acquisition in controlled environments, few studies systematically evaluate whether gains are maintained over time or transferred to everyday settings such as schools (Maskey et al., 2019; Shahidi et al., 2023; Tokarskaya et al., 2020; Zhao et al., 2016), homes (Fan et al., 2015; Moon et al., 2024; Porayska-Pomsta et al., 2018; Wade et al., 2015), workplaces (Koirala et al., 2019; Shahidi et al., 2023), or community contexts (Koirala et al., 2019; Maskey et al., 2019; Wedyan et al., 2021). The scarcity of longitudinal follow-ups and ecologically valid outcome measures limits the ability to establish sustained and practical benefits.

Methodological heterogeneity also complicates the synthesis of findings. The reviewed studies employed diverse outcome measures and task designs (Boyd et al., 2022; Jeyarani et al., 2024; Salimi et al., 2024), usability assessments (Boyd et al., 2022; Schwarze et al., 2019; Zhao et al., 2017), and physiological indicators (Porayska-Pomsta et al., 2018; Tokarskaya et al., 2020; Zhao et al., 2018a), often without standardization (Herrero and Lorenzo, 2020; Krishnan et al., 2024; Zhang et al., 2015), making cross-study comparisons and meta-analytic integration challenging (Herrero and Lorenzo, 2020; Krishnan et al., 2024; Zhang et al., 2015).

Technological limitations further constrain intervention effectiveness. Hardware restrictions (Ide-Okochi et al., 2024; Koirala et al., 2019; Salimi et al., 2024), usability challenges (Akhtar and Guha, 2019; Moon et al., 2024; Zhao H. et al., 2021), and sensory tolerability issues (Fan et al., 2015; López-Bouzas et al., 2024; Porayska-Pomsta et al., 2018) were frequently reported, especially for younger (Frolli et al., 2021; Tokarskaya et al., 2020) or lower-functioning participants (Koirala et al., 2021; Tran, 2024). Additionally, the adoption of adaptive or personalized VR systems remains limited, with few studies fully leveraging multimodal data, including gaze (Boyd et al., 2022; Schwarze et al., 2019), physiological (Shahidi et al., 2023; Zhao et al., 2018a), and motion tracking (Carreon et al., 2024b; Ghanouni et al., 2021), to support individualized intervention pathways or automated progress monitoring (Carreon et al., 2024a; Smith et al., 2024).

The field lacks standardized outcome measures and multimodal assessment frameworks. Most studies rely on small-scale pilot designs (Koirala et al., 2021; Schwarze et al., 2019; Tran, 2024), self-reports (Frolli et al., 2021; Tokarskaya et al., 2020), or limited sensor metrics (Boyd et al., 2022; Zhao et al., 2018a), limiting reproducibility (Herrero and Lorenzo, 2020; Krishnan et al., 2024) and cross-study comparisons (Boo et al., 2022; Zhang et al., 2015).

In addition to design limitations, the reported intervention outcomes highlight both promise and variability in immersive technology effectiveness for children and adolescents with ASD. Across studies, quantitative results indicate improvements in social and emotional skills, such as affective expression, social reciprocity, and emotion recognition (e.g., pre–post increases of 9%–15% or large effect sizes, d > 1.0) (Adjorlu et al., 2017; Amat et al., 2021; Frolli et al., 2021), and cognitive, motor, and attentional gains, including faster task completion times, higher accuracy, and enhanced problem-solving (e.g., task duration decreased 17%–66%; errors reduced; classification accuracy 77%–97%) (Cecil et al., 2017; Gali et al., 2024; Ju et al., 2024; Schwarze et al., 2019; Tran, 2024). However, many studies relied on small sample sizes (often fewer than 10–20 participants per group) and non-randomized designs, limiting the statistical power and precision of effect estimates (Koirala et al., 2021; Tokarskaya et al., 2020; Tran, 2024). Qualitative and feasibility reports frequently noted usability challenges, such as difficulty with avatar control, calibration issues, and device comfort (HMD tolerability 62%–95%; mild side effects 7%–30%) (Akhtar and Guha, 2019; Boyd et al., 2022; Moon et al., 2024) while engagement and perceived utility were generally positive (median enjoyment 3.6–4.0/5; task compliance >70%) (Boyd et al., 2022; Moon et al., 2024; Zhao et al., 2018a). The integration of advanced technologies, including gaze tracking, motion capture, and AI-based adaptive feedback, was still limited, often restricted to single-task performance or post hoc classification, rather than supporting fully individualized or real-time intervention adaptation (Boyd et al., 2022; Carreon et al., 2024a; Carreon et al., 2024b; Schwarze et al., 2019; Shahidi et al., 2023; Smith et al., 2024). Collectively, these findings underscore the need for larger, multi-site trials with standardized outcome measures and robust assessment frameworks to establish reliable efficacy, optimize usability, and support long-term generalization of skills to naturalistic settings (Boo et al., 2022; Cecil et al., 2017; Dow et al., 2024; Herrero and Lorenzo, 2020; Krishnan et al., 2024; Zhang et al., 2015).

3.8 Trends over time (RQ9–RQ10)

Over the past decade, research on immersive technologies for individuals with ASD has shown a clear trajectory toward increasing sophistication, scale, and application diversity. Early studies (2015–2017) predominantly utilized VR, often in the form of HMDs, semi-immersive VR, or CAVE, targeting primarily children and adolescents aged 6–18 years (Fan et al., 2015; Gelsomini, 2016; Ghanouni et al., 2021; Herrero and Lorenzo, 2020; Intawong et al., 2025; Ip et al., 2020; Jeyarani et al., 2024; Koirala et al., 2019; Koirala et al., 2021; Krishnan et al., 2024; Maskey et al., 2019; Salimi et al., 2024; Schwarze et al., 2019; Tentori et al., 2015; Thomsen and Adjorlu, 2021; Tokarskaya et al., 2020; Wade et al., 2015; Wade et al., 2016; Yang and Lee, 2022; Zhao et al., 2016). Sample sizes were generally small, typically fewer than 60 participants, and augmented reality (AR) was largely limited to prototype or experimental forms (Koirala et al., 2019; Schwarze et al., 2019; Zhao et al., 2016). Outcome measures during this period were mainly qualitative, relying on teacher observations, parent interviews, and session video recordings to assess engagement, social behavior, and task performance years (Fan et al., 2015; Gelsomini, 2016; Ghanouni et al., 2021; Herrero and Lorenzo, 2020; Intawong et al., 2025; Ip et al., 2020; Jeyarani et al., 2024; Koirala et al., 2019; Koirala et al., 2021; Krishnan et al., 2024; Maskey et al., 2019; Salimi et al., 2024; Schwarze et al., 2019; Tentori et al., 2015; Thomsen and Adjorlu, 2021; Tokarskaya et al., 2020; Wade et al., 2015; 2016; Yang and Lee, 2022; Zhao et al., 2016).

Between 2018 and 2020, immersive technologies expanded. VR remained dominant, but MR, haptic-enhanced VR, and full-body interaction systems emerged, supporting collaborative and sensorimotor-enriched interventions (Adjorlu et al., 2017; Akhtar and Guha, 2019; Boo et al., 2022; Boyd et al., 2022; Carreon et al., 2024a; Dhamodharan et al., 2020; Frolli et al., 2021; Frolli et al., 2022; Ide-Okochi et al., 2024; Ju et al., 2024; Krishnappa Babu and Lahiri, 2020; López-Bouzas et al., 2024; Lorenzo et al., 2016; Malihi et al., 2020a; Moon et al., 2024; Narayanan, 2019; Porayska-Pomsta et al., 2018; Samantaray et al., 2024; Sayis et al., 2019; 2020; Shahidi et al., 2023; Tran, 2024; Wedyan et al., 2021; Zhang et al., 2015; Zhang et al., 2017; Zhang et al., 2020; Zhao et al., 2017; Zhao H. et al., 2021). Sample sizes increased, and age range broadened to include younger children (as young as 2 years) and older adolescents and adults (López-Bouzas et al., 2024; Sayis et al., 2020; Shahidi et al., 2023). Early applications of AR and motion-gaming systems were also introduced (Akhtar and Guha, 2019; Samantaray et al., 2024; Zhang et al., 2017). Outcome measures evolved toward multimodal assessment, combining behavioral observation with physiological and cognitive metrics such as EEG, ECG, EDA, PPG, accelerometry, and gaze tracking (Alabood et al., 2023; Bozgeyikli et al., 2018; Carreon et al., 2024a; Wongwan et al., 2025). Standardized behavioral and cognitive instruments (e.g., SRS, VABS, PEP-3, MABC-2, SSP, and SCAS) were increasingly adopted (Lorenzo Lledó, 2025; Sayis et al., 2019; Tzanavari et al., 2015; Yeh and Meng, 2020).

From 2021 onward, interventions further diversified and increased in complexity. VR and MR platforms became more interactive and immersive, incorporating haptic feedback, multi-user collaboration, hybrid AR/VR solutions, and participant characteristics expanded, including nonspeaking children, individuals with intellectual disability, and those with comorbid conditions (Abd El-Sattar, 2025; Al Kheraif et al., 2024; Alcañiz et al., 2022; Alcaniz Raya et al., 2020; Al-Saddi et al., 2023; Amat et al., 2021; Babu et al., 2022; Cao et al., 2025; Carreon et al., 2024a; Elkin et al., 2022; Gali et al., 2024; Gali-Perez et al., 2021; Ip et al., 2024; Junaidi et al., 2020; Keshav et al., 2019; Liu et al., 2021; Liu et al., 2022; Lorenzo Lledó, 2025; Lu et al., 2018; Meng and Yeh, 2022; Nalepka et al., 2022; Nazari et al., 2024; Samantaray et al., 2023; Sanku et al., 2023; Sayis et al., 2022; Smith et al., 2024; Soltiyeva et al., 2023; Tzanavari et al., 2015; Wongwan et al., 2025; Yan et al., 2025; Yeh and Meng, 2025; Zhao et al., 2018a; Zhao J. Q. et al., 2021). Sample sizes reached up to 176 participants, reflecting a trend toward larger-scale, applied interventions (Al-Saddi et al., 2023; Alvarado et al., 2023; Meng and Yeh, 2022).

The trend toward AI and adaptive features shows a clear evolution from static, pre-scripted simulations toward intelligent, closed-loop systems that respond in real time to user performance, behavior, and physiological state. Early adaptive systems modulated difficulty based on task success or repetition logic (Carreon et al., 2024b; Kouhbanani et al., 2021) while modern implementations use multimodal sensor fusion, including eye tracking (Al Kheraif et al., 2024; Amat et al., 2021; Cecil et al., 2017; De Luca et al., 2021; Ip et al., 2020; Johnston et al., 2019; Koirala et al., 2019; Liu et al., 2021; Meng and Yeh, 2022; Narayanan, 2019; Porayska-Pomsta et al., 2018; Wade et al., 2015; Wade et al., 2016; Yeh and Meng, 2020), EEG (Amat et al., 2021; Ip et al., 2020; Koirala et al., 2019; Sayis et al., 2020; Tentori et al., 2015; Wang et al., 2021), physiological signals (Al Kheraif et al., 2024; Amat et al., 2021; Cao et al., 2025; De Luca et al., 2021; Dow et al., 2024; Junaidi et al., 2020; Lee and Yang, 2025; Liu et al., 2022; Sayis et al., 2020; Wang et al., 2021), and motion tracking (Al Kheraif et al., 2024; De Luca et al., 2021; Johnston et al., 2019; Junaidi et al., 2020; Lee and Yang, 2025; Lu et al., 2018; Narayanan, 2019; Silva et al., 2025), to infer cognitive load, engagement, and affect. Machine learning enables automated assessment and diagnosis using gaze, movement, or multimodal data (Alcañiz et al., 2022; Amat et al., 2021; Johnston et al., 2019; Koirala et al., 2019; Lu et al., 2018; Malihi et al., 2020a; Narayanan, 2019; Porayska-Pomsta et al., 2018; Silva et al., 2025; Tentori et al., 2015; Yeh and Meng, 2025), prediction of skill development (Amat et al., 2021; Ip et al., 2020; Koirala et al., 2019; Sayis et al., 2020; Tentori et al., 2015; Wade et al., 2015; Wang et al., 2021; Zhang et al., 2020) and automated behavior coding (Al Kheraif et al., 2024; De Luca et al., 2021; Johnston et al., 2019; Junaidi et al., 2020; Lee and Yang, 2025; Malihi et al., 2020a; Narayanan, 2019). AI-powered intelligent virtual agents include collaborative partners (Al Kheraif et al., 2024; Alcañiz et al., 2022; Johnston et al., 2019; Kouhbanani et al., 2021; Tran, 2024; Zhao et al., 2017), tutoring/guidance agents (Bauer et al., 2023; Biffi et al., 2018; Cao et al., 2025; Dow et al., 2024; Frolli et al., 2021; Frolli et al., 2022; Junaidi et al., 2021; Kouhbanani et al., 2021; Li et al., 2025; Shahab et al., 2022; Yan et al., 2025; Yang and Lee, 2022; Zhao J. Q. et al., 2021), and proactive assistive agents (Frolli et al., 2022; Liu et al., 2021; Zhao J. Q. et al., 2021). Sophisticated personalization is achieved via adaptive difficulty algorithms (Al Kheraif et al., 2024; Bozgeyikli et al., 2018; Carreon et al., 2024b; De Luca et al., 2021; Johnston et al., 2019; Kouhbanani et al., 2021; Lu et al., 2018; Malihi et al., 2020a; Meng and Yeh, 2022; Porayska-Pomsta et al., 2018; Shahab et al., 2022), dynamic prompting (Frolli et al., 2022; Lu et al., 2018; Moon et al., 2024; Porayska-Pomsta et al., 2018; Zhao J. Q. et al., 2021), and sensor-driven biofeedback to regulate anxiety, affect, or cognitive load (Amat et al., 2021; Dow et al., 2024; Koirala et al., 2019; Sayis et al., 2020; Wang et al., 2021; Yeh and Meng, 2020; Zhao et al., 2016). Recent innovations include generative AI and large language models for dynamic social simulations and modeling human supporter behavior (Biffi et al., 2018; Frolli et al., 2022).

Outcome measures have shifted from qualitative observations toward quantitative, multimodal, and system-generated metrics, combining VR/AR task performance, eye tracking, motion tracking, standardized behavioral scales, physiological monitoring, and machine learning-based analyses (Alcañiz et al., 2022; Carreon et al., 2024a; Oki et al., 2025). Social interaction and communication metrics remain central, including observational coding, speech/language analysis, and parent/teacher questionnaires (Keshav et al., 2019; Porayska-Pomsta et al., 2018; Sayis et al., 2022; Schwarze et al., 2019). User experience and engagement are increasingly measured via SUS, ITC-SoPI, Fun Toolkit, and physiological correlations such as heart rate variability and motion synchrony (Bauer et al., 2023; Lee and Yang, 2025; Schmidt et al., 2023). Task-specific and adaptive metrics now support motor coordination, cognitive mapping, daily living skills, and emotion recognition (Bian et al., 2015; Ip et al., 2024; Zhang et al., 2017).

Despite these advances, many studies highlight the need for larger and more diverse samples, including female participants, lower-functioning individuals, and participants with co-occurring conditions, to improve statistical power and generalizability (Adjorlu et al., 2017; Frolli et al., 2021; Frolli et al., 2022; Jeyarani et al., 2024; Koirala et al., 2021; López-Bouzas et al., 2024; Tokarskaya et al., 2020; Tran, 2024). Longitudinal studies are scarce, limiting understanding of skill maintenance and transfer to real-world contexts such as schools, homes, workplaces, and community settings (Alabood et al., 2023; Koirala et al., 2019; Maskey et al., 2019; Porayska-Pomsta et al., 2018; Shahidi et al., 2023; Wade et al., 2015; Zhao et al., 2016). Standardized, ecologically valid, and multimodal assessment frameworks are rarely implemented, constraining reproducibility and cross-study comparison (Boyd et al., 2022; Ghanouni et al., 2021; Ide-Okochi et al., 2024; Krishnan et al., 2024; Schwarze et al., 2019; Zhao H. et al., 2021).

Future interventions should adopt adaptive and personalized approaches, including personalized avatars, variable difficulty levels, and multimodal tracking (gaze, physiological, motion) to support individualized learning pathways and automated progress monitoring (Carreon et al., 2024a; Narayanan, 2019; Tentori et al., 2015; Thomsen and Adjorlu, 2021; Tokarskaya et al., 2020; Zhao et al., 2018a). Expanding VR content to cover complex social scenarios, life skills, and hybrid AR/VR environments is recommended (Gelsomini, 2016; Ju et al., 2024; Moon et al., 2024; Salimi et al., 2024; Wade et al., 2016; Wedyan et al., 2021) alongside investigating skill transfer between virtual and real-world environments and integrating emerging technologies such as brain-computer interfaces, haptic feedback, AI-driven personalization, and portable VR platforms (Akhtar and Guha, 2019; Fan et al., 2015; Porayska-Pomsta et al., 2018; Smith et al., 2024; Zhang et al., 2020; Zhao et al., 2017). Well-designed randomized controlled trials, crossover studies, and comparisons with traditional interventions (e.g., CBT), using multimodal and objective outcome measures, are essential to establish robust evidence of effectiveness (Frolli et al., 2022; Koirala et al., 2021; López-Bouzas et al., 2024; Shahidi et al., 2023; Tokarskaya et al., 2020; Tran, 2024).

Based on a systematic analysis of 141 studies (2015–2025), the trajectory of future research reveals a progression from foundational proof-of-concept to demands for robust, inclusive, and scalable implementation. The foremost need is rigorous longitudinal, controlled studies that definitively establish real-world skill generalization and long-term efficacy, as early studies identified this gap. However, evidence remains largely confined to in-system performance. Research must embrace true personalization and diversity, including representative samples across the full autism spectrum, accounting for sensory profiles, co-occurring conditions, and underrepresented demographics.

3.9 Study quality and methodological rigor

The methodological quality of the included studies was assessed using two complementary approaches: RoB and MMAT, rather than redundant methods. The MMAT provided an overall appraisal of methodological quality across diverse study designs, whereas the adapted risk-of-bias framework enabled evaluation of specific sources of bias, including selection bias, performance bias, attrition bias, and reporting bias. Findings from both assessments informed the interpretation of the strength and reliability of the evidence.

Manual assessment using a five-item Yes/No checklist revealed that most studies were at high risk of bias. Specifically, 111 studies (78.7%) were classified as high risk (RoB 60%–80%), 16 studies (11.3%) as moderate risk (RoB 40%), and 14 studies (10.0%) as low risk (RoB 20%) (Figure 8). Using the MMAT 2018 criteria, studies were graded according to the proportion of criteria fulfilled. Six studies achieved Grade A (highest quality), 18 studies were graded as B, 68 studies as C, and 49 studies received Grade D (lowest quality) (Figure 9).

Taken together, these findings suggest that although immersive technology interventions for ASD frequently report positive outcomes, the overall evidence base should be interpreted with caution. The predominance of studies with high risk of bias and moderate-to-low methodological quality indicates that reported intervention effects may be influenced by limitations such as small sample sizes, non-randomized designs, lack of control groups, short intervention durations, and limited follow-up assessments. Consequently, while the findings of this review support the potential value of immersive technologies for improving social, emotional, cognitive, and functional outcomes, the strength of the evidence varies across studies and outcome domains.

At the same time, several high-quality studies with a low risk of bias demonstrate that rigorous research is achievable in this field. These studies provide stronger support for the feasibility and effectiveness of immersive technology interventions and highlight the importance of adopting more robust methodological practices in future research. Overall, the quality assessment findings underscore the need for larger randomized controlled trials, standardized outcome measures, and longer-term follow-up evaluations to strengthen the evidence base and improve confidence in reported intervention effects.

4 Discussion

The observed temporal growth in VR/AR/MR/XR research involving children and adolescents with ASD reflects a broader maturation of immersive technologies and their increasing accessibility within clinical and educational contexts. The marked increase in publications after 2019, with peaks in 2022 and 2024, coincides with advances in consumer-grade head-mounted displays, improved motion tracking, and the integration of intelligent adaptive features (Al-Saddi et al., 2023; Alvarado et al., 2023; Amat et al., 2021; Cao et al., 2025; Ide-Okochi et al., 2024; Meng and Yeh, 2022; Narayanan, 2019; Sanku et al., 2023; Sayis et al., 2020), making immersive interventions more feasible for pediatric neurodevelopmental populations (Carreon et al., 2024a; Cecil et al., 2017; Junaidi et al., 2020; Malihi et al., 2020a). This trend also aligns with heightened interest in technology-mediated interventions following the COVID-19 pandemic, during which remote, simulated, and technology-supported therapeutic approaches gained prominence (Al Kheraif et al., 2024; Alcañiz et al., 2022; Bozgeyikli et al., 2018; Carreon et al., 2024a; Smith et al., 2024). Collectively, these findings suggest a shift from exploratory feasibility studies toward more structured intervention research (Akhtar and Guha, 2019; Boyd et al., 2022; Koirala et al., 2021; Schwarze et al., 2019; Tran, 2024) although the sustained variability in annual output indicates that the field is still evolving rather than fully consolidated (Herrero and Lorenzo, 2020; Krishnan et al., 2024; Zhang et al., 2015).

Geographically, the concentration of studies in North America, Europe, and parts of Asia highlights substantial regional disparities in research capacity and adoption of immersive technologies for ASD interventions (Al-Saddi et al., 2023; Alvarado et al., 2023; Amat et al., 2021; Carreon et al., 2024a; Carreon et al., 2024a; Jeyarani et al., 2024; Junaidi et al., 2020; Koirala et al., 2019; Maskey et al., 2019; Meng and Yeh, 2022; Narayanan, 2019; Sanku et al., 2023; Sayis et al., 2020; Shahidi et al., 2023; Thomsen and Adjorlu, 2021; Tzanavari et al., 2015; Yeh and Meng, 2020; Zhao et al., 2016; Zhao et al., 2017). The dominance of the United States, China, Spain, and India likely reflects a combination of technological infrastructure, research funding availability, and established ASD research networks (Jeyarani et al., 2024; Meng and Yeh, 2022; Narayanan, 2019; Shahidi et al., 2023; Tzanavari et al., 2015; Zhao et al., 2016; Zhao et al., 2017). In contrast, the limited representation from Africa, South America, and Oceania underscores persistent global inequities in access to advanced therapeutic technologies and research dissemination (Amat et al., 2021; Carreon et al., 2024a; Maskey et al., 2019; Sayis et al., 2020; Thomsen and Adjorlu, 2021). This imbalance raises concerns regarding the cultural, contextual, and socioeconomic generalizability of current evidence, as intervention design, usability, and acceptability may vary significantly across regions (Al-Saddi et al., 2023; Carreon et al., 2024a; Junaidi et al., 2020; Koirala et al., 2019). The underrepresentation of low- and middle-income countries also suggests that existing findings may not fully capture implementation challenges related to cost, technical support, or contextual appropriateness (Alvarado et al., 2023; Yeh and Meng, 2020; Zhao et al., 2017).

The participant characteristics reported across the included studies reveal substantial heterogeneity and several recurring methodological limitations. The predominance of small-scale samples, with more than 80% of studies involving fewer than 30 participants, reflects the logistical, financial, and recruitment challenges associated with immersive technology research in pediatric ASD populations (Amat et al., 2021; Jeyarani et al., 2024; Koirala et al., 2021; Meng and Yeh, 2022; Narayanan, 2019; Sayis et al., 2020; Schwarze et al., 2019; Shahidi et al., 2023; Thomsen and Adjorlu, 2021; Tokarskaya et al., 2020; Tzanavari et al., 2015; Zhao et al., 2016; Zhao et al., 2017). This imbalance raises concerns regarding the external validity of the current evidence base, as the usability, acceptability, and therapeutic response to immersive interventions may differ across sexes, potentially limiting the generalizability of findings to the broader ASD population. While such sample sizes are common in early-stage or feasibility-oriented VR research, they constrain statistical power and limit the generalizability of findings (Al-Saddi et al., 2023; Alvarado et al., 2023; Carreon et al., 2024a; Frolli et al., 2021; Junaidi et al., 2020; Koirala et al., 2019; López-Bouzas et al., 2024; Porayska-Pomsta et al., 2018). The strong male predominance mirrors known ASD prevalence patterns but also highlights the continued underrepresentation of female participants, which may obscure gender-specific responses to immersive interventions (Amat et al., 2021; Carreon et al., 2024a; Frolli et al., 2022; Koirala et al., 2021; Narayanan, 2019; Schwarze et al., 2019; Shahidi et al., 2023; Tokarskaya et al., 2020; Tzanavari et al., 2015; Zhao et al., 2017).

Although most studies employed established diagnostic tools (Al-Saddi et al., 2023; Alvarado et al., 2023; Amat et al., 2021; Carreon et al., 2024a; Frolli et al., 2022; Jeyarani et al., 2024; Koirala et al., 2021; López-Bouzas et al., 2024; Meng and Yeh, 2022; Narayanan, 2019; Porayska-Pomsta et al., 2018; Sanku et al., 2023; Sayis et al., 2020; Schwarze et al., 2019; Shahidi et al., 2023; Thomsen and Adjorlu, 2021; Tokarskaya et al., 2020; Tzanavari et al., 2015; Yeh and Meng, 2020; Zhao et al., 2016; Zhao et al., 2017), reporting of ASD severity, functional level, and comorbid conditions was often inconsistent or incomplete (Amat et al., 2021; Carreon et al., 2024a; Frolli et al., 2022; Jeyarani et al., 2024; Koirala et al., 2021; Meng and Yeh, 2022; Narayanan, 2019; Schwarze et al., 2019; Shahidi et al., 2023; Tokarskaya et al., 2020; Tzanavari et al., 2015; Zhao et al., 2017). Given the high prevalence of co-occurring neurodevelopmental and psychiatric conditions in ASD, this lack of systematic reporting limits interpretability and hampers cross-study comparisons (Amat et al., 2021; Frolli et al., 2022; Koirala et al., 2021; Narayanan, 2019; Schwarze et al., 2019; Shahidi et al., 2023; Tokarskaya et al., 2020; Tzanavari et al., 2015; Zhao et al., 2017). Exclusion of participants with common comorbidities further narrows the ecological validity of findings and may result in overly optimistic assessments of intervention feasibility or effectiveness (Al-Saddi et al., 2023; Alvarado et al., 2023; Carreon et al., 2024a; Junaidi et al., 2020; Sanku et al., 2023; Sayis et al., 2020; Schwarze et al., 2019; Shahidi et al., 2023). Collectively, these results indicate a need for larger, more diverse samples and more standardized reporting of participant characteristics, including comorbidities and functional profiles, to strengthen the evidence base and improve the clinical relevance of immersive technology-based interventions for ASD.

The technological landscape of immersive technology-based interventions for children and adolescents with ASD is substantially diverse, with a clear dominance of immersive VR and HMD-based systems. VR-based interventions dominated the literature, accounting for more than half of all included studies (55.3%), whereas AR (12.8%) and MR (8.5%) were substantially less represented. Consequently, most evidence regarding intervention effectiveness originates from VR-based systems, limiting robust comparisons across immersive technology modalities. Fully immersive HMD-based systems were the predominant technological approach, whereas projection-based and CAVE environments represented only a small proportion of studies (4.3%). This distribution suggests that current evidence primarily reflects highly immersive interventions, with comparatively limited evidence regarding lower-immersion or projection-based approaches. This predominance reflects the relative maturity, affordability, and flexibility of VR platforms such as Oculus and HTC Vive, along with the widespread adoption of game engines such as Unity3D and Unreal Engine, which enable rapid prototyping and customization of therapeutic environments (Al-Saddi et al., 2023; Koirala et al., 2021; Meng and Yeh, 2022; Narayanan, 2019; Porayska-Pomsta et al., 2018; Sanku et al., 2023; Schwarze et al., 2019; Shahidi et al., 2023; Tzanavari et al., 2015; Zhao et al., 2017). The comparatively lower uptake of AR and MR technologies suggests that these approaches, while promising for blending real-world context with virtual cues, remain constrained by hardware cost, usability challenges, and limited empirical validation in pediatric ASD populations (Akhtar and Guha, 2019; Amat et al., 2021; Koirala et al., 2019; Samantaray et al., 2024; Schwarze et al., 2019; Zhang et al., 2017; Zhao et al., 2016). Similarly, the smaller number of studies employing haptic-enhanced, projection-based, or CAVE systems highlights ongoing trade-offs between ecological validity and practical feasibility, as these systems often require specialized infrastructure and technical expertise (Akhtar and Guha, 2019; Amat et al., 2021; Junaidi et al., 2020; Koirala et al., 2019; Koirala et al., 2021; Samantaray et al., 2024; Sayis et al., 2020; Tzanavari et al., 2015). The emergence of collaborative virtual environments and sensor-integrated systems indicates a growing shift toward more interactive, socially embedded, and data-rich intervention designs (Al-Saddi et al., 2023; Alvarado et al., 2023; Amat et al., 2021; Carreon et al., 2024a; Meng and Yeh, 2022; Narayanan, 2019; Porayska-Pomsta et al., 2018; Sanku et al., 2023; Shahidi et al., 2023; Tzanavari et al., 2015; Zhao et al., 2017). However, the heterogeneity in hardware configurations and simulation types complicates cross-study comparisons and limits conclusions regarding technology-specific effectiveness (Amat et al., 2021; Ide-Okochi et al., 2024; Junaidi et al., 2020; Koirala et al., 2021; Narayanan, 2019; Schwarze et al., 2019; Shahidi et al., 2023; Tzanavari et al., 2015; Zhao et al., 2017). Overall, these findings suggest that current research prioritizes technological accessibility and immersion over standardization, underscoring the need for clearer reporting, comparative evaluations across technology types, and greater alignment between simulation design and targeted therapeutic objectives.

The increasing integration of AI-enabled components within immersive ASD interventions reflects a shift toward more responsive, personalized, and data-driven therapeutic systems. Emotion recognition, adaptive algorithms, and multimodal behavioral analysis were commonly employed to enhance ecological validity and tailor intervention delivery in real time (Abd El-Sattar, 2025; Bian et al., 2015; Cao et al., 2025; Cecil et al., 2017; Dhamodharan et al., 2020; Ip et al., 2018; Johnston et al., 2022; Junaidi et al., 2020; Minissi et al., 2023; Moya-Albor et al., 2018; Yan et al., 2025; Yeh and Meng, 2020). Compared with conventional immersive interventions that relied primarily on fixed scenarios and pre-programmed feedback, AI-enabled systems offered greater capacity for real-time adaptation based on user behavior, engagement, and performance. This shift suggests a movement from static intervention delivery toward more individualized and responsive therapeutic experiences. These approaches align well with the heterogeneity of ASD presentations, as AI-driven adaptation allows task demands and feedback to be dynamically adjusted to individual performance and engagement levels (Dow et al., 2024; Romadlon Junaidi et al., 2022; Yeh and Meng, 2020; Zhao et al., 2018a).

The frequent use of eye tracking, motion capture, and multimodal sensing further underscores a growing emphasis on fine-grained behavioral measurement, enabling objective assessment of social attention, joint attention, and interaction patterns that are difficult to capture using traditional observational methods. It is important to recognize that the AI-enabled interventions identified in this review represent a continuum of intelligent and adaptive technologies rather than a single homogeneous category. Across studies, AI-related components ranged from machine learning-based emotion recognition and multimodal behavioral analysis to adaptive feedback mechanisms informed by eye tracking and motion-sensing data. Consequently, the sophistication and degree of autonomy of these systems varied considerably. In some cases, sensing technologies primarily supported behavioral monitoring and adaptive feedback rather than advanced artificial intelligence in the strict sense. This heterogeneity complicates direct comparisons and suggests that the current AI evidence base should be interpreted as encompassing varying levels of intelligent functionality and technological maturity.

Another consideration is the validity and reliability of AI-driven emotion recognition and behavioral analysis systems. Given the substantial heterogeneity of ASD presentations, variations in facial expressions, communication styles, sensory profiles, and behavioral characteristics may affect the accuracy and generalizability of AI-based inferences across individuals and contexts. Relatively few studies reported detailed information regarding model training datasets, validation procedures, or algorithmic performance, making it difficult to evaluate reliability, fairness, and robustness. However, despite their promise, AI-enabled features were often implemented in highly bespoke systems, with limited transparency regarding model training, validation, or robustness across participants and contexts (Bian et al., 2015; Cao et al., 2025; Dhamodharan et al., 2020; Ip et al., 2018; Li et al., 2015; Minissi et al., 2023; Moya-Albor et al., 2018; Newbutt et al., 2020; Yuan and Ip, 2018). The limited reporting of training and validation data also raises concerns regarding potential algorithmic bias, particularly if AI models are developed using datasets that are not representative of the broader ASD population. Such biases may reduce system performance for specific subgroups and affect the equity and reliability of intervention delivery.

Few studies evaluated the independent contribution of AI components relative to non-adaptive or rule-based systems, making it difficult to determine whether observed benefits were attributable to intelligent adaptation or to increased immersion more broadly (Cao et al., 2025; Johnston et al., 2022; Minissi et al., 2023; Moya-Albor et al., 2018). Consequently, although AI-enabled interventions appear to enhance personalization and responsiveness, the current evidence base does not permit firm conclusions regarding whether AI-integrated systems consistently outperform comparable non-AI immersive interventions.

Ethical considerations also warrant greater attention. Many AI-enabled immersive systems collect sensitive behavioral, physiological, gaze tracking, and interaction data from children and adolescents, creating important challenges related to privacy, data security, informed consent, and long-term data governance. In addition, some AI-driven adaptive systems operate with limited transparency regarding how decisions or recommendations are generated, potentially reducing interpretability for clinicians, educators, caregivers, and end users. Moreover, the reliance on specialized hardware and computational resources raises concerns about scalability and real-world deployment, particularly in clinical or educational settings with limited technical support (Akhtar and Guha, 2019; Alvarado et al., 2023; Dow et al., 2024; Ip et al., 2016; Miller et al., 2020; Newbutt et al., 2020; Peng et al., 2019; Silva et al., 2025; Tzanavari et al., 2015; Wongwan et al., 2025). Finally, it is important to distinguish between experimental research prototypes and clinically validated intervention tools. While many AI-enabled immersive systems have shown promising preliminary results, most were evaluated in small-scale research settings and have not yet undergone extensive clinical validation or large-scale real-world implementation.

The distribution of comparator conditions across studies highlights a field that remains largely exploratory, with a strong reliance on single-arm and pre–post designs. While these designs are appropriate for early-stage feasibility, usability, and proof-of-concept evaluations, their predominance limits causal inference regarding the effectiveness of immersive VR-based interventions. Within-subject and crossover designs offered improved control over individual variability and consistently demonstrated gains following intervention exposure (Al-Saddi et al., 2023; Carreon et al., 2024b; Falivene et al., 2025; Frolli et al., 2022; Junaidi et al., 2020; Malihi et al., 2020a; Naranjo et al., 2017; Nazari et al., 2024; Tokarskaya et al., 2020; Wedyan et al., 2021; Zhang et al., 2020); however, they remain vulnerable to learning effects and temporal confounds. The relatively small proportion of studies using active comparators, treatment-as-usual, or waitlist controls suggests that rigorous efficacy testing against established interventions is still limited.

Comparisons with typically developing peers were primarily used to benchmark baseline performance and characterize ASD-specific profiles rather than evaluating intervention efficacy, which constrains their interpretive value for clinical outcomes. A notable pattern across the evidence base was the predominance of single-arm and pre–post study designs, which accounted for the largest proportion of included studies. While these studies frequently reported improvements relative to baseline, the absence of comparator conditions limits causal inference and increases the possibility that observed gains may reflect factors such as maturation, repeated testing, or participant expectancy effects. Encouragingly, studies that included passive or active control conditions generally reported superior outcomes for VR-based interventions, supporting their potential added value. However, controlled studies represented a substantially smaller proportion of the literature and often varied considerably in comparator type, intervention duration, and outcome assessment. As a result, although positive findings were observed across both controlled and uncontrolled designs, the evidence is generally stronger for studies incorporating active or passive comparison groups than for pre–post evaluations alone. Nevertheless, the heterogeneity of comparator types, outcome measures, and intervention targets complicates cross-study synthesis and precludes firm conclusions about relative effectiveness. Overall, these findings underscore the need for more robust comparative designs, including adequately powered randomized controlled trials and standardized comparator frameworks, to advance the field from feasibility toward evidence-based clinical implementation.

The distribution of therapeutic objectives across the reviewed studies underscores the versatility of immersive VR technologies in addressing a wide spectrum of developmental and functional needs in children and adolescents with ASD. The strong emphasis on social and emotional skills reflects longstanding intervention priorities in ASD research and highlights VR’s capacity to simulate socially complex yet controlled environments for practicing interaction, emotion recognition, and regulation (Ju et al., 2024; Koirala et al., 2021; Schwarze et al., 2019; Thomsen and Adjorlu, 2021). Cognitive, educational, and daily living skill interventions further demonstrate the adaptability of immersive systems for supporting learning, independence, and real-world task rehearsal, particularly where repetition and contextual realism are critical (Al-Saddi et al., 2023; Amat et al., 2021; Boo et al., 2022; Lu et al., 2018; Sanku et al., 2023). Across therapeutic domains, a broadly consistent pattern emerged: immersive interventions were associated with improvements in the specific skills targeted by the intervention. Positive findings were reported for social communication, emotional understanding, cognitive performance, educational outcomes, daily living skills, and motor-related competencies, suggesting that the potential benefits of immersive technologies are not confined to a single developmental domain. However, the strength of evidence varied across domains, with social and emotional interventions supported by a larger and more mature evidence base than educational, motor, or daily living applications. Across domains, reported outcomes were generally positive, suggesting gains in targeted skills, engagement, and confidence; however, outcome measures were often task-specific and short-term, with limited evidence of sustained effects or real-world generalization (Fan et al., 2015; Koirala et al., 2019; Maskey et al., 2019; Moon et al., 2024; Porayska-Pomsta et al., 2018; Shahidi et al., 2023; Tokarskaya et al., 2020; Wade et al., 2015; Wedyan et al., 2021; Zhao et al., 2016).

Despite consistent positive findings across domains, comparisons of effectiveness remain challenging because studies differed substantially in intervention duration, outcome measures, participant characteristics, and methodological quality. Consequently, it is difficult to determine whether certain domains benefit more from immersive technologies than others. Importantly, the breadth of intervention targets contrasts with the relatively narrow focus of many individual studies, which typically examine single skill domains in isolation. This fragmentation limits understanding of how improvements in one area may transfer to broader functional outcomes or quality of life. Moreover, reliance on heterogeneous outcome measures complicates cross-study comparison and synthesis (Boyd et al., 2022; Jeyarani et al., 2024; Ju et al., 2024; Salimi et al., 2024; Schwarze et al., 2019; Zhao et al., 2017; Zhao et al., 2018a).

Collectively, these findings indicate that while VR-based interventions show promise across multiple therapeutic domains, future research should prioritize longitudinal designs, standardized outcome frameworks, and evaluation of generalization to everyday contexts to better establish clinical significance and real-world impact. Importantly, many included studies primarily evaluated feasibility, usability, engagement, or proof-of-concept outcomes rather than therapeutic efficacy. Consequently, although positive findings were frequently reported, these results should be interpreted cautiously and should not necessarily be considered evidence of sustained or clinically meaningful treatment effects.

The predominance of non-randomized and small-scale quantitative designs highlights a critical methodological limitation in immersive technology-based ASD intervention research. Although feasibility and exploratory studies are essential for testing novel technologies, the limited number of randomized controlled trials constrains the ability to draw causal inferences regarding intervention effectiveness. Moreover, the underrepresentation of female participants, lower-functioning individuals, and participants with co-occurring conditions indicates that current evidence may not generalize across the broader autism spectrum. Longitudinal evaluation and real-world generalization remain largely unexplored, with few studies assessing whether skill gains persist beyond controlled settings or transfer to everyday contexts such as schools, homes, and community environments. The heterogeneity of outcome measures, task designs, and usability assessments further complicates cross-study synthesis and meta-analytic integration. Technological and usability constraints, including hardware limitations, sensory tolerability, and underutilization of adaptive, multimodal systems, limit the potential for personalized intervention delivery. Collectively, these patterns underscore the need for more rigorous, standardized, and ecologically valid research designs that leverage advanced immersive technologies while ensuring inclusive and generalizable findings across diverse ASD populations.

The quality assessment reveals a clear tension between feasibility and rigor in immersive technology-based ASD research. While a minority of studies met high methodological standards (MMAT Grade A; low RoB), most had a high risk of bias and moderate to low quality, reflecting common challenges such as small sample sizes, non-randomized designs, and heterogeneity in outcome measures. This variability underscores the difficulty of balancing technological innovation with rigorous evaluation, particularly in complex immersive interventions involving children and adolescents with diverse abilities and co-occurring conditions. Importantly, the presence of high-quality studies demonstrates that methodologically robust VR research is achievable, providing exemplary design, reporting, and multimodal assessment strategies.

Overall, across immersive modalities, the findings suggest a general trend toward positive outcomes; however, the strength and consistency of evidence vary substantially by technology type, study design, and outcome domain. VR-based HMD systems dominate the evidence base and therefore disproportionately shape the overall conclusions, while AR, MR, and projection-based systems remain underrepresented, limiting meaningful cross-modality inference. Importantly, differences in immersion level, interactivity, sensory complexity, and system adaptability (including AI integration) were rarely isolated as independent variables, making it difficult to determine whether observed effects are driven by modality-specific mechanisms or by shared characteristics of immersive engagement. Furthermore, the predominance of small-scale, feasibility-oriented, non-randomized, and single-arm studies, together with the moderate-to-high risk of bias identified in many studies, limits the strength of causal inference and necessitates cautious interpretation of reported benefits. Consequently, current evidence supports cautious optimism regarding immersive interventions for ASD but does not yet allow definitive conclusions about which technologies, configurations, or design features are most effective across participant profiles and developmental domains.

5 Research implications and limitations

Based on the patterns identified in this systematic review, future research on immersive VR-based interventions for children and adolescents with ASD should move beyond feasibility-driven exploration toward methodologically rigorous and theoretically grounded evaluation. The predominance of small samples, single-arm designs, and short-term outcome assessments underscores the need for adequately powered randomized controlled trials that enable stronger causal inference and more precise estimation of intervention effects. Adopting multi-site trial designs and shared intervention and assessment protocols would be particularly valuable for reducing fragmentation in the current evidence base and improving comparability across studies, as consistently emphasized in prior evidence-based reviews of VR interventions for ASD (Mesa-Gresa et al., 2018).

In parallel, the substantial heterogeneity in participant characteristics observed across studies highlights the need for more systematic reporting of ASD severity, functional level, and co-occurring conditions, alongside deliberate efforts to include underrepresented populations such as female participants and individuals with comorbid neurodevelopmental or psychiatric disorders. Improved transparency and inclusivity in participant characterization would enhance ecological validity and better align VR research with real-world clinical populations, concerns that have been raised in state-of-the-art reviews of immersive and digital interventions for ASD (Parsons and Cobb, 2011). Furthermore, meta-analytic evidence indicates that while VR-based interventions demonstrate overall effectiveness across multiple skill domains, greater methodological consistency and outcome standardization are necessary to support robust clinical interpretation and translation (Karami et al., 2021).

The review further indicates that while AI-enabled features (e.g., adaptive difficulty, emotion recognition, and behavioral sensing) are increasingly incorporated, their independent contribution to intervention effectiveness remains unclear. Future studies should therefore explicitly compare AI-adaptive systems with non-adaptive or rule-based VR environments to determine whether intelligent personalization provides measurable added value beyond immersion alone. Greater transparency regarding AI model design, training, and validation is also necessary to support interpretability, replication, and ethical deployment, consistent with broader recommendations in human–AI interaction research (Shneiderman, 2020).

From a practical perspective, the findings suggest that immersive VR technologies are most effectively positioned as complementary tools within existing educational and therapeutic programs rather than as standalone interventions. The strong emphasis on social and emotional skills across the reviewed studies supports the use of VR as a structured and controlled environment for practicing socially complex behaviors, such as emotion recognition, social reciprocity, and communication, that may be difficult to rehearse safely or repeatedly in real-world settings. This positioning aligns with evidence from a recent systematic review of VR-based social skills training for children and adolescents with ASD, which highlights the use of immersive, repeatable social scenarios in clinical and school-related contexts to support social learning and engagement (Liu, 2023).

However, the dominance of bespoke systems and specialized hardware observed in this review highlights potential barriers to scalability and adoption. To support real-world implementation, practitioners and developers should prioritize technologically accessible platforms, such as commercially available HMDs and modular software architectures, alongside simplified setup and maintenance requirements. These considerations are particularly important for resource-constrained clinical and educational settings, where technical expertise and infrastructure may be limited.

The increasing use of behavioral sensing (e.g., eye tracking and motion capture) and AI-driven adaptation offers promising opportunities for objective assessment and personalized intervention delivery, but their deployment should be accompanied by careful attention to usability, sensory tolerability, and ethical considerations. Clinicians and educators should be involved early in system design to ensure that adaptive features support, rather than complicate, therapeutic goals and decision-making processes. Importantly, the review indicates that current VR interventions often target isolated skill domains, which may limit broader functional impact. Practitioners should therefore consider integrating VR-based activities within multi-domain intervention programs, linking gains in social, emotional, or cognitive skills to real-world practice and reinforcement. Clear guidance on session structure, dosage, and transfer activities may help maximize generalization beyond the virtual environment.

Despite adopting a systematic and comprehensive approach, this review has several limitations. First, only peer-reviewed studies published in English were included, which may have excluded relevant research in other languages or unpublished sources, introducing potential publication and language bias. Second, substantial heterogeneity in study designs, participant characteristics, intervention modalities, and outcome measures precluded formal meta-analysis and limited quantitative synthesis. Third, the rapidly evolving nature of VR/AR/MR/XR technologies means that the reviewed interventions may not fully reflect the latest hardware, software, or AI-enabled capabilities. Finally, the review relies on the completeness and accuracy of reporting in the primary studies, which was often inconsistent, particularly regarding participant functional levels, comorbidities, and technical specifications, limiting the ability to draw fully generalizable conclusions.

6 Conclusion and future directions

Immersive technologies, particularly VR, demonstrate considerable potential as intervention tools for children and adolescents with ASD. This systematic review synthesizes evidence that these technologies can produce significant, quantifiable improvements across core developmental domains, including social-emotional skills, cognitive function, daily living abilities, and motor coordination. Critically, meta-analytic evidence confirms large effect sizes (e.g., SMD >1.4) in social and emotional outcomes (Mittal et al., 2024), while integrated AI systems show high technical precision, with machine learning models achieving up to 97% accuracy in emotion recognition within controlled virtual environments (Amat et al., 2021; Dhamodharan et al., 2020). The field exhibits strong participant engagement, with task compliance often exceeding 70% and usability ratings reflecting generally positive acceptance (Ghanouni et al., 2021; Zhao H. et al., 2021; Zhao et al., 2018a).

However, this promising efficacy evidence exists within a literature still defined by methodological nascency and constraint. The predominance of small-scale, non-randomized studies with heterogeneous designs limits causal inference and obscures the true magnitude and consistency of treatment effects. Furthermore, the field’s focus on high-functioning, predominantly male participants, coupled with inconsistent reporting of comorbidities and functional levels, restricts the ecological validity and generalizability of findings. Crucially, while in-system performance metrics and short-term gains are encouraging, evidence of skill generalization to naturalistic settings and long-term maintenance remains strikingly limited. Collectively, these findings underscore a clear dichotomy: immersive interventions show strong in-principle efficacy and technological promise, yet the evidence base remains insufficient to support their widespread, evidence-based implementation. The transition from compelling proof-of-concept to robust clinical application will require a foundational shift toward greater methodological rigor, representativeness, and ecological validation.

Future research on immersive VR/AR/MR/XR interventions for children and adolescents with ASD must move beyond promising in-system outcomes toward robust evidence of real-world effectiveness. Priority should be given to adequately powered, longitudinal randomized controlled trials employing standardized and ecologically valid outcome measures to establish causal efficacy, long-term skill retention, and generalization to everyday contexts such as schools, homes, and community settings. Studies should also recruit more representative samples that reflect the full heterogeneity of the autism spectrum, including girls and young women, lower-functioning and nonspeaking individuals, and those with co-occurring conditions, to enhance clinical relevance and equity. Methodological advancement should focus on closed-loop, AI-personalized immersive systems that leverage multimodal behavioral data (e.g., eye tracking, physiological signals, and motion capture) for real-time adaptation and objective progress monitoring, while prioritizing translational feasibility through cost-effective, modular platforms co-designed with clinicians, educators, and autistic stakeholders. Future AI-enabled immersive systems should also move beyond simple emotion recognition and adaptive feedback mechanisms toward explainable and transparent AI architectures that allow clinicians, educators, and caregivers to understand how intervention decisions are generated. Greater emphasis should be placed on developing and validating AI-driven digital biomarkers that can objectively track social attention, engagement, emotional regulation, and learning trajectories over time.

Future studies should directly compare AI-enabled and non-AI immersive interventions to determine the unique clinical value contributed by intelligent adaptation, personalization, and predictive analytics. Establishing standardized reporting frameworks for AI model development, training datasets, validation procedures, fairness assessment, and algorithmic performance will be essential to improve reproducibility, reduce bias, and facilitate cross-study comparison. Finally, the responsible scaling of immersive interventions will require clear ethical and implementation frameworks addressing data privacy, transparency, safety, and personalized deployment across resource-diverse settings. Particular attention should be given to privacy-preserving AI approaches, secure management of behavioral and physiological data, and regulatory pathways for translating experimental AI-enabled prototypes into clinically validated and deployable intervention tools.

Generative AI statement

The author(s) declared that generative AI was used in the creation of this manuscript. During the preparation of this work, the authors used ChatGPT in order to correct language errors and suggest restyling when the language was deemed difficult to read by some authors. After using this tool/service, the authors reviewed and edited the content as needed and take full responsibility for the content of the publication.

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

Publisher’s note

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

Supplementary material

The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/frvir.2026.1850096/full#supplementary-material

Supplementary material Supplementary material at the source

Footnotes

Edited by: Elias Manjarrez, Meritorious Autonomous University of Puebla, Mexico

Reviewed by: Yücel Altın, Bursa Uludag Universitesi, Türkiye

Rui Silva, ISPGAYA, Portugal