Abstract
Early detection of Autism Spectrum Disorder (ASD) is critical for effective intervention, yet clinical diagnoses often occur years after symptoms first emerge. This research proposes a standardized screening protocol using the NAO humanoid robot to elicit social behaviors in toddlers. By pairing these interactions with automated computer vision analysis, the system provides clinicians with objective, quantifiable data on social cues such as joint attention and imitation. Preliminary findings suggest that while robotic interactions offer a consistent stimulus for screening, protocols must be carefully tailored to the developmental age of the child.
Introduction
Early intervention is widely recognized as the "gold standard" for treating ASD, with the most significant developmental gains occurring when children are identified by their second year. Despite this, the average age of diagnosis in many regions remains as high as five years. Current diagnostic tools, such as the Autism Diagnostic Observation Schedule (ADOS), rely on semi-structured play and subjective human observation, which are time-intensive and prone to variability.
This research seeks to bridge the gap between clinical needs and robotic capabilities by creating a reproducible, portable protocol. By integrating robotics into clinical workflows, the study aims to provide a "flagging" system that assists clinicians in processing video footage more efficiently and identifying symptoms earlier.
Methodology
The core of the proposed system is the NAO robot, a 2-foot-tall humanoid with 25 degrees of freedom. The robot is programmed to lead toddlers through three primary activities designed to elicit specific social cues:
- "Simon Says": Targets imitation skills through actions like arm flapping or hand clapping.
- "I Spy": Focuses on joint attention and object recognition.
- Song and Dance: Encourages reciprocal social behavior and general engagement.
Data Capture and Analysis
The system utilizes three synchronized video streams: a close-up of the child from the robot's shoulder, a wide-angle room view, and a first-person view from the robot’s internal cameras. To analyze this data, the researchers employ computer vision algorithms—specifically Histograms of Oriented Gradients (HOG) and Support Vector Machines (SVM)—to automatically detect "target symptoms" such as repetitive body movements, asymmetric gait, and eye contact avoidance.
Study Design
The ongoing study at the University of Minnesota targets a sample of approximately 150 subjects. The initial cohort includes 30 participants from the Autism Spectrum and Neurodevelopmental Disorders (AS/NDD) clinic. Longitudinal tracking allows for the comparison of child-robot interaction data over months and years, providing a robust dataset for validating the automated detection system.
Key Insights and Preliminary Results
A central premise of this work is the "Robot Attraction" phenomenon: children with ASD often find robots more engaging and less intimidating than human partners. This allows the robot to elicit behaviors that might be suppressed in standard human-to-human assessments. Furthermore, the robot provides a perfectly consistent stimulus, reducing the experimental noise inherent in human-led observations.
Preliminary results have highlighted the importance of developmental calibration. Initial visits indicated that standard NAO routines were occasionally too complex for 2-year-olds, suggesting that robotic interactions must be tiered based on developmental age. Technically, the computer vision stack has shown promise in classifying symptoms like arm asymmetry and visual tracking, building on previous successful implementations of Object Cloud Models.
Limitations and Future Directions
The current iteration of the program is "static," following a pre-programmed script regardless of the child's reactions. Future work will focus on making the robot more dynamic, such as changing eye LED colors or repeating games based on child interest. Additionally, addressing verbal barriers in toddlers (ages 2–5) is a priority, requiring robust "default behaviors" for non-verbal responses. The ultimate goal remains refining the "flagging" logic to minimize false positives and negatives, ensuring the tool serves as a reliable assistant to clinical experts.
Summary updated