A Child-Aware Setup for Automatic Evaluation of Interaction with Robots in NAO Autism Therapy

The use of social robots in autism therapies has shown remarkable growth, attributable to the potential of robots to increase engagement with pediatric patients. This study presents an experimental setup with a well-defined procedure to evaluate the level of interaction between children with autism spectrum disorder (ASD) and NAO robots. The experimental setup consists of a robot placed in front of the patient and remotely controlled by a human therapist. The proposed methodology allows the child's proximity to the robot to be measured quantitatively using a depth camera and Computer Vision techniques to calculate the distance and direction of patient's gaze. In particular, we use YOLOWorld in open-vocabulary scenarios to enable object identification and integrate it with Gaze-LLE, a visual foundation model for estimating gaze targets. The system's outcome can be used as a proxy indicator of the patient's engagement, by exploiting the distance and direction of gaze between the child and objects used in therapy. Experiments have been carried out on two children, in the context of the Stella Project carried out in partnership with the Early Start Center in Chiaromonte (Italy) afferent to the Stella Maris Mediterraneo Foundation. The results show the effectiveness of the approach presented through qualitative analysis. The code of our approach is publicly released to the community at https://github.com/michelebri/gaze-estimation