Online Affect Detection and Adaptation in Robot Assisted Rehabilitation for Children with Autism

This paper presents a novel affect-sensitive human-robot interaction framework for rehabilitation of children with autism spectrum disorder (ASD) where the robot can detect the affective cues of the children implicitly and response to them appropriately. Psychophysiological analysis is performed that uses subjective reports of the affective states from a clinical observer. Comprehensive physiological indices are investigated that may correlate with the affective states of children with ASD. A robot uses a support vector machines based affect model to detect the affective cues. A reinforcement learning based adaptation mechanism is employed to allow the robot to adjust its behaviors autonomously as a function of the predicted children's affective state. Four adolescents diagnosed with high-functioning autism participated in the experiments. This is the first time, to our knowledge, that the affective states of children with ASD have been detected via physiology-based affective modeling technique in real-time. This is also the first time that impact of affect-sensitive interaction between a robot and children with ASD in closed loop has been demonstrated experimentally.