For a parent concerned about their child’s development, waiting months or longer for an autism evaluation can feel like an eternity. Early diagnosis opens the door to interventions that can improve long-term outcomes, but access to trained specialists remains limited, and many clinical assessments rely on subjective observations that can be expensive and difficult to scale.

Researchers at Penn Engineering and the Kennedy Krieger Institute are working to change that.

In a new study published in IEEE Transactions on Biomedical Engineering, René Vidal, Rachleff University Professor in Electrical and Systems Engineering in Penn Engineering and Director of the Innovation in Data Engineering and Science (IDEAS) Initiative, collaborated with lead author and Ph.D. candidate Kaleab Kinfu and Stewart H. Mostofsky, co-author and director of the Center for Neurodevelopmental and Imaging Research at Kennedy Krieger Institute, to develop an artificial intelligence system that measures a child’s ability to imitate simple body movements using nothing more than ordinary video. Their technology is referred to as Computerized Assessment of Motor Imitation, or CAMI, with CAMI-2DNet as their most recent iteration.

“The technology offers a faster, more objective way to assess an important behavioral marker associated with autism,” says Vidal, who also holds a joint appointment in Radiology in Penn Medicine. “We hope to see this tool eventually make high-quality screening available in more clinics, schools and homes.”

One of the earliest ways children learn is by imitating the people around them. They wave goodbye, clap their hands and mimic playful dance moves, gradually building the motor and social skills that help them interact with others.

Researchers have long known that children with autism often imitate these movements differently than their neurotypical peers, making motor imitation an important clue during developmental evaluations.

“Motor imitation is a critical building block for social development,” says Mostofsky. “This technology provides an objective way to measure imitation skills using tools that are far more accessible than specialized laboratory equipment. Our long-term goal is to better understand individual differences in autism and help inform interventions that address those differences.”

Traditionally, measuring motor imitation has required specialized motion-capture systems or painstaking manual analysis by trained experts, limiting where and how often these assessments can be performed.

CAMI-2DNet changes that equation.

Using video from a standard camera, the AI tool automatically tracks a child’s movements as they imitate an instructor performing simple motions. It then compares the movements of the instructor and the child and generates an objective score, eliminating hours of manual review while requiring none of the expensive equipment used in traditional motion analysis.

https://www.engineering.upenn.edu/wp-content/uploads/2026/07/CAMI-Autism-Video-1-compressed.mp4

CAMI-2DNet builds on a series of studies that have steadily moved computerized motor imitation assessment closer to clinical use.

Earlier work established that children’s ability to imitate simple movements could serve as an objective marker of autism. In a 2025 clinical study of 183 children, researchers showed that a one-minute Computerized Assessment of Motor Imitation (CAMI) could distinguish autistic children from their neurotypical peers with an 80% true positive rate and differentiate autism from ADHD with a 70% true positive rate. The latter is an especially important finding because the two conditions often co-occur and can be difficult to tell apart.

However, previous versions of CAMI depended on specialized motion-tracking systems or required extensive manual preprocessing and annotation of video data before analysis. CAMI-2DNet removes those barriers by using deep learning to analyze ordinary video directly, eliminating much of the manual work while achieving performance comparable to the team’s more resource-intensive 3D motion-capture system.

“Our goal with CAMI-2DNet was to make that science practical,” says Vidal. “We are developing an AI system that can perform the assessment automatically from standard video, bringing us much closer to a tool that could be used in everyday clinical settings.”

To develop the system, the researchers needed to teach a computer to recognize key body joint positions and use an AI model that could focus on how a child moves while ignoring irrelevant differences like camera angle, lighting or body size.

“A key challenge was making the system robust enough to work in real-world settings,” says Kinfu. “We wanted the AI tool to pay attention to the quality of the movement itself, not where the camera was placed or who was performing the action. That makes the technology much more practical outside of specialized research labs.”

This project required Penn Engineers’ expertise in both computer vision and machine learning to demonstrate how advances in AI can help translate sophisticated engineering into tools that address real healthcare challenges.

“The exciting part of this work is advancing upon state-of-the-art AI to solve a problem with real clinical impact,” says Vidal. “By replacing specialized hardware and manual analysis with intelligent computer vision, we can help make objective behavioral assessments more accessible to clinicians and families.”

When the researchers evaluated CAMI-2DNet, the system performed as well as sophisticated 3D motion-capture technology and matched, or even outperformed, expert human raters.

Because it relies only on conventional video, the technology has the potential to bring high-quality motor assessments to settings where motion-capture laboratories simply aren’t practical.

CAMI-2DNet is not designed to diagnose autism on its own or replace the expertise of clinicians. Instead, it provides an objective measurement that can complement existing evaluations, helping clinicians make more informed decisions and allowing researchers to track changes over time.

 “Our hope is that tools like CAMI-2DNet can reduce barriers to timely, high-quality assessment,” says Kinfu. “By making these measurements easier to obtain, we can support larger studies, improve clinical workflows and ultimately help more children receive the evaluations and services they need as early as possible.”

The Penn Engineering team is now working to further improve their technology’s approach and expand the evaluation across multiple cohorts, clinical sites and movement types while collaborators at Kennedy Krieger Institute and Washington University in St. Louis are using CAMI in their clinical research, helping to evaluate its performance across different populations and settings.

Their next steps include continued validation and broader dissemination of the technology to other research groups and clinical labs to allow them to adopt and integrate CAMI into their own workflows.

Learn more about the work being done in Rene Vidal’s lab here and watch the video on CAMI technology here.

This work was supported by the National Science Foundation (grant numbers 2124276 to S.H.M., principal investigator, and 2124277 to R.V., principal investigator), National Institutes of Health R21 (grant numbers MH127501, R01 MH106564 and P50 HD103538 to S.H.M., coinvestigator), Simons Foundation (grant number 724867 to S.H.M., principal investigator), and The Eagles Foundation (grant number WU-24-0499 to S.H.M., co-investigator).