Modeling of ASD/TD Children's Behaviors in Interaction with a Virtual Social Robot During a Music Education Program Using Deep Neural Networks

Researchers from Sharif University of Technology have developed a novel AI system that uses deep neural networks to analyze and model the behaviors of children with Autism Spectrum Disorder (ASD) and neurotypical (TD) children. The system operates within a virtual music education program where children interact with a social robot, capturing both impact data and motion signals. The model achieved an impressive 81% accuracy and 96% sensitivity in distinguishing between ASD and TD children based on their behavioral patterns during these interactions.

Beyond diagnosis, the team designed a transformer-based network capable of generating realistic simulated behaviors for both groups. In validation tests, expert clinicians could only differentiate between real and AI-generated behaviors with 53.5% accuracy—essentially random chance—demonstrating the model's high fidelity. This dual capability for both identification and simulation opens new avenues for therapist training, personalized intervention development, and deeper research into behavioral patterns associated with ASD.

The research, detailed in a preprint paper on arXiv, builds on previous work from the university's Social and Cognitive Robotics Laboratory. By creating a system that not only identifies complex behavioral signatures but also replicates them, the team provides a powerful tool for the autism research and therapeutic communities. This approach could lead to more accessible, scalable, and data-driven methods for understanding and supporting neurodiverse individuals.