Autism Characteristics, Adaptive Skills, and Performance in an Emotion-Recognition Teaching Programme with a Humanoid Robot

Background: Several existing projects have shown promise in using robot-assisted interventions for teaching social and academic skills to autistic children, including emotion recognition. Most such work, however, has included older, so-called ‘high functioning’ autistic children and has therefore neglected to address the feasibility of robot-assisted emotion interventions for young autistic children, those with intellectual disability, or limited spoken language. Given claims that robot-assisted interventions could present lower, less complex social demands than human-led interventions, it is particularly important to investigate their feasibility for children whose social, daily life, or language skills may present barriers to participation in “traditional” interventions.

Objectives: This project tested the feasibility of an emotion-recognition training programme in developing the potential of robot-assisted interventions for a large sample of autistic children in Serbia and in the UK.

Methods: 128 autistic children aged 5-12 years were assessed in a 6-step emotion-training programme based on “Teaching Children with Autism to Mind Read” (Howlin, Baron-Cohen, & Hadwin, 1999). In the robot-assisted condition (n=64, 10 females), a Robokind R25 humanoid robot (“Zeno”; see Figure 1) helped to deliver the programme controlled covertly by the adult, whilst 64 children (9 females) participated in the therapist-assisted comparison condition. Parents completed the Vineland Adaptive Behaviour Scales – 2^nd edition (Sparrow et al., 2005), and researchers completed the CARS2-ST (Schopler et al., 2010) based on direct observation and parent information.

Results: Children in the robot- and therapist-assisted conditions were closely matched on autism severity and adaptive behaviours (Table 1). Emotion recognition performance, operationalised as the proportion of correct answers in steps 1-2 of Howlin et al.’s emotion-training programme, was inversely correlated with CARS-2 scores in both the robot-assisted (r = -.40, p < .001) and therapist-assisted conditions (r = -.47, p < .001). Higher Vineland composite scores (i.e. better adaptive behaviours) were significantly positively correlated with task performance for children in the therapist-assisted (r = .38, p < .01), but not the robot-assisted condition (r = .16, p = .27). Using the Vineland communication subscale as a proxy for language skills, better language was positively associated with performance in both conditions, but more strongly for the children in the therapist-assisted (r = .50, p < .001) than in the robot-assisted condition (r = .28, p < .05).

Conclusions: This study examined the relationship between everyday adaptive behaviours and emotion recognition performance within robot-assisted and therapist-assisted activities, as part of establishing the feasibility of using such interventions with a younger and less able participants than those studied thus far. Results suggest that while it is generally feasible for this group to participate, individual performance moderately correlated with lower CARS-2 scores and higher adaptive behaviour and language scores, but less so in the robot-assisted condition. This pattern may support claims that robots impose lower social demands, or are less complex to work with than a person. This study was the first phase of data collection for a large-scale project, with further studies iteratively developing a new robot-assisted emotion teaching intervention.