Could Interaction with Social Robots Facilitate Joint Attention of Children with Autism Spectrum Disorder?

This research addressed whether interactions with social robots could facilitate joint attention of the autism spectrum disorder (ASD). Two conditions of initiators, namely ‘Human’ vs. ‘Robot’ were measured with 15 children with ASD and 15 age-matched typically developing (TD) children. Apart from fixation and gaze transition, a new longest common subsequence (LCS) approach was proposed to analyze eye-movement traces. Results revealed that children with ASD showed deficits of joint attention. Compared to the human agent, robot facilitate less fixations towards the targets, but it attracted more attention and allowed the children to show gaze transition and to follow joint attention logic. This results highlight both potential application of LCS analysis on eye-tracking studies and of social robot to intervention.

💡 Research Summary

This study investigated whether interactions with a social robot can facilitate joint attention in children with autism spectrum disorder (ASD). Fifteen ASD children (average age ≈ 5 years) and fifteen age‑matched typically developing (TD) children were presented with two types of initiators: a human actor and a NAO humanoid robot. In each condition, a short video showed the initiator greeting the participant and then turning its head toward one of three toy trucks placed on a table, thereby attempting to elicit joint attention toward the target object. Eye movements were recorded with a Tobii X3‑120 eye‑tracker (120 Hz). Areas of interest (AOIs) included the initiator’s face, body, each toy, a framing rectangle, and the background.

Fixations were defined as gaze points within 1.5° for at least 100 ms. Traditional metrics—first fixation latency and fixation duration percentage for each AOI—were computed. In addition, the authors introduced a novel analysis based on the Longest Common Subsequence (LCS) algorithm. They constructed an ideal “joint‑attention sequence” (e.g., face → target → background) and measured the similarity between each participant’s actual gaze sequence and this ideal using LCS, yielding a score from 0 (no similarity) to 1 (perfect match). This approach captures the dynamic, logical flow of attention rather than static dwell times.

Statistical analysis employed repeated‑measures ANOVA with factors Group (ASD vs. TD) and Initiator (Human vs. Robot). Key findings were:

1. Fixation distribution – ASD children spent less time fixating on the target toy in the robot condition than in the human condition (≈ 12 % vs. ≈ 18 % of total screen time). However, they showed relatively higher fixation on the robot’s face and body compared with the human, suggesting that the robot captured visual interest but was less effective at directing gaze to the intended object.

2. Gaze transitions – The number of transitions between AOIs was significantly higher for ASD participants in the robot condition than in the human condition, indicating that the robot promoted more active scanning and shifting of attention.

3. LCS similarity – The LCS scores for ASD participants were higher in the robot condition (≈ 0.68) than in the human condition (≈ 0.53). TD children achieved high LCS scores in both conditions, with the human condition yielding the highest similarity. Thus, the robot helped ASD children follow the prescribed logical sequence of joint‑attention cues more closely than a human did.

Overall, the robot did not increase raw fixation on the target object, but it enhanced overall attentional engagement, encouraged more frequent gaze transitions, and facilitated a more orderly, logic‑driven pattern of joint attention as quantified by the LCS metric. The study demonstrates two important contributions: (1) social robots can serve as effective scaffolds for joint‑attention training in ASD, especially by structuring the dynamics of visual attention, and (2) the LCS‑based dynamic eye‑tracking analysis provides a powerful tool for evaluating complex attentional behaviors beyond traditional fixation metrics.

Limitations include the modest sample size, potential confounds from differences in voice, facial expression, and the fact that stimuli were presented on a screen rather than in a fully embodied interaction. Future work should expand participant numbers, compare multiple robot platforms, and explore longitudinal effects of robot‑mediated joint‑attention training on language and social development. Nonetheless, the findings suggest that integrating humanoid robots into early intervention programs, coupled with sophisticated dynamic eye‑tracking analyses, holds promise for improving joint‑attention skills in children with ASD.