For many families, the path to an autism evaluation is not quick or simple. Parents may notice developmental differences early, yet still face long waits, limited specialist availability, high costs, and uncertainty about where to turn next.
A recent Penn Engineering article highlights one research effort aimed at reducing those barriers. Researchers from Penn Engineering and the Kennedy Krieger Institute are developing an artificial intelligence system that uses ordinary video to assess a child’s ability to imitate simple body movements. The technology is called Computerized Assessment of Motor Imitation, or CAMI, with CAMI-2DNet as the team’s most recent iteration.
The goal is not to replace clinicians or diagnose autism on its own. Instead, the research points toward a more accessible way to measure an important behavioral marker that may support developmental evaluation, clinical workflows, and larger research studies.
A behavioral marker that can be hard to measure consistently
Motor imitation is one of the ways children learn from the people around them. A child may wave goodbye, clap along, copy a gesture, or mimic a playful movement. These actions are not only motor skills. They are also part of how children engage socially.
Researchers have long observed that some children with autism imitate movements differently than neurotypical peers. That makes motor imitation a useful area of study during developmental evaluations, but measuring it can be difficult.
“Motor imitation is a critical building block for social development.”
— Stewart H. Mostofsky, Director of the Center for Neurodevelopmental and Imaging Research at Kennedy Krieger Institute
Traditional assessment of motor imitation may rely on expert observation, manual scoring, or specialized motion-capture systems. These approaches can be valuable, but they are not always easy to scale. They may require trained personnel, expensive equipment, research-lab conditions, or significant time for video review.
That creates a practical problem. If an assessment method is too resource-intensive, it may be less available in schools, community clinics, rural settings, or lower-resource environments where families already face barriers to care.
Ordinary video as the starting point
CAMI-2DNet is designed to make motor imitation assessment more practical. According to the Penn Engineering article, the system uses video from a standard camera. A child imitates an instructor performing simple movements, and the AI tool tracks the child’s movement, compares it with the instructor’s movement, and generates an objective score.
That shift matters because it removes some of the constraints of traditional motion analysis. The system does not require specialized 3D motion-capture equipment, and it reduces the need for hours of manual review.
The researchers describe this as a move from proof of concept toward a tool that could eventually be useful in everyday clinical settings. Earlier work had already shown that a one-minute Computerized Assessment of Motor Imitation could distinguish autistic children from neurotypical peers with an 80% true positive rate in a 2025 clinical study of 183 children. The same study found that CAMI could differentiate autism from ADHD with a 70% true positive rate, an important distinction because the two conditions can co-occur and may be difficult to separate clinically.
CAMI-2DNet builds on that work by making the process less dependent on specialized equipment and manual preprocessing.
Computer vision meets developmental assessment
The technical challenge is more complex than simply recording a child on video. The system needs to identify key body joint positions, follow movement over time, and compare a child’s motion with an instructor’s motion.
It also needs to avoid being distracted by factors that should not determine the score, such as camera angle, lighting, body size, or background differences. Lead author Kaleab Kinfu describes the challenge as making the system robust enough for real-world settings, so the AI attends to movement quality rather than irrelevant recording conditions.
That point is important for any screening-related technology. A tool meant to expand access cannot work only in ideal laboratory conditions. It needs to perform across varied settings, populations, cameras, and environments.
The Penn Engineering team reports that CAMI-2DNet performed as well as sophisticated 3D motion-capture technology and matched, or even outperformed, expert human raters. Those findings suggest potential, but continued validation across multiple cohorts and sites remains essential.
A tool to complement, not replace, clinical judgment
The Penn Engineering article is clear that CAMI-2DNet is not designed to diagnose autism by itself. That distinction should remain central in any discussion of AI and developmental screening.
Autism evaluation is complex. It involves developmental history, caregiver input, clinical observation, standardized tools, communication patterns, social behavior, sensory differences, co-occurring conditions, and professional judgment. A video-based motor imitation score can add useful information, but it cannot capture the full clinical picture.
The more realistic promise is that AI tools may help make one part of the evaluation process more objective, consistent, and accessible. They may support clinicians, help researchers track changes over time, and reduce the burden of manual scoring.
That is a meaningful role. It is also a limited one. The value of AI in this context depends on how well it integrates into responsible clinical workflows.
Access is the central issue
The strongest part of this research is its focus on access. Families often wait months or longer for autism evaluations. Those delays can be especially difficult when parents already sense that something is different in their child’s development.
Early diagnosis can open the door to interventions, services, and supports that may improve long-term outcomes. But access to trained specialists is uneven. Some families live far from specialty centers. Others face cost, scheduling, transportation, insurance, or language barriers.
A tool that uses ordinary video could help reduce some of those barriers if it is validated and implemented responsibly. It could support clinics that do not have specialized motion-capture labs. It could help schools or community settings collect useful developmental information. It could allow researchers to study larger and more diverse populations.
That does not mean AI alone solves access. Technology still needs trained interpretation, privacy protections, equitable deployment, and a clear pathway for follow-up. But a lower-barrier tool may help more children reach the next step sooner.
The equity question
Any technology meant to expand screening access must also be tested for equity. AI systems can perform differently across populations if the data used to develop them is not diverse enough. Camera quality, lighting, physical environment, clothing, mobility differences, and cultural variation in movement or interaction could all affect performance.
The Penn Engineering article notes that the team is working to expand evaluation across multiple cohorts, clinical sites, and movement types. That is exactly the kind of work needed before a tool like CAMI-2DNet can be widely adopted.
For clinicians and health systems, the question should not be only whether the tool performs well in a study. It should also be whether it performs consistently across the children most likely to experience delayed evaluation today.
Equitable screening technology must be accurate, accessible, explainable, and connected to care. Otherwise, it risks creating another layer of assessment without solving the underlying access problem.
What this means for schools and community settings
Schools are often where developmental differences first become visible outside the home. Teachers may notice differences in imitation, motor coordination, social participation, communication, or peer interaction. School-based teams may also be the first to encourage families to seek evaluation.
A video-based tool such as CAMI-2DNet could eventually be useful in settings where specialist access is limited, provided it is validated for those environments. It could help collect objective information that supports referrals, research, or monitoring.
Still, schools and community programs would need clear boundaries. A screening-support tool should not label a child without a full evaluation. Results would need to be communicated carefully, with respect for families and without creating unnecessary alarm.
The strongest use case may be as part of a larger pathway: observation, screening support, referral, clinical evaluation, and access to services when needed.
A broader shift in healthcare AI
CAMI-2DNet is part of a larger movement toward using AI to make specialized assessments more scalable. In this case, advanced computer vision is being applied to a real clinical bottleneck: the difficulty of measuring behavior objectively and efficiently.
That shift can be valuable when the technology is used carefully. AI can help process complex movement data quickly. It can reduce the burden of manual scoring. It can create standardized measurements that may be useful for research and longitudinal tracking.
But AI should not be treated as neutral simply because it is automated. Its outputs depend on the data, design choices, validation process, and clinical context behind it. The role of researchers and clinicians is to make sure those tools improve care rather than oversimplify it.
A practical step toward more accessible assessment
The promise of CAMI-2DNet is not that it turns an ordinary video into a full autism diagnosis. Its promise is more practical: it may help measure one meaningful developmental behavior in a faster, more objective, and more scalable way.
That is still important. For families waiting for answers, any responsible tool that helps improve access, reduce delays, or support better referrals deserves attention.
The next stage will determine how far the technology can go. Continued validation across sites, populations, and movement types will be essential. So will thoughtful integration into clinical and educational workflows.
For now, CAMI-2DNet offers a glimpse of how AI may support developmental care when it is focused on a real problem: making high-quality assessment easier to reach.
