Notwithstanding intensive research and many scientific advances, diagnosing autism spectrum disorders remains a slow and tedious process. Due to the absence of any physiological tests, the outcome depends solely on the expertise of the clinician, which takes years to acquire. Complicating the matter further, research has shown that inter-rater reliability can be very low, even among experienced clinicians. As an attempt to facilitate the diagnostic process and make it more objective, this paper proposes a robot-assisted diagnostic protocol. The expected benefit of using a robot is twofold: the robot always performs its actions in a predictable and consistent way, and it can use its sensors to catch aspects of a child's behavior that a human examiner can miss. In this paper, we describe four tasks from the widely accepted ADOS protocol, that have been adapted to make them suitable for the Aldebaran Nao humanoid robot. These tasks include evaluating the child's response to being called by name, symbolic and functional imitation, joint attention and assessing the child's ability to simultaneously communicate on multiple channels. All four tasks have been implemented on the robot's onboard computer and are performed autonomously. As the main contribution of the paper, we present the results of the initial batch of four clinical trials of the proposed robot assisted diagnostic protocol, performed on a population of preschool children. The results of the robot's observations are benchmarked against the findings of experienced clinicians. Emphasis is placed on evaluating robot performance, in order to assess the feasibility of a robot eventually becoming an assistant in the diagnostic process. The obtained results indicate that the use of robots as autism diagnostic assistants is a promising approach, but much work remains to be done before they become useful diagnostic tools.

Autism spectrum disorders (ASD) is a term used to describe a range of neurodevelopmental disorders affecting about 1% of the population, with increasing prevalence. Due to the absence of any physiological markers, diagnostics is based purely on behavioral tests. The diagnostic procedure can in some cases take years to complete, and the outcome depends greatly on the expertise and experience of the clinician. The predictable and consistent behavior and rapidly increasing sensing capabilities of robotic devices have the potential to contribute to a faster and more objective diagnostic procedure. However, significant scientific and technological breakthroughs are needed, particularly in the field of robotic perception, before robots can become useful tools for diagnosing autism. In this paper, we present computer vision algorithms for performing gesture imitation. This is a standardized diagnostic task, usually performed by clinicians, that was implemented on a small-scale humanoid robot. We describe the algorithms used to perform object recognition, grasping, object tracking and gesture evaluation in a clinical setting. We present an analysis of the algorithms in terms of reliability and performance and describe the first clinical trials.