Summary: Researchers applied artificial intelligence (AI) to track how infants shift from seemingly random movements to intentional actions. Using motion capture data from a classic baby-mobile experiment, AI models—especially a deep learning model called 2D-CapsNet—successfully classified infant movement patterns and revealed that changes in foot movement are the clearest sign infants are learning to affect their environment.
The study found that infants increased exploratory behavior after being disconnected from the mobile, suggesting a drive to reestablish interaction with the world. These results demonstrate how AI can reveal subtle dynamics of early motor development and learning that are otherwise difficult to detect.
Key Facts:
- AI classified infant movement states with up to 86% accuracy, with foot movements showing the strongest signals.
- Infants explored more after losing control of the mobile, indicating efforts to reconnect with the environment.
- Machine and deep learning methods provide fresh, objective ways to study early motor development and infant learning.
Source: FAU
Advances in computing and artificial intelligence combined with experimental approaches to infant learning now make it possible to study how purposeful action emerges from early motor behavior.
While much prior work has distinguished spontaneous infant movements—such as fidgety versus non-fidgety activity—there remains limited understanding of how infants intentionally engage with objects and people in their environment and what learning processes guide this transition.
To address this gap, researchers at Florida Atlantic University and collaborators used the baby-mobile experiment, a well-established paradigm in developmental psychology. In this setup a colorful mobile is gently tethered to an infant’s foot so that a kick causes the mobile to move. That causal link between action and visual feedback lets researchers study how infants discover and control a relationship with an external object.
The team recorded infant movements with a Vicon 3D motion capture system and converted joint trajectories into pose-based descriptors. They then trained multiple AI classifiers to determine which stage of the experiment a short movement clip came from—capturing transitions from spontaneous activity to goal-directed kicking and responses to the mobile’s motion.
Across the tested architectures—k-Nearest Neighbour (kNN), Linear Discriminant Analysis (LDA), fully connected networks, 1D convolutional models and capsule networks, 2D convolutional networks, and 2D-CapsNet—the AI systems reliably distinguished experimental states. Overall, deep learning models performed best on whole-body features, with the 2D-CapsNet delivering the highest classification performance.
A consistent and notable finding was that foot movements provided the clearest signal of change across experiment stages. The feet—being the direct point of contact with the mobile—showed the largest and most coherent alterations in movement patterns. The 2D-CapsNet model reached about 86% accuracy when classifying foot movement snippets and typically outperformed other methods by roughly 20% compared to hands, knees or whole-body features.
“The AI systems were not given any information about which limb was connected to the mobile, yet they identified feet as the most informative,” said Scott Kelso, Ph.D., co-author and Glenwood and Martha Creech Eminent Scholar in Science at FAU’s Center for Complex Systems and Brain Sciences. “This indicates that an infant’s connection with the environment manifests most strongly where the organism contacts the world—here, ‘feet first.’”
The study also observed that many infants increased exploratory movements after the mobile was disconnected. According to co-author Aliza Sloan, Ph.D., this suggests that losing control prompts infants to search for ways to reestablish interaction. Some infants, however, continued movement patterns learned earlier in the experiment even while disconnected, suggesting those infants had formed a stronger expectation that their actions would produce outcomes.
Researchers propose that when AI detects sustained high classification accuracy during disconnection, it may reflect learning that occurred during prior interaction. Different movement types and their temporal patterns could therefore indicate distinct forms of learning or expectation formation.
“Studying infants is uniquely challenging because they cannot verbally report intentions,” said Nancy Aaron Jones, Ph.D., co-author and director of the FAU WAVES Lab. “AI helps us quantify subtle changes in motion and stillness, offering a window into how infants think and learn before they can speak. These movement signatures also help explain the wide individual variation in early development.”
By examining how AI classification accuracy evolves for each infant, researchers gain a sensitive measure of when and how infants begin to engage intentionally with their surroundings. Kelso notes that integrating theory-driven experiments with AI classification can move the field beyond simple spontaneous-movement labels toward assessments that are context-sensitive and clinically relevant for identifying developmental risks.
Study co-authors include lead author Massoud Khodadadzadeh, Ph.D., and Damien Coyle, Ph.D. Funding sources included high performance computing resources, research scholarships and fellowships in the U.K., support from the FAU Foundation, and grants from the U.S. National Institutes of Health.
Funding: The research was supported by Tier 2 High Performance Computing resources provided by the Northern Ireland High-Performance Computing facility funded by the U.K. Engineering and Physical Sciences Research Council; the U.K. Research and Innovation Turing AI Fellowship (2021–2025) funded by the Engineering and Physical Research Council, Vice Chancellor’s Research Scholarship; the Institute for Research in Applicable Computing at the University of Bedfordshire; the FAU Foundation (Eminent Scholar in Science); and United States National Institutes of Health.
About this AI and neurodevelopment research news
Author: Gisele Galoustian
Source: FAU
Contact: Gisele Galoustian – FAU
Image: The image is credited to Neuroscience News
Original Research: Open access. “Artificial intelligence detects awareness of functional relation with the environment in 3 month old babies” by Scott Kelso et al., published in Scientific Reports
Abstract
Artificial intelligence detects awareness of functional relation with the environment in 3 month old babies
This experiment examined how purposeful action begins in early infancy by manipulating an infant’s functional connection to an object—a colorful mobile tethered to the foot. Researchers used Vicon motion capture to record multiple joints and transformed these 3D spatial trajectories into Histograms of Joint Displacements (HJDs) to create pose-based descriptors.
Using HJDs as input, a range of machine and deep learning models were trained to classify the experimental state from short movement snippets. Architectures included kNN, LDA, fully connected networks, 1D convolutional and capsule networks, 2D convolutional networks, and 2D-CapsNet. Sliding-window analyses explored temporal changes in movement topology related to the functional context.
Classical methods such as kNN and LDA performed well when using single-joint features, while deep learning approaches—particularly 2D-CapsNet—showed higher accuracy using full-body descriptors. For every architecture tested, foot-related measures displayed the most pronounced and coherent changes across experimental stages, indicating that interaction with the environment alters infant behavior most strongly at the site of organism–world contact.