Joint Attention Behavior for Children with Autism Spectrum Disorder (ASD) Interacted with Social Robots

Background: Joint attention has been the target of interventions of the autism spectrum disorder (ASD). Researches, however, did not yet find consistent evidence to support a joint attention deficit in the ASD population (Guillon, Hadjikhani, Baduel, & Rogé, 2014). Recently, social robots have been involved in the diagnosis and intervention of the ASD in social behaviors, language, imitation, and stereotyped behaviors, but robots seemed to reduce the joint attention behaviors when interacting with ASD population compared to the human agent. To address the problem whether interaction with social robots could improve joint attention of the autism spectrum disorder (ASD), we introduced a novel algorithm based on longest common subsequence (LCS) to quantify how the participant’s gaze dynamically follow the given logic of the videos, in order to further capture the dynamic feature of eye movement data collected from the joint attention tasks.

Objectives:

By utilizing a commercial humanoid robot NAO, this study compared the difference of joint attention behaviors induced by both the human and robot stimuli. Based on the analysis of fixation time, gaze transitions and LCS algorithm, the study could reveal both the static and dynamic process underlying the joint attention ability and compare the difference between the ASD and typically developing (TD) children.

Methods:

15 ASD (mean age: 4.96±1.10) and 15 age-matched TD children (4.53±0.90) participated in this eye- tracking study. Video stimuli were used to induce joint attention behaviors of participants. Each video consisted of an agent (human or robot) sitting behind the table together with three objects on the left, middle and right side of the table. The agent would turn the head to one of the three truck toys, and the eye movement was recorded by the Tobii X3-120 eye-tracker system. Besides the traditional fixation time and gaze transition analysis, we introduced an LCS algorithm to measure the complexity of transition distributions between area of interests (AOIs).

Results:

Data analysis showed that the TD group fixated longer on the agent face (p=0.002) than the ASD group, and the ASD group gazed on the frame (p=0.045) and non-target (p=0.036) longer than the TD group. All participants spent more time on face area in the “Robot” condition (p=0.038), and more on target (p=0.004) and non-target in the “Human” condition (p=0.045). However, we did not found any significant difference in gaze transition analysis between the ASD and TD group. The LCS analysis revealed a significant effect of group (F(1, 28)=11.18, p=0.002, partial η²=0.898), indicating that the TD participants have significantly higher LCS scores than the ASD. It implied the TD could better follow the logic of the video.

Conclusions:

This research showed a complex joint attention profile of the participants with ASD. Compared to the human agent, the robot attracted more interests of all participants, but was not able to facilitate the gaze transitions towards the targets or following the stimuli’s logic. It implies a negative impact on joint attention behavior when children were interacted with social robots.