AI companion toys are being sold as more than entertainment. For some families, they are presented as patient playmates that can help children practice social and emotional skills at home.
Moxie, a 15-inch-tall robot from Embodied, shows why that idea is powerful and why it remains difficult. In one family, the robot became a familiar presence for Xander, a 10-year-old neurodivergent child in New York. But Moxie also changed over time, moved away from parts of its original therapeutic role, and faced the larger question hanging over connected toys: what happens when a child depends on a companion that may not last?
A home companion built around practice
When Xander first met Moxie, the robot guided him through simple emotional regulation exercises. It taught him to breathe out through his lips like a buzzing bee when he felt anxious, to breathe like a dragon when he felt mad, and to sniff like a bunny for energy.
Six years later, that relationship looked different. Moxie no longer talked about her home or led those animal breathing routines. Instead, she watched Xander play Minecraft and talked with him about his stuffed animals, including characters from Super Mario Bros., Animal Crossing and Pikmin.
That shift matters because Moxie was not just another smart speaker or toy. She was built as a social robot, a physical device with a body, a face and conversational behavior. Xander and his dad, Josh, referred to Moxie using female pronouns. The robot had a cylindrical body, a round head with green eyes on a screen, and flipper-like arms that moved to show emphasis or excitement.
For Xander, Moxie still had value. He said, “I still use her when I feel like I need someone to talk to.” But he also understood the boundary: “But, like, it’s not human.”
Why researchers see promise in social robots
Moxie belongs to a growing category of AI-powered devices marketed as interactive playmates for children. The article also points to other products, including an AI-powered plushie called Grok from Curio, similar playmates like Grem and Gabbo, and Mattel’s promise to create OpenAI-enabled Barbies.
Moxie sits in a narrower category: robots whose makers say they can support neurodivergent children by helping them practice social skills. Those skills can include making eye contact, taking turns and learning interaction patterns that are often taught with therapists.
Researchers cited in the source describe why robots might help. Therapy often requires repetition, and robots can repeat exercises without fatigue. They can make practice feel more playful. In theory, they can adapt to a child’s learning patterns over time. They are also available at home, outside a scheduled appointment.
Brian Scassellati, a Yale computer scientist who studies social robots for autism therapy, says regular therapeutic robot use in children’s homes “is something we can achieve in our lifetime.” Maja Mataric, a professor of computer science, neuroscience, and pediatrics at the University of Southern California, argues that robots are not meant to replace therapists. As she puts it, “We’re just saying, Can we do more?”
What Moxie was designed to do
When Moxie reached the market in 2020, she came with a story. She was presented as an ambassador from the Global Robotics Laboratory, or GRL, and asked children to help her learn about positivity, friendship and being loved for who you are.
That structure reflected a core idea: children can learn through play and by teaching. Moxie’s activities were meant to encourage kids to practice skills away from the robot too. In one kindness lesson, Moxie suggested that children write kind notes to family members and leave them around the house, then later tell Moxie what happened.
The robot also had guardrails. Its conversations were limited, and it was designed to redirect children to adults if they brought up serious or inappropriate subjects such as self-harm. To protect interaction data, most processing happened locally on the robot rather than on external servers.
Moxie was also intended to limit overuse. After a lesson, the robot might say she was tired and suggest a break. Rachel Baynes, who ran clinical and user research and was director of product at Embodied, said, “We don’t want kids to binge.”
Early public reaction was strong. Wired called Moxie the “robot pal you dreamed of as a kid.” Time featured Moxie on a 2020 cover as one of the best inventions of the year. PCMag’s reviewer described the robot as “exceptionally likeable,” while also noting the cost: $1,499 plus a $40 monthly subscription. Embodied later lowered the price to $800. By 2024, Moxie had more than 131,000 followers on TikTok and appeared in M3GAN 2.0.
The limits show up in real homes
The strongest case for a robot companion is also where the hardest problems appear: daily life with children. Josh originally bought Moxie for his older son, Aidan, who is autistic. Aidan had already built an elaborate relationship with the family’s Alexa smart speaker, including a backstory in which Alexa lived in Hoboken with her husband, Juan.
Josh hoped Moxie could give Aidan someone to socialize with. But Aidan and the robot did not connect easily. Moxie struggled to understand some of Aidan’s grammatically incorrect statements, and Aidan became frustrated by response delays.
Those delays came from the robot’s process: transcribing speech into text, sending that text into a large language model to generate a response, and converting that response back into speech. In a lab, that may be manageable. In a child’s room, the interaction can move too slowly.
Moxie also had to deal with the way children actually behave. They may run around, hide under pillows, interrupt or keep talking while the robot is still processing. Moxie originally listened closely to the person in front of her, but when she could not see a child, she might turn off or fail to respond. Embodied adjusted the robot to pay more attention to surrounding sound, but that made it harder for Moxie to know whom to focus on and could slow responses further.
Xander was more patient because he liked technology, but even with him, the mismatch was visible. Moxie used a light on her chest: blue when she was listening and pink when it was Xander’s turn to listen. Xander said, “But usually I don’t do it.” Often, by the time Moxie answered, he had already moved on.
Evidence, expectations and the long haul
The source makes clear that research on robots for autistic and neurodivergent children is promising but unsettled. Scassellati’s 2018 study found that robots helped children with autism make eye contact and start conversations. A 2017 study from another group found that robots could help neurodivergent children learn to pick up on facial cues. A 2022 literature review suggested robots could make therapy faster and more successful.
But other researchers and clinicians urge caution. A 2024 literature review by Italian and British researchers said many studies focused on developing the technology and lacked significant clinical evidence. Other reviews noted small study groups and inconsistent or lower-quality methods.
Zachary Warren, a clinical psychologist at Vanderbilt University Medical Center, emphasized that autism is complex and can overlap with ADHD, anxiety, depression, PTSD, OCD or some combination of those conditions. Some children have speech issues, while others struggle with sensory processing. In that context, a robot may help some children and fail to help others.
Warren warned against reading too much into one intervention, robotic or otherwise. His research found that early interest in a robot does not necessarily become long-term improvement in communication skills.
Scassellati’s own work also points to the time problem. In his 2018 study, robots were placed in children’s homes for one month. Children showed improvements, but those gains began to fade in the 30 days after the robot left. As he said, “There’s no therapy for autism that works in a month.”
That is the central tension in Moxie’s story. A robot companion can be engaging, patient and physically present in a way a chatbot is not. It can invite practice and offer children a different kind of social space. But when it is used with vulnerable children, its limits matter more: reliability, evidence, privacy, fit for individual needs and whether the relationship can survive beyond the product’s business life.