Introduction

Autism Spectrum Disorders (ASD) are characterized by persistent difficulties in social communication and interaction, often accompanied by restricted, repetitive behaviors1. The severity and manifestation of these symptoms vary significantly across individuals, leading some researchers to conceptualize ASD as a form of neurodiversity rather than a singular disorder2. Given its complexity, ASD necessitates a wide range of therapeutic interventions tailored to each individual’s unique developmental profile3,4. Among the most established approaches, Applied Behavior Analysis (ABA) has gained recognition due to its evidence-based strategies for enhancing cognitive, social, and adaptive skills through systematic behavior modification5,6. One critical skill often impaired in individuals with ASD is Theory of Mind (ToM), the ability to understand others’ thoughts, beliefs, and intentions, which is essential for successful social interactions7. Traditional ToM interventions frequently involve role-playing games with dolls or imaginary characters, aiming to reduce the stress of real-life interactions8,9. However, these methods have limitations because they lack ecological validity, as they do not involve real-time interaction with a responsive social partner.

Technological advancements in social robotics offer innovative solutions to address these limitations. Humanoid robots provide dynamic, interactive environments that can replicate social scenarios in controlled and predictable ways, making them particularly suitable for children with ASD10,11. Unlike human therapists, robots can perform highly consistent social behaviors, such as gaze-following, pointing, and turn-taking, which are critical for ToM development. Moreover, their programmable nature enables therapists to adapt robot behavior in real time, offering personalized and scalable interventions12,13. Building on previous advancements in robot-assisted therapy, our study extends the current state of the art by incorporating a human-assisted control condition and conducting trials within a real clinical environment using a rigorous two-period crossover design. Prior research has explored the benefits of socially assistive robots like iCub and Kaspar in fostering social and cognitive skills in children with developmental disorders14,15,16. However, these interventions were often limited to controlled laboratory settings, raising concerns about ecological validity. In contrast, our study sought to bridge this gap by creating ecologically rich, real-world simulation scenarios, such as social role-plays in simulated cafes or cinemas, providing a familiar, yet structured, context for children to practice ToM-related tasks like recognizing intentions, initiating requests, and taking turns. While previous work has emphasized the importance of physical engagement and predictable robot behavior for enhancing attention and reducing social anxiety11,17, many of these studies lacked direct human comparison groups, making it challenging to isolate the specific contributions of robotic interventions. To address this, we introduced an active human-assisted control, where trained clinicians replicated the robot’s actions. This allowed us to disentangle the unique therapeutic effects attributable to the robot, including its consistency, predictability, and non-threatening social presence.

Several studies have previously compared robot-assisted interventions with traditional human-led approaches, particularly in domains like joint attention and cognitive flexibility. For example, David et al. (2018) employed a single-case alternating treatment design to explore the effects of robot-enhanced versus standard therapy sessions, focusing on joint attention outcomes32. Costescu et al. (2015) similarly examined reversal learning performance in children with ASD under both robot- and human-mediated conditions, reporting increased engagement and enjoyment during robot interactions33. While these studies underscore the promise of socially assistive robots, they typically relied on comparisons with broader or less controlled human-led sessions, rather than using tightly matched human-assisted conditions designed to mirror the robot’s behavior over multiple sessions. Additionally, these studies were typically limited to narrowly focused training outcomes (e.g., joint attention or task-switching), and lacked higher-order cognitive outcomes like Theory of Mind. Building on this foundation, our study implemented a full crossover design in a real-world clinical setting and introduced a structured human-assisted control condition. This allowed us to more precisely isolate the therapeutic contribution of the robot’s embodiment, consistency, and predictability—particularly in supporting Theory of Mind development. Among existing humanoid robots, the iCub robot has emerged as a well-suited platform for clinical applications due to its anthropomorphic design and advanced motor capabilities18. Robots like iCub can generate human-like behaviors such as head movements, pointing gestures, and gaze-following, which are pivotal for social cognition training. These features allow robots to serve as both interactive partners and therapeutic tools, providing consistent, engaging experiences that can be challenging to achieve with human-led interventions alone. In the clinical context, ABA principles can be seamlessly integrated with robot-assisted interventions. Robots can deliver structured learning experiences while minimizing human error, ensuring consistency in task presentation and feedback. Additionally, they allow for automated tracking of behavioral responses, facilitating real-time performance monitoring and enabling adaptive learning19. This integration creates an optimal learning environment where children can practice social skills repeatedly without the unpredictability of human interactions, which may be overwhelming for some children with ASD.

Our study builds on these technological and clinical foundations by evaluating the effectiveness of a robot-assisted Theory of Mind (ToM) training protocol compared to two human-led interventions. We used a two-period crossover design involving 32 children with Autism Spectrum Disorder (ASD), all formally diagnosed using the Autism Diagnostic Interview–Revised (ADI-R) and the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2). Each child received both Robot-Assisted Training (RAT) and their ongoing Standard Therapy (ST). To further test whether the robot itself—rather than the training protocol—drove observed improvements, we included a separate Human-Control (HC) group (N = 14) who received the same training protocol delivered by a clinician. This group was analyzed separately to minimize potential carryover effects and to avoid overburdening participants, many of whom were already involved in multiple weekly therapies. The training sessions employed simulated real-world scenarios, such as ordering at a café or purchasing movie tickets, designed to mimic everyday social interactions. These scenarios provided a familiar yet structured context for children to practice ToM-related tasks, such as initiating requests, recognizing others’ intentions, and taking turns. Each interaction was guided by clinical psychologists who controlled the robot’s behavior remotely, ensuring task accuracy and participant safety. To promote learning progression, tasks increased in complexity, moving from explicit to more subtle social cues like gaze-following, head orientation, and facial expressions. Further supporting this focus on interaction dynamics, research on social robots in educational and clinical settings has demonstrated that robotic interventions can effectively enhance social-emotional learning through structured, predictable interactions that reduce anxiety and promote engagement. For example, Kewalramani et al. (2024) reviewed studies using robots like Nao, Kaspar, and Zeno to teach emotional recognition, turn-taking, and joint attention, emphasizing that such interactions sustain attention and improve learning outcomes through repeated practice20. However, they also highlighted the limited generalization of skills beyond clinical environments, reinforcing the need for real-world application research—a gap our study directly addressed. Additionally, research on social robots in educational settings has shown that their physical embodiment can yield cognitive and affective learning gains comparable to human tutoring. Robots like Nao and Keepon have demonstrated significant improvements in both engagement and learning outcomes due to their ability to elicit human-like interactions through social presence and physical embodiment21. These findings align with our approach of creating ecologically rich simulations while maintaining experimental rigor through a crossover design.

In addition to its therapeutic potential, the robot’s involvement addressed key clinical challenges. By automating repetitive tasks, robots allowed clinicians to focus on higher-level activities such as observing children’s behavior and adjusting the therapy’s direction based on real-time feedback. This dual approach highlights how humanoid robots can complement traditional therapies, enhancing overall clinical efficacy while preserving the essential human element in rehabilitation settings. Our study not only evaluates the clinical utility of robot-assisted ToM training but also explores its broader implications for integrating robotics into therapeutic practices. By leveraging advanced robotics technology in a clinical framework, we aim to establish a scalable, replicable model for ASD interventions that combines the strengths of human expertise with the precision and adaptability of social robots.

Importantly, our approach is not intended to replace human therapists, but rather to augment the therapy by introducing consistent, programmable tools that can enhance clinical outcomes. The robot-assisted protocol was designed for integration into existing therapeutic frameworks, under direct supervision of trained clinicians. This reflects a broader perspective that technological interventions should serve as complements—not substitutes—to professional expertise.

Results

The study evaluated the effects of robot-assisted ToM training (using the humanoid robot iCub) compared to both standard therapy and an active human-assisted control condition. The intervention was delivered over a period of 6 to 8 weeks, with two sessions per week. Each session lasted approximately 10 to 15 min., depending on the child’s engagement and availability. The primary outcome measure was the change in ToM-related scores assessed using the NEPSY-II test (see the Methods section). Data were analyzed using linear mixed-effects models, and all results are reported with corresponding statistical metrics and figure references.

Robot-Assisted Training vs. Standard Therapy.

Analysis of NEPSY-II score improvements revealed a significant main effect of the training protocol on children’s ToM abilities, indicating that robot-assisted training yielded greater improvements than standard therapy. A mixed-effects model analysis showed a statistically significant interaction between training type and assessment phase (F(1, 62) = 43.6, p < .001). Specifically, planned comparisons revealed that children exposed to robot-assisted training demonstrated a mean score increase of βEstimate = 3.53, t(31) = 6.60, p < .001, surpassing gains observed in the standard therapy group. Further analysis of the tri-phase assessment confirmed a significant increase in NEPSY-II scores across training phases (F(2, 60) = 20.63, p < .001). Before the intervention (T0), scores were comparable between the two groups (βEstimate = 1.50, t(31) = 1.15, p = .858). However, by the first assessment after robot-assisted training (T1), children in the robot-assisted group scored significantly higher (βEstimate = 6.04, t(31) = 4.63, p < .001), while those receiving only traditional therapy showed no meaningful improvement (βEstimate = 0.13, t(31) = 0.25, p = .999). See Fig. 1 for a comparison of NEPSY-II score improvements across groups.

Fig. 1
figure 1

Box-plot summarizing the fixed effects of the robot-assisted trainings in comparison with the Standard Therapy (ST) on NEPSY-II subscale Theory of Mind. The plot on the left (A) represents pre–post delta scores for each condition; the plot on the right (B) shows performance across three assessment phases (T0–T1-T2). The group that first received RAT improved significantly after the first training phase (pink bars; T0 vs. T1), while the group that initially received ST showed no improvement during that phase (turquoise bars; T0 vs. T1) but improved following RAT (T1 to T2). Horizontal lines indicate statistically significant comparisons (***p < .001).Crossover Effects and Training Retention.

Full size image

The crossover design allowed for the evaluation of ToM improvements and retention effects. Following the crossover, the previously ST-only group received robot-assisted training from T1 to T2. A significant increase in NEPSY-II scores was observed in this group (βEstimate = 2.67, t(31) = 5.02, p < .001), aligning with the gains seen in the original robot-assisted group (Fig. 1B). Notably, after the crossover, there were no significant differences between the two groups (βEstimate = 2.79, t(31) = 2.14, p = .292), indicating that both cohorts ultimately benefited equally from robot-assisted training. Importantly, children from the initial robot-assisted training group maintained their ToM improvements through the second evaluation phase (T1 to T2), with no observable regression in scores (βEstimate = 0.47, t(31) = 1.17, p = .850), suggesting that the benefits of robot-assisted training persisted beyond the intervention period.

Robot vs. Human-Control

To isolate the specific contribution of the robot beyond the structure of the training protocol, we included a separate Human-Control (HC) group (N = 14) that completed the same role-play activities with a trained clinician rather than the robot. NEPSY-II Theory of Mind scores in the HC group showed no significant improvement between T0 and T1 (βEstimate = 0.36, t(13) = 1.05, p = .315), in contrast to the RAT group, which demonstrated significant gains. An independent samples t-test on NEPSY-II delta scores (Δ) confirmed a statistically significant difference between groups, favoring robot-assisted training (d = − 1.99, t(28) = − 5.44, p < .001). The mean improvement in the RAT group was M = 4.31 ± 2.50, compared to M = 0.71 ± 2.56 in the HC group (see Fig. 2). This result suggests that the training protocol alone was not sufficient to drive the observed benefits and that the robot’s embodiment and interaction dynamics played a critical role in producing measurable cognitive improvements.

Fig. 2
figure 2

Box-plot illustrating pre–post NEPSY-II Theory of Mind scores in the Human-Control (HC) group. Children in this group received the structured training protocol delivered by a trained clinician, without robotic assistance. No statistically significant improvement was observed between T0 and T1.

Full size image

Qualitative observations

Anecdotal reports from clinical staff suggested that children remained more consistently engaged and attentive during robot-assisted sessions than in Human-Control (HC) training. Unlike the HC sessions, where children occasionally lost focus, the robot’s interactive and predictable behavior appeared to sustain their interest throughout. This finding underscores the robot’s ability to maintain high levels of engagement, a critical factor in enhancing learning outcomes.

Discussion

The present study aimed to investigate the clinical utility of a humanoid robot in enhancing social cognition in children with Autism Spectrum Disorder (ASD), specifically focusing on Theory of Mind (ToM) development. Using an innovative rehabilitation protocol grounded in Applied Behavior Analysis (ABA), children participated in structured role-playing activities with the humanoid robot iCub across a range of simulated social scenarios. The study demonstrated that robot-assisted training significantly improved ToM abilities, as assessed by the NEPSY-II test, compared to both standard therapy and a Human-Control (HC) condition. The NEPSY-II, widely used in clinical practice, comprehensively evaluates ToM by assessing abilities such as understanding false beliefs, emotional comprehension, and social perspective-taking. Importantly, improvements observed in the robot-assisted group generalized beyond the specific tasks within the role-playing games, indicating meaningful gains in social cognition that were transferable to broader social contexts. This generalization supports prior research on robot-based interventions22,23, emphasizing the importance of using standardized clinical assessments to validate the efficacy of novel therapeutic approaches.

One of the study’s most notable findings was the superior improvement in ToM skills observed in the robot-assisted group compared to the HC group, where a trained clinician replicated the robot’s role and no improvement was observed. This result highlights the specific contribution of the robot’s technological features, including its anthropomorphic design, consistent behavior, and programmable social interactions. The data suggest that the robot’s role extended beyond being a passive facilitator; it acted as an active, engaging social agent that mediated role-playing activities. Children in the robot-assisted group maintained higher levels of engagement and attention, a finding corroborated by anecdotal reports from clinical staff. In contrast, children in the HC group often lost interest, suggesting that the robot’s presence heightened motivation and prolonged focus. This aligns with existing literature indicating that children with ASD are often more receptive to technology-based interventions presumably due to the predictability and consistency that the technology provides24. The robot’s interactive capabilities likely triggered sustained social engagement, contributing to greater skill acquisition.

The iCub humanoid robot’s technological capabilities likely played a critical role in promoting ToM development. Importantly, no children in the robot-assisted group withdrew from the study, reinforcing findings that children with ASD tend to be highly engaged by technology-enhanced interventions. Its ability to perform social actions such as gaze-following, head turning, and facial expression simulation created an immersive environment where children could practice recognizing and responding to social cues. This capacity to simulate complex social behaviors in a consistent manner, free from human variability, likely facilitated the children’s ability to grasp social communication that are typically challenging for individuals with ASD. Additionally, the robot’s precision in delivering prompts allowed therapists to focus on higher-level clinical tasks such as monitoring progress and tailoring interventions to individual needs. This adaptability supports ABA principles by enabling personalized learning while reducing the cognitive load on clinicians responsible for managing repetitive training tasks. Similar findings have been reported in related work22, highlighting how technological precision can complement clinical expertise. While our findings specifically reflect the use of the iCub humanoid robot, we recognize that many of the observed benefits—particularly enhanced engagement and immersion—may not be unique to this platform alone. The iCub’s human-like morphology and capacity to physically interact with the environment and manipulate objects likely contributed to its effectiveness by capturing children’s attention and enabling ecologically grounded role-play. However, these features are not exclusive to iCub. Prior literature has highlighted similar outcomes using platforms such as Nao34, Kaspar35, and Zeno36. As such, we cautiously hypothesize that the protocol described here could be adapted to other robots capable of basic social behaviors and physical interaction with objects, though empirical validation is needed to confirm this generalizability.

Despite following the same protocol, the HC condition did not yield comparable improvements. Human interactions, even when scripted, inherently involve variability, which can hinder learning in children with ASD who often benefit from predictable, structured environments. Even the most experienced clinicians may unintentionally deviate from standardized task delivery, whereas the robot’s consistent performance ensured uniform task presentation. This finding underscores a broader implication: robots can augment therapeutic protocols by assuming roles that require precision, repetition, and predictability. By automating these elements, clinicians can direct their attention to qualitative aspects of the intervention, such as monitoring progress and adjusting strategies based on individual responses.

Although our study included a multidisciplinary, ABA-informed standard therapy condition (passive control), the lack of observed improvements in ToM during this period may appear inconsistent with existing literature on ABA efficacy (see, for example31), . However, it is essential to consider two factors. First, ABA is typically structured around long-term, cumulative learning processes, often requiring extended timelines to capture gains in high-order social cognition such as Theory of Mind. Our study focused on short-term effects within a 6- to 8-week window, which may not have been sufficient to detect the more gradual progress associated with ABA. Second, while ABA principles were systematically applied—such as individualized programming, consistent therapist oversight, and integration with structured supports like TEACCH (Treatment and Education of Autistic and related Communication-handicapped Children) and the CAT-kit (Cognitive Affective Training kit)—the outcomes assessed in this study centered on Theory of Mind (ToM), as measured by NEPSY-II scores. These improvements often emerge later in ABA trajectories. Our intent was not to compare robot-assisted training to ABA in terms of comprehensive efficacy, but rather to examine how short, targeted robotic interventions could complement established protocols by accelerating specific skill acquisition or enhancing engagement for technology-responsive children.It remains to be examined in future work, whether such short-term gains as those observed in our training persist over longer timescale – a timescale at which ABA has been proven efficacious.

Important to note is that the successful integration of the robot into a clinical setting highlights the importance of designing technological tools that complement, rather than replace, human expertise. Far from diminishing the role of clinicians, the robot acted as a supportive tool, enabling therapists to focus on diagnostic and adaptive intervention processes. This collaborative model aligns with emerging recommendations for incorporating robotic technologies into clinical practice24. The clinical staff expressed positive attitudes toward the robot, recognizing its potential to enhance therapeutic practice through standardized, replicable task execution. These outcomes support prior research on technology adoption in healthcare25, suggesting that humanoid robots can be effectively integrated into clinical settings when supported by appropriate training and infrastructure. However, it is important to highlight that our findings should be interpreted in the context of clinician-robot collaboration. While the robot provided consistency, predictability, and engagement, it was never autonomous in guiding therapy. Rather, it functioned as a clinical tool under professional supervision. This reflects a growing consensus in the field that robots are most effective when embedded within human-led practices. Surveys on public perception have also shown hesitancy toward fully replacing human therapists with robotic agents, emphasizing the importance of maintaining a human-centered model of care37]– [38. Our protocol aligns with this perspective by enhancing, rather than replacing, the therapist’s role.

In summary, this study demonstrated that robot-assisted training can effectively enhance ToM skills in children with ASD. By combining advanced robotic technologies with evidence-based scientifically-grounded clinical practices, we developed a scalable, replicable protocol based on ABA principles. The improvements observed in NEPSY-II scores, supported by clinical staff feedback, highlight the potential of integrating humanoid robots into therapeutic frameworks. The robot’s ability to act as a consistent, engaging, and adaptive partner underscores its unique value as a complementary tool in ASD interventions. Future research should continue exploring how robotic technologies can augment human-led therapy, fostering more comprehensive, individualized treatment plans for neurodiverse populations.

Limitations and future directions

Despite these promising results, several challenges remain. Widespread implementation of robot-assisted training will depend on factors such as scalability, financial feasibility, and sustained collaboration among clinicians, researchers, and technology developers. Future studies should examine not only long-term therapeutic effects and cross-cultural adaptability but also the cost-effectiveness of robotic interventions in real-world settings. Importantly, fostering clinician buy-in will require structured training and support to ensure that therapists feel confident and competent in incorporating robotics into their therapeutic practices. It is essential that future deployments of robot-assisted training remain grounded in clinician-led practice. Public concerns about automation in healthcare are not unfounded, and previous studies have documented resistance to the idea of robots replacing human therapists. For this reason, our protocol was intentionally designed to support therapists by providing structured, repeatable interaction formats—tools to be used within, not outside of, the clinical context. Sustained clinician involvement is not only ethically necessary but functionally integral to the successful implementation of such technologies.

Technological adaptability is another critical consideration. While the iCub robot successfully supported scripted, scenario-based interventions, some children required real-time adjustments or repeated exposures to specific tasks. These situations emphasized the importance of equipping future robotic systems with adaptive learning capabilities, enabling them to respond flexibly and more autonomously to diverse developmental profiles and dynamic clinical needs.

Methodologically, our study employed a two-period crossover design to compare Robot-Assisted Training (RAT) and Standard Therapy (ST), supplemented by a separate Human-Control (HC) group. This hybrid structure was chosen to balance scientific rigor with the practical and ethical considerations inherent in working with neurodivergent children. While a full three-condition crossover design—where each child undergoes all three treatments—might offer stronger internal validity, such an approach would significantly extend the intervention timeline. This could introduce fatigue, reduce engagement, and increase dropout risk, particularly in pediatric clinical populations already managing multiple concurrent therapies. Our decision to separate the HC group, while mirroring the RAT protocol as closely as possible, allowed us to isolate the robot’s contribution without overburdening participants.

Future research with larger and more diverse samples may be better positioned to implement more comprehensive within-subject designs, including multi-arm crossover or factorial trials. These would allow for direct comparisons across robotic, human-control (HC), and standard therapy (ST) conditions within a single study framework. Such work would also help identify which specific features—such as embodiment, consistency, or expressiveness—drive therapeutic gains. Understanding how different technologies can best support specific skills (e.g., emotional understanding, joint attention, perspective-taking) will be essential to advancing personalized, scalable neurodevelopmental interventions. While long-term outcomes were not the primary focus of this study, the crossover design provided some insight into short-term retention. Children who received robot-assisted training during the first phase maintained their gains at the second evaluation, suggesting some stability in acquired Theory of Mind skills. However, future studies should include delayed post-tests to more robustly assess the durability of these improvements over time.

Another important direction for future research involves exploring the adaptability of this protocol to other robotic platforms. Although we used the iCub robot—a research-grade humanoid with advanced motor capabilities—we acknowledge that iCub is not yet widely available in clinical settings. Commercially available robots that support physical interaction and exhibit simplified anthropomorphic features could offer more scalable alternatives. Investigating whether these systems can elicit similar therapeutic outcomes will be critical to assessing the protocol’s generalizability across different contexts and technological infrastructures.

Lastly, although ABA-informed therapy was used as a standard of care in our clinical setting, the duration and scope of our study were not designed to assess the full impact of long-term behavioral interventions. Rather, our aim was to evaluate whether short, targeted, robot-assisted activities could enhance hypothesized outcomes—Theory of Mind—when integrated alongside ongoing therapy. Future longitudinal work is needed to explore how such targeted, technology-supported modules might reinforce broader therapeutic frameworks or accelerate progress in selected cognitive domains.

Materials and methods

This trial was prospectively registered with the ISRCTN registry (ISRCTN15341724) on May 6, 2025, before the enrollment of participants. The registration record is publicly accessible at: https://www.isrctn.com/ISRCTN15341724.

Participants

The study included 32 children with a formal diagnosis of Autism Spectrum Disorder (ASD), confirmed using the Autism Diagnostic Interview–Revised (ADI-R) and the Autism Diagnostic Observation Schedule, Second Edition (ADOS-2). These participants were recruited from the Boggiano Pico Center in Genova, a clinical facility affiliated with Don Orione Italia and Piccolo Cottolengo Genovese di Don Orione. The only inclusion criteria were: (1) the child’s ability to understand the instructions of the training games, and (2) the family’s willingness to participate. This inclusive approach ensured ecological validity while maintaining relevance to real-world clinical practice. The mean age of the participants was 7.53 years (± 1.32), and 7 of the children were female. In addition to the 32 children participating in the main crossover trial, an independent group of 14 children (mean age = 7.42 ± 1.70 years, 2 females) was recruited for the Human-Control (HC) condition. These participants were drawn from a different clinical institution, the Philos Academy in Genova, Italy, and received a structured, clinician-delivered intervention based on the same protocol used in the robot-assisted sessions. While children in the HC group later participated in robot-assisted training sessions to ensure equitable clinical access, only data from the human-led phase were analyzed to assess the protocol’s efficacy independently of the robot. This analytical decision was made to avoid overparameterizing the statistical model and introducing carryover or novelty effects, especially given the relatively small sample size. We opted to structure the HC condition as a separate group rather than integrate it into the main crossover design in order to reduce the burden on participants, many of whom were already engaged in multiple weekly therapies. This approach preserved analytical clarity while minimizing practical and ethical strain on families.

Informed consent

was obtained from all parents or legal guardians, who were fully informed about the study’s objectives and methodology. The inclusion criteria were non-discriminatory, ensuring a broad representation of children with ASD regardless of their race, ethnicity, culture, gender, or socioeconomic status. Specific demographic data on racial/ethnic identification and socioeconomic background were not collected, as these were not relevant to the aims of the study.

The sample size was determined based on previous studies employing similar methods and measures, ensuring sufficient power to detect differences between the experimental and control conditions. Ethical approval for the study was obtained from the Comitato Etico Regione Liguria, and the study was conducted in full accordance with the principles outlined in the Declaration of Helsinki, ensuring the rights, dignity, and well-being of all participants.

Standard therapy (ST)

In the clinical practice where this study was conducted, each child with autism received an individualized intervention tailored to their specific needs and strengths. Due to the heterogeneity of ASD, a one-size-fits-all protocol is not feasible. In line with a multidisciplinary approach, the personalized rehabilitation plan typically included Applied Behavior Analysis (ABA), speech therapy, and psychoeducational interventions such as TEACCH (Treatment and Education of Autistic and related Communication-handicapped Children). Psychological support and occupational therapy were also commonly provided to enhance motor skills and daily living activities, along with structured socialization sessions to promote peer interaction and social skills development. All participants in the study engaged in Cognitive Affective Training (CAT-kit) protocols, which aimed to enhance perspective-taking and Theory of Mind (ToM). These interventions were delivered by a team of 22 certified therapists specializing in areas such as speech therapy and motor rehabilitation. Throughout the study, children receiving standard therapy utilized the Emotional Toolkit developed by Attwood, focusing on enhancing social and communication skills. ABA principles were consistently integrated to address individualized goals, with sessions conducted twice a week from T0 to T2, independent of the robot-assisted activities. Importantly, the ABA-based therapy in this study was conducted within a clinical center fully recognized by the Italian national healthcare system. The interventions were delivered by a multidisciplinary team of professionals—including psychologists, educators, speech therapists, and psychomotricity specialists—ensuring adherence to national and international standards of ABA practice. The inclusion of this standard-of-care ABA framework was not intended to serve as a direct comparator in terms of relative effectiveness, but rather to ensure that all children in the study were receiving consistent, structured therapeutic support from experienced professionals. The robot-assisted activities were introduced as an addition to, not a replacement for, this care—providing a focused, short-term intervention targeting ToM-related outcomes. With children taking part in the robot-assisted activities, the training was delivered twice weekly over a 6-to-8-week period, depending on individual availability. Each session lasted between 10 and 15 min., with duration tailored to the child’s capacity for engagement.

Robot-Assisted training (RAT)

In this study, we utilized the iCub humanoid robot, noted for its anthropomorphic design and capability for interactive behaviors26. Our protocol involved structured interactions designed to mimic everyday social situations, aiming to enhance Theory of Mind (ToM) skills in children with Autism Spectrum Disorder (ASD). Each session comprised several role-play trials that were conducted in familiar environments such as restaurants and cinemas, environments where social interactions typically occur (see Fig. 3 for session setups).

Fig. 3
figure 3

The figure illustrates two examples of the setups employed for the activity. Stickers and backgrounds were altered in each session to introduce novel stimuli, thereby mitigating potential boredom among the children throughout the sessions.

Full size image

Initially, the iCub robot demonstrated specific social behaviors—making requests, initiating greetings, and responding to questions—that the children were encouraged to observe and later mimic. The robot displayed various social gestures aiming to promote engagement in joint attention. This approach facilitated an active learning process by allowing the children to engage in both observation and direct interaction, with role reversals integrated into the sessions to enable children to practice and apply learned behaviors actively (see Fig. 4 for the trial sequence).

The training took place in a specially designed simulation room that resembled a mini theater, equipped with props and backdrops relevant to each scenario.

Fig. 4
figure 4

The figure depicts a flowchart representing the various phases of the role-playing game carried out by the child and the robot. On the left (A), an example of the phase where the child is asked to respond to a behavioral request, which is executed by the robot and repeated if the child fails to respond; if successful, the robot reinforces the response with positive feedback (i.e. the robot raises its arms and hands in a gesture of victory, reinforcing the child with verbal prompts such as ‘let’s keep it up’ or ‘you’re doing great’). On the right (B), an example of the phase where the child is asked to perform a behavioral response, which may be prompted by the robot. If the child immediately complies with the request, the trial proceeds and the robot promptly responds, otherwise, the prompt to execute the request is given a second time.

Full size image

This immersive environment was optimized to minimize distractions and maximize engagement, making use of visual and auditory cues to guide the children through the role-play activities effectively. To enhance realism and engagement, the iCub was programmed to perform a variety of human-like gestures and facial expressions synchronized with verbal interactions. These included eye movements, head tilts, and hand gestures, which were all crucial for creating a believable and interactive social agent.

Throughout these sessions, the effectiveness of the interactions was continuously evaluated through direct behavioral observations and feedback from the children. Real-time adjustments were made to the robot’s behavior and the complexity of the scenarios based on this ongoing assessment, ensuring the training was tailored to the individual needs and progress of each child. These robot-assisted sessions represent a significant step forward in therapeutic interventions for ASD, combining technological precision with the flexibility required for effective personalized therapy. By methodically integrating the iCub robot into structured social training, we aim to provide a replicable and impactful model that can significantly improve social cognition in children with ASD.

Cross-over design

Our study followed a two-period crossover design, in line with CONSORT guidelines, to ensure transparent and unbiased reporting of randomized controlled trials. A total of 32 participants were assessed for eligibility, and all were randomized into two groups. Group 1 initially received robot-assisted training from baseline (T0) to the first assessment (T1), while Group 2 began with standard therapy during the same period. After the first assessment, the groups switched treatments, allowing all participants to experience both robot-assisted and standard therapy by the second assessment point (T2). This crossover design minimized potential confounding effects from the treatment order, supporting a robust comparative analysis of the interventions’ effectiveness. Throughout the study, evaluators remained blinded to which treatment condition participants were assigned to maintain the integrity of the results. Random allocation was conducted manually by blinded clinical staff to ensure fairness, avoiding automated randomization methods. Importantly, no participants were excluded, lost to follow-up, or discontinued from the study. This was likely due to the participants’ regular attendance at the clinical center for ongoing therapy. Careful monitoring throughout the study helped prevent bias, with the clinical staff responsible for enrolling participants and maintaining blinding during treatment assignment. The trial ran smoothly with no dropouts, ensuring complete data for analysis across both groups.

The two groups consist of individuals with similar demographic characteristics. Group 1 has an average age of 7.25 years (± 1.34) and includes 13 males and 3 females, while Group 2 has an average age of 7.81 years (± 1.28) with 12 males and 4 females.

Human-control group

A comparable protocol, closely mirroring the robot-assisted intervention, was administered to an independent cohort of children (Human-Control group), involving an expert clinician who replaced the robot as the social interaction partner. The clinician received specific training to emulate the sequence, structure, and behavioral cues used in the robot-assisted sessions—such as initiating greetings, prompting turn-taking, and providing verbal feedback—while conducting the same role-play scenarios in simulated environments (e.g., cafés, cinemas).

However, due to the dynamic nature of live human interaction and the clinical context, the clinician was not expected to mimic the exact demeanor of the robot. Instead, clinicians were encouraged to remain responsive to each child’s individual needs—providing additional support, clarification, or reinforcement when necessary. This approach balanced fidelity to the structured training design with the flexibility required for real-world clinical engagement. In line with principles of Applied Behavior Analysis (ABA), the goal was to preserve consistency in task structure while avoiding undue discomfort or rigidity that could interfere with child engagement.

The decision to conduct this human-assisted condition as a separate experiment—rather than as part of the crossover design—was guided by both scientific and ethical considerations. First, it allowed us to evaluate whether the specific therapeutic benefits observed were attributable to the robot itself, beyond the novelty of its presence or the structure of the training protocol. The HC group thus controlled for the protocol content alone, allowing us to test whether the combination of structured training and robotic embodiment produced outcomes superior to the protocol delivered by a human. Second, incorporating a full within-subject design comparing all three conditions (ST, RAT, and HC) would have required substantially more participants and longer testing timelines to ensure statistical power and mitigate fatigue, novelty, and carryover effects—challenges especially acute in pediatric ASD populations already enrolled in multiple therapies. To avoid overburdening participants, we opted to implement the HC group as a parallel cohort.

It is important to note that introducing a robot into any clinical environment inherently alters the therapeutic context. Its novelty, physical presence, and interactive behavior can elevate attention or engagement, potentially confounding comparisons with standard therapy. Thus, a comparison with passive controls alone would not be sufficient to isolate the robot’s specific contribution. This motivated our inclusion of an active control condition, structured to match the robot-assisted training as closely as possible, minus the robot’s presence.

Finally, to ensure ethical fairness, participants in the HC group were invited to engage in robot-assisted training after completing the NEPSY-II assessments, thereby allowing them access to the full intervention without introducing contamination into the outcome data. This design maintained the integrity of the control comparison while aligning with the broader clinical goal of equitable access to potentially beneficial treatments.

Measures of interest

Throughout the study, the social cognition and neuropsychological development of participants were assessed using the NEPSY-II test27, a comprehensive tool designed for children aged 3 to 16 years. This tool evaluates various cognitive domains through 32 subtests, including Attention, Executive Functioning, Language, Memory, Sensorimotor, Social Perception, and Visuospatial Processing. Our primary focus was on the Social Perception domain, particularly the Theory of Mind (ToM) subscale, which measures the ability to attribute mental states to others—a crucial skill often challenged in children with ASD. The ToM subscale includes tasks such as identifying characters’ beliefs or intentions based on short illustrated stories and answering perspective-taking questions about others’ knowledge, thoughts, or feelings. These structured, age-normed tasks are designed to assess core components of social cognition in a standardized, clinically valid format.

The NEPSY-II test was administered at three key points: baseline (T0), after the first treatment period (T1), and at the study’s conclusion (T2), as depicted in Fig. 3, which illustrates the timing of assessments. These assessments were carried out by qualified personnel blind to the participants’ treatment conditions, ensuring unbiased results. The raw scores obtained were standardized according to age-specific norms provided in the NEPSY-II manual, producing scaled scores ranging from 1 (lowest performance) to 19 (highest performance). This standardization allows for accurate comparisons of individual performance over time and across different treatment conditions.

Analyses

Our analysis utilized a linear mixed-effects model (LMM) to assess the impact of the interventions on NEPSY-II scores, considering both fixed effects (robot-assisted vs. standard therapy) and random effects (individual differences). This model, suitable for the repeated measures design used in our study, was executed using the lme4 package in R28, allowing for an evaluation of the progression in social cognition as measured by changes in NEPSY-II scores between baseline (T0), mid-study (T1), and study conclusion (T2).

To ensure robustness in our findings, residuals were checked for normal distribution and homogeneity of variance through Q-Q plots and residual-fitted plots. Model validity was confirmed by adhering to diagnostics recommended by29. Additionally, the Satterthwaite approximation was used to adjust p-values for degrees of freedom30. The main analytical focus was on the delta (Δ) values between successive evaluations (T1 – T0 and T2 – T1), indicating improvements or regressions in participant scores. Differences in improvement rates between the two treatment conditions were further examined using an independent samples t-test, with effect sizes reported as Cohen’s d to quantify the strength of the intervention effect. These comprehensive analytical steps ensured the reliability and validity of the results, providing a clear indication of the efficacy of the interventions in enhancing Theory of Mind capabilities among participants.

Community involvement statement

In the design and implementation of this study, we actively involved clinicians who specialize in Autism Spectrum Disorder (ASD) to ensure that the protocol was both clinically relevant and practical. These clinicians contributed to the development of the robot-assisted training activities, helping to tailor the interventions to the specific needs of the children. Additionally, we conducted meetings with the families of all participants, both collectively and individually, to ensure that they were fully informed about the study’s objectives and procedures. These meetings provided an opportunity to address any questions or concerns the families had, thereby fostering an environment of transparency and collaboration. Our primary goal was to engage the children in the robot-assisted activities without disrupting their standard therapeutic routines. We ensured that the interventions were seamlessly integrated into the existing therapy schedules, allowing the children to benefit from the additional support without interference with their ongoing treatments.