At a glance
Social robots like Kaspar, NAO and QTrobot are now tested against standard therapy for autistic children. What the trials, costs and experts actually show.
A robot helped a boy at a South Carolina elementary school hold a back-and-forth conversation for the first time in five years. The remarkable part isn't the machine. It's what the machine didn't do. It never got tired, and it never let frustration change its tone when the boy looked away.
That consistency is the quiet engine behind two decades of research into whether robots can help autistic children build communication skills. Autism spectrum disorder (ASD) is a lifelong developmental condition that can make social interaction and both verbal and non-verbal communication harder to read and harder to produce. The World Health Organization estimates that about one in 100 children worldwide has autism. The US Centers for Disease Control and Prevention puts the figure at one in 36 eight-year-olds.
For a child who finds faces unpredictable, a robot offers something rare: the same face and the same predictable reaction, every single time. That is why the field keeps returning to machines. The question worth asking now isn't whether a child will talk to a robot. It's whether any of it sticks once the robot leaves the room.
Why a machine, of all things
The idea is older than it sounds. Researchers date the first known therapeutic use of a robot with an autistic child to 1976. The modern field began in the late 1990s with the Aurora project, and out of that lineage came Kaspar, a child-sized humanoid built at the University of Hertfordshire and the first social robot designed specifically for autism. Kaspar's face is deliberately stripped of fine detail, so nothing pins down an age or a mood, and a human operator controls its speech and movements from behind the scenes. Researchers call that setup Wizard of Oz.
The robots in these studies are social robots, sometimes called socially assistive robots (SARs): machines built to interact through speech and gaze rather than to perform a physical task. Why children respond to them is still partly hypothesis, but the leading explanation is blunt. Robots are social without being too social. A robot makes eye contact and takes turns without the weight of a person's expectations, and it does not bring the sensory and emotional noise a person carries into the room.
The behaviour researchers measure supports that story. In a Dutch special school study, nine children maintained attention longer with Kaspar than with their teacher, and they reached out to touch the robot more often. In a Greek school, seven children across 111 sessions showed gains in prompted speech and focus. In Malaysia, 25 students improved on measures of social interaction and communication after just three sessions.
The strongest trials, read closely
Early enthusiasm now has a second, harder layer of evidence: randomized controlled trials that measure robots against standard therapy rather than against no treatment at all.
The most consequential arrived in 2025 in Science Robotics. Two trials, one with 69 children averaging 4.4 years old and another with 63 children averaging 5.9, tested robot-assisted therapy in clinics and in simplified school-and-home setups. The result was equivalence, not magic: children matched the outcomes of conventional therapy while showing significantly higher engagement, sometimes in fewer sessions. The two randomized trials suggest robots could extend care where specialist clinicians are scarce.
A single-blind randomized study in Italy sharpened the picture for minimally verbal children. Thirty-seven children aged 4 to 12 received either NAO-assisted therapy or standard speech therapy across 12 weekly 45-minute sessions. The robot group showed larger gains in verbal communication, and their spontaneous requests rose from an average of 6.8 per session to 16.7.
NAO, a 58-centimetre humanoid, is the workhorse of this research. A systematic review of 44 studies found consistent positive effects on communication, with the more recent studies from 2020 onward reporting larger gains than earlier ones.
The field's best-known home study comes from Yale, where 12 children used an autonomous robot called Jibo for 30 minutes a day over a month. Joint attention, the ability to share focus on an object with another person, improved by roughly a third, and caregivers reported less prompting and more communication. "The children showed improved performance across the board," lead author Brian Scassellati said at the time.
A 2025 meta-analysis of 14 studies and 408 children calculated a large overall effect of 0.829. One finding in it deserves more attention than the headline number: the effect shrank as teacher involvement decreased, and longer sessions were linked to better outcomes.
Photo by Igor Omilaev on Unsplash
Here's the catch
Generalization is the whole game, and it is the weakest part of the evidence.
Most gains fade when the robot leaves the room. In the Yale study, joint attention improvements began to slip within 30 days of the robot being removed. The average study in this field involves around a dozen children, runs for weeks, and uses a different robot and outcome measure from the one before it. A 2024 review concluded that robots improve behaviour in the short term but that longer-term evidence is still missing.
The equivalence finding cuts both ways. Matching standard therapy is a real achievement, but no trial has yet shown that a robot reliably beats a skilled human clinician. Some reported effects may also be novelty: children engage with a new machine because it is new, and that excitement can inflate early numbers.
Educators who would actually deploy these tools say it plainly. In a 2019 study of 31 autism education staff in England, teachers and speech therapists were receptive but wary, and one theme kept returning: children need to interact with people, not robots. The educators interviewed were not uncritical fans of the technology. The risk isn't that a robot fails. It's that a child comes to prefer the machine's perfect consistency to the messier, more demanding back-and-forth of a human conversation.
Then there's the price. The NAO robot used in many trials costs thousands of euros before anyone programs it, and a Milo package for a school district has run about $10,000.
Hype versus reality
Strip away the marketing and the scorecard looks like this.
- Engagement: strong. Across robots and countries, children consistently attend longer and initiate more with a robot than with a human stranger. This is the most replicated finding in the field.
- Measured communication gains: positive but modest. Trials show improvement, often equivalent to standard therapy, with the largest effects in longer programs and among children who are verbal or minimally verbal.
- Transfer to real people: unproven at scale. Individual cases are striking, the South Carolina boy's first conversation in five years, but controlled evidence of durable generalization is thin.
- Cost and access: weak. Hardware is expensive, funding is scarce, and most families cannot buy what their child's school or clinic might borrow.
The verdict is neither a miracle nor a failure. Robots are a useful adjunct, not a replacement. Nobody should sit a child alone in front of a robot and call it therapy.
The price problem
Demand keeps growing while the hardware stays expensive. The CDC counted one in 36 eight-year-olds in 2020, and the World Health Organization's estimate of one in 100 globally hides wide gaps between countries.
The clearest evidence that robots can work in everyday settings comes from a school rather than a lab. Rodney House School in Manchester has used QTrobot, a small robot with an expressive screen face, for several years, and the school was featured in a UNESCO report on inclusive education. Teachers there reported that two out of three children in one class went from saying little to saying hello and goodbye unprompted, and that attention spans stretched from a few minutes to 15 minutes within six weeks.
Real-world programs remain scattered. Ontario's Trillium Lakelands District School Board, which serves 16,000 students across 54 schools, brought in Milo to give children a consistent starting point. South Carolina once sent 21 Milos into its schools. Moberly, Missouri, spent about $10,000 for one robot plus curriculum. Each example is a pilot with a champion behind it. That pattern is the problem: scale depends on grant money and enthusiastic staff, not on a routine, reimbursable service.
Photo by Alex Knight on Unsplash
What happens next
Three changes would move this field from pilot to practice. Longer trials with shared outcome measures, so studies can actually be compared. Cheaper, simpler robots that a school can repair without calling a specialist, researchers have already built working prototypes for under $300. And clear rules on data privacy, because a robot in a child's bedroom is also a recording device.
None of that is exotic. It's study design and procurement, which is less glamorous than a dancing humanoid and far more likely to help a child.
Scassellati's explanation of why robots work is also the best argument for keeping expectations modest. Robots are "social enough to get people to respond to them, but they're not too social that they provoke any anxiety." A tool that borrows just enough social presence to open a door, then steps aside for the humans who should walk through it. That's the whole promise, and the whole caution, in one sentence.
- child-development
- social-robots
- assistive-technology
- autism-communication
- robot-assisted-therapy

