Children as young as 3 years old can instinctively read intentions and preferences in a human's eyes, but they completely fail to recognize the same nonverbal communication when a humanoid robot stares at an object. This finding from an international study in developmental psychology is forcing engineers and AI researchers to fundamentally rethink how physical robots should be designed to interact with children.

The research, coordinated by Professor Antonella Marchetti at Università Cattolica in Milan and published in the International Journal of Child-Computer Interaction, tested Italian children aged 3 to 5 years old. Researchers showed each child either a person or a humanoid robot looking at a specific object, then asked whether the child could figure out which item the agent "preferred".

The results revealed a stark cognitive divide. When children watched a human look at an object, they naturally assumed the person liked it. But when a robot performed the identical action, the gaze conveyed no psychological meaning to the child. The robot's stare was treated as a meaningless mechanical movement, not as a window into preference or desire.

Why Does a Robot's Gaze Feel So Different to Children?

The answer lies in how children's brains search for evidence of a mind behind the eyes. A human gaze signals intentionality, thought, and feeling. A robot's gaze, even when perfectly mimicked, registers as an empty gesture. Children instinctively understand that a person looking at something reveals their inner preferences, but they don't extend that same logic to machines.

This distinction matters enormously for the field of embodied AI, which focuses on integrating artificial intelligence into physical systems like humanoid robots. The study demonstrates that simply copying isolated human behaviors, like eye movement, is insufficient to create genuine communication between a robot and a child.

"Simply imitating a single human signal, such as gaze, in a robotic artifact is not enough to make it truly communicative in a child's eyes. Designing robots and intelligent technologies for children requires richer, more natural, and developmentally appropriate interactions: made up of words, gestures, reciprocity, context, and shared presence," explained Professor Antonella Marchetti, Director of the Department of Psychology at Università Cattolica.

Professor Antonella Marchetti, Director of the Department of Psychology, Università Cattolica

What Does This Mean for Embodied AI Development?

The findings carry significant implications for how engineers approach physical AI. Many AI systems today focus on verbal output and text-based responses, but this research highlights that communication involves far more than words. For children especially, the presence and shared context of a physical, interactive system matter deeply.

Embodied AI, which integrates artificial intelligence into physical bodies that can move and interact in real environments, represents a crucial dimension for helping children attribute mental states like intentions and beliefs to technology. However, the study shows that embodied systems must go beyond surface-level mimicry to achieve genuine communicative connection.

How Can Designers Build Better Child-Robot Interactions?

• Multimodal Communication: Combine gaze with words, physical gestures, and reciprocal responses rather than relying on any single signal to convey meaning or intention.
• Developmental Appropriateness: Design interactions that match the cognitive and social abilities of the target age group, recognizing that children process robot behavior differently than adult behavior.
• Shared Context and Presence: Create situations where the robot and child engage in genuine back-and-forth interaction, building a sense of shared understanding rather than one-way observation.
• Intentionality Signals: Incorporate multiple cues that signal the robot has thoughts, preferences, and goals, not just mechanical responses to stimuli.

The research also has direct clinical applications. Children on the autism spectrum often struggle with gaze interpretation and shared attention, which are vulnerable dimensions of social development. Understanding how children perceive robot gaze can help design more effective therapeutic interventions.

In response to these findings, the Don Carlo Gnocchi Foundation and Università Cattolica are launching the ROBIN project (ROBot-based Neuropsychomotor INtervention) in June 2026. This initiative will use humanoid robots to promote imitation skills and socio-communicative rehabilitation in young children with autism spectrum disorder, applying the research insights to real-world therapeutic settings.

How Is the Physical AI Industry Scaling Training Data Collection?

While researchers work to improve how robots communicate with children, the broader physical AI industry is racing to gather the massive amounts of real-world data needed to train more capable embodied systems. Human Archive, a startup focused on collecting training data for robotics, recently raised $8.2 million in seed funding from Wing Venture Capital, NVP Capital, Y Combinator, and angel investors from frontier AI labs.

The company develops hardware and mobile systems that collect and synchronize multiple types of data, including video, audio, sensor readings, and long-duration activity recordings. These datasets are gathered across diverse real-world environments including homes, hotels, restaurants, agriculture, construction sites, industrial facilities, and retail locations.

Human Archive has already deployed more than 1,000 data-collection headsets, primarily in India, and is expanding into Southeast Asia and the United States. The company is adding tactile gloves, motion-capture systems, and other sensors to capture richer datasets that go beyond video alone.

"Despite decades of research, we still barely understand ourselves. Our goal is to learn how humans interact with the world, and over the past 6 months, our team's made enormous progress toward that alongside leading AI labs," noted Raj Patel, co-founder of Human Archive.

Raj Patel, Co-founder, Human Archive

The funding reflects growing recognition that understanding human physical behavior is essential for automating manual labor and developing more capable physical AI systems. Patel emphasized that the company's technology could become foundational infrastructure for advancing both robotic automation and our understanding of human intelligence itself.

These two developments, taken together, illustrate the dual challenge facing embodied AI: designing systems that can meaningfully interact with humans, especially children, while simultaneously gathering the vast amounts of real-world training data needed to make those systems intelligent and capable. The Marchetti study shows that the first challenge requires far more than mechanical mimicry, while the Human Archive funding demonstrates that solving the second challenge demands significant investment and global coordination.