The Kick That Gave Learning Away

I once taped cheap pressure sensors under an old sneaker and spent an afternoon stomping around my apartment. On the screen, walking stopped looking like “walking.” It became a messy stream of pressure shifts, hesitations, and tiny corrections.

That is roughly the problem developmental scientists face with babies. An infant kicks, pauses, squirms, and kicks again. Somewhere in that commotion, exploration may be turning into learning. But where?

Machine learning can help researchers spot the transition—not by reading a baby’s mind, but by finding organization in movement that human eyes may miss.

Let’s build the experiment

Here is the setup from a wonderfully hands-on study of 3-month-old infants.

A ribbon connected an infant’s leg to a colorful mobile. When the baby kicked, the mobile moved. Researchers also recorded periods when kicking and mobile movement were not linked. Motion-capture equipment tracked how the infant’s joints moved through space (Khodadadzadeh et al., 2024).

Now imagine watching the baby.

At first: kick, wiggle, pause, accidental mobile dance.

Then, perhaps: a more organized burst of movement that reliably makes the interesting thing happen.

The researchers translated joint motion into compact descriptions of how the body’s pose changed. They then trained machine-learning systems to distinguish movement recorded when the infant was connected to the mobile from movement recorded when the infant was not.

The models could tell those conditions apart. In the researchers’ interpretation, that shift reflected movement becoming organized around the action–outcome relationship: my kick makes that move (Khodadadzadeh et al., 2024).

This is AI acting less like a robot teacher and more like a lab instrument. A conventional microscope reveals structures too small to see. This computational microscope reveals patterns too distributed across time and joints to notice easily.

Fun historical twist: the paper connects this approach to Alan Turing’s interest in the “child machine.” Instead of asking only whether machines can imitate adult intelligence, turn the question around: can a machine help us see intelligence assembling itself?

From wiggling to brain waves

The same basic strategy works with neural signals, although the sensor changes.

In another study, 5-month-old infants experienced live face-to-face social interaction while researchers recorded brain activity using magnetoencephalography, or MEG. The interaction included the tightly coordinated ingredients babies usually encounter together: infant-directed speech, eye contact, and contingent responses.

Activity in attention and sensorimotor regions during this social experience was associated with later language growth into toddlerhood (Bosseler et al., 2024). The sensorimotor result is especially interesting to me. Language learning did not show up as a disembodied word counter. It involved systems that help a baby attend, move, predict, and participate.

Put the studies side by side and a useful pipeline appears:

1. Capture a signal. Joint trajectories, gaze, speech pauses, or brain activity.
2. Ask a narrow question. Can the pattern distinguish linked action from unlinked action? Does activity during social exchange relate to later development?
3. Test beyond the training data. A useful model must work on observations it did not simply memorize.
4. Return to the child. Researchers still need experiments and developmental theory to interpret what the pattern means.

That last step is the important one. A classifier can detect a difference without explaining it. It may use a meaningful coordination pattern—or an unnoticed feature of the recording setup. Prediction is a clue, not a tiny digital neuroscientist delivering a verdict.

What this does—and does not—say about your child

For parents, the exciting takeaway is not that an app should score every kick or brain wave. Please, no baby dashboard with a red “learning” light.

The takeaway is that development can be visible in the shape of change. A movement that looks repetitive may be an experiment. A pause may be a prediction. A delighted repeat may be a child testing whether the world still follows the rule.

You can make room for that everyday science:

• Let safe actions have noticeable consequences. A spoon makes a sound; a block topples; a scarf changes the view.
• Pause before demonstrating the solution. Exploration needs a little runway.
• Respond to bids. A look, sound, or gesture can become part of a back-and-forth learning loop.
• Treat variability as information, not failure. Learning often looks untidy while a child searches for a controllable pattern.

None of this turns a household into a laboratory, and ordinary differences in movement or language are not diagnoses. If you have concerns about your child’s development, a pediatrician or qualified developmental professional can evaluate the whole child rather than a single signal.

AI gives developmental science a sharper way to inspect motion and brain activity. But the model does not discover meaning alone. Researchers create the task, choose the measurements, inspect the errors, and decide whether the pattern survives a better experiment.

The machine may catch the kick that gives learning away. Understanding why it matters still takes humans watching carefully.