Designing a LLM-driven Avatar System to Enhance Social Skills for Autistic Children in DTT Learning

Autism Spectrum Disorder (ASD) is a psychological condition with unknown causes, challenging treatment, and an increasing prevalence rate. A typical characteristic of this disorder is impaired social interaction or communication. In current autism treatments, there are significant challenges and obstacles, such as a shortage of professionals, long treatment cycles, and substantial financial costs. To address these issues, we design a avatar system for individual training sessions for autistic children. We aim to reduce economic burdens while helping autistic children develop core social skills—specifically, self-introduction—while also alleviating the workload of autism therapists and improving treatment efficiency. In this paper, we employ the Discrete Trial Teaching (DTT) method, which is the most commonly used technique in Applied Behavior Analysis (ABA). We enable our avatar figures with the rapidly advancing large language models (LLMs) and speech services in artificial intelligence, to create a comprehensive and high-performance individual training session for teaching self-introduction to children with autism. Our evaluation results show that this system is fully functional with excellent performance. We also plan to explore future improvements including a comprehensive facial expression dataset for children with autism and applying the system into clinic practices.