Robots in Autism Therapy: What the Evidence Actually Says

Two decades of studies suggest children often light up around robots, but the hard evidence is thinner than the headlines, and autistic researchers are asking whether the field has been solving the right problem.

The founding hunch

Autism was one of the very first places social robotics went looking for a real job. The intuition was appealing: if human social signals are fast, subtle and sometimes overwhelming, perhaps a machine that offers simpler, slower, more predictable cues could become a comfortable bridge. Many early studies seemed to bear this out, with videos of children who rarely engaged with adults suddenly laughing, pointing and taking turns with a small robot.

In 2012 Brian Scassellati, Henny Admoni and Maja Matarić took stock of a decade of this work in the Annual Review of Biomedical Engineering. Their verdict was encouraging but notably cautious. Robots did seem to attract attention and draw out social behaviour that therapists found interesting. Yet most studies were descriptive case reports involving only a handful of children, and the authors stressed that nobody really knew why the effect happened. They listed several candidate explanations: robots might be less overstimulating, more predictable, free of the negative associations some children have with human interaction, or simply more exaggerated in their social prompts. Understanding the mechanism, they argued, was perhaps the most important open question for the field.

Taking the robot home

One persistent criticism was that lab sessions are short, artificial and novel. In 2018 Scassellati and colleagues at Yale published a study in Science Robotics that tried to address this by leaving an autonomous robot in family homes for a month. Twelve children aged between six and twelve sat down with a caregiver and the robot for about thirty minutes a day, working through games on emotional storytelling, perspective taking and sequencing. The robot adjusted the difficulty of the activities to each child's past performance.

The results were promising. Engagement held up across the month, children showed better joint attention with adults in assessments where the robot was not present, and caregivers reported needing to prompt less and seeing more communication overall. That last point matters, because the whole point of an intervention is what happens when the machine is switched off. Still, twelve children, no control group and a one-month horizon add up to a feasibility study rather than proof of a treatment.

Pooling the trials

What happens when you gather the randomised trials together? Athanasia Kouroupa and colleagues did exactly this in a 2022 systematic review and meta-analysis in PLOS ONE. They found forty eligible studies, of which seventeen were randomised controlled trials, and twelve of those could be pooled. For social outcomes, robot-mediated interventions produced a small but statistically significant benefit, with an effect size of around a third of a standard deviation and no sign of disagreement between trials. Emotional and motor outcomes showed no reliable effect, though only a few trials measured them. Younger children appeared to benefit more.

The caveats are as instructive as the headline. Outcome assessors were unblinded in the vast majority of studies, follow-up periods were short, adverse events were rarely reported and most research came from the US and Europe. In other words, the evidence is real but modest, and it is built on studies that would struggle to satisfy a demanding clinical reviewer.

The critique from inside the community

A different kind of question has come from autistic researchers themselves. In a critical review presented at CHI 2024, Naba Rizvi and colleagues analysed 142 papers on HRI and autism published between 2016 and 2022. They argued that much of this work treats autism primarily as a deficit to be corrected, relies on narrow assumptions about gender and age, and builds in power imbalances between researchers and participants. By their count, roughly nine in ten studies did not include the perspectives of autistic people. A 2024 metareview by Hundt and colleagues made a related point, contrasting the field's focus on interventions that aim to make autistic people appear more typical with what autistic people actually say they want.

This is not a fringe complaint. If a robot is used to train a child to hold eye contact, for example, one has to ask who benefits and at what cost to the child, a question that the neurodiversity movement has raised about many conventional therapies as well. The critics are not saying robots can never help. They are saying the goals should be set with autistic people rather than for them.

Where that leaves us

A fair summary is that robots reliably capture interest, can sustain engagement over weeks, and appear to produce small gains in some social skills, especially in younger children. They are not a proven standalone therapy, and the best studies so far are small, short and hard to blind. The most interesting future work will probably look less like robots drilling children in neurotypical behaviour and more like tools that autistic people help design, for goals they choose, evaluated in trials large and long enough to tell us something definitive.

For a field that began with a hunch about why children respond to machines, that is an oddly fitting place to end up: still asking why, but now also asking whom the answer is for.

References

1. Scassellati, B., Admoni, H. and Matarić, M. (2012). Robots for Use in Autism Research. Annual Review of Biomedical Engineering, 14, 275-294. Link
2. Scassellati, B., Boccanfuso, L., Huang, C.-M., Mademtzi, M., Qin, M., Salomons, N., Ventola, P. and Shic, F. (2018). Improving social skills in children with ASD using a long-term, in-home social robot. Science Robotics, 3(21). Link
3. Kouroupa, A., Laws, K., Irvine, K., Mengoni, S., Baird, A. and Sharma, S. (2022). The use of social robots with children and young people on the autism spectrum: A systematic review and meta-analysis. PLOS ONE. Link
4. Rizvi, N., Wu, W., Bolds, M., Mondal, R., Begel, A. and Munyaka, I. N. S. (2024). Are Robots Ready to Deliver Autism Inclusion?: A Critical Review. Proceedings of the CHI Conference on Human Factors in Computing Systems. Link
5. Hundt, A. et al. (2024). Love, Joy, and Autism Robots: A Metareview and Provocatype. arXiv preprint. Link

Feature drafted with AI from the published research listed above, for social-robotics.com.