Designing Social Robots for Social-Cognition Training with Autistic Adults

For many adults on the autism spectrum, the transition from school to the wider world brings a specific kind of exhaustion. While their technical skills or intellectual abilities may be sharp, the unwritten rules of adult life—navigating ambiguous conversations, reading the subtle emotional currents between colleagues, or knowing exactly when to speak and when to listen—can feel like a foreign language without a dictionary. These challenges do not disappear with age; in fact, they often intensify as social environments become less structured and more unpredictable. For decades, scientists have looked to social robots as a potential bridge across this gap. The idea is that a machine, with its predictable behavior and lack of judgment, could offer a safe space to practice these difficult interactions. However, almost all previous work in this field has focused on children, leaving the needs of autistic adults largely unaddressed. Furthermore, the robots that have been built were designed by researchers and clinicians, often without asking the very people who would use them what they actually need.

This gap in understanding is what a team of researchers at Ben-Gurion University of the Negev set out to fix. They gathered five autistic adults for a ninety-minute online session, not to test a robot, but to design one from scratch. The group included a researcher who is also autistic, ensuring that the conversation was guided by lived experience rather than just observation. Using a shared digital whiteboard, the participants moved past abstract ideas to concrete details, discussing what a robot should do, how it should behave, and, crucially, when it should not be present. The result was a clear, unified vision that challenges many current assumptions in the field. The participants did not want a robot companion to live with them or a social assistant to speak for them. Instead, they described a tool for private rehearsal: a configurable partner that helps them practice for real life, but is designed to eventually become unnecessary.

The most striking finding from the session was a strong consensus against the idea of a robot that simply replaces human interaction. The participants expressed a fear that relying on a machine could lead to reduced practice of vital social abilities rather than gaining them. They envisioned the robot as a training aid, similar to the extra wheels on a child's first bicycle. The goal is not to ride with the extra wheels forever, but to learn how to balance on two wheels so the support can be removed. The robot should start by offering heavy guidance and then gradually step back as the user becomes more confident, ultimately teaching the person how to live and interact without it. This approach flips the script on many current designs, which often focus on keeping the user engaged with the machine for as long as possible.

Another major insight concerned the personality of the robot. The participants were critical of the overly polite, agreeable tone found in many current artificial intelligence systems. They argued that a robot that always agrees with you or apologizes too quickly is a poor teacher because real people do not behave that way. In the real world, people disagree, they get impatient, and they have their own opinions. The participants wanted a robot that could hold a position, offer a different viewpoint, and even disagree with them. This "authenticity over comfort" approach means the robot should not be a sycophant. Instead, it should be a realistic training partner that can simulate the friction of real conversation, allowing the user to practice standing their ground or navigating conflict in a safe environment. To make this work, the robot's identity should be adjustable, allowing the user to practice with a virtual partner who looks and sounds like an older man, a formal official, or someone from a different cultural background, depending on the situation they are preparing for.

The conversation also revealed a deep need for feedback that is tailored to the individual's specific struggles. The participants identified a wide range of skills they wanted to improve, from the timing of their speech and the use of eye contact to reading implied emotions and understanding body language. They rejected the idea of a simple score or a grade, which they felt could not capture the complexity of human interaction. Instead, they preferred feedback that was personalized and delivered in a way that suited them, whether that was written text or a specific type of signal. Crucially, they noted that the robot could offer a kind of honesty that human friends cannot. Because the robot is a machine, it does not have to worry about hurting feelings or damaging a relationship, which allows it to give direct, objective feedback that a human might soften or avoid entirely.

A significant portion of the discussion focused on the internal experience of the participants, particularly a condition called alexithymia, which makes it difficult to identify and describe one's own emotions. Many of the participants noted that they often feel something is wrong but cannot put it into words. They suggested that a robot should not rely on them to say "I am feeling anxious" before it can help. Instead, the robot could infer their emotional state from other clues, such as the music they are listening to or their behavior, bypassing the need for difficult self-reporting. This design feature would allow the robot to support users who struggle to articulate their internal state, offering help before they even realize they need it.

The participants were also very clear about where the robot should and should not be used. They drew a firm line between the private space of preparation and the public space of performance. They stated that a robot should never be present during intimate moments, such as a date, or in situations where they are trying to interact naturally with others. The visibility of a robot whispering advice in their ear would carry a stigma, signaling a dependency that they do not want to display in public. The robot belongs in the "backstage," a private space where they can rehearse, make mistakes, and refine their skills. Once they are ready, they step onto the "front stage" of real life to perform independently. The success of the robot, therefore, should not be measured by how long the user interacts with it, but by how well they can handle social situations without it.

Finally, the group imagined the physical form of this tool. They wanted a robot that was portable, perhaps small enough to carry, and one that could be repaired or maintained by the user, addressing the economic barriers that many autistic adults face. They also wanted the robot's expressions to be adjustable, allowing them to control how human-like or how clear the robot's emotions appeared. This level of control ensures that the tool adapts to the user's needs rather than forcing the user to adapt to the machine.

The study concludes that for social robots to be truly useful for autistic adults, they must be designed as temporary scaffolds rather than permanent companions. They must be authentic enough to challenge the user, private enough to protect their dignity, and flexible enough to adapt to the complex, shifting demands of adult life. By listening directly to the people who will use these systems, the researchers have outlined a path forward that prioritizes independence and real-world competence over simple engagement. The vision is not of a robot that solves social problems for the user, but of a tool that helps the user solve them for themselves.