Curiosity made robots learn like children — stumbling and all

Tell a 2-year-old to fetch the blue cup, and they will — even if they have never heard the phrase "blue cup" before. Knowing what blue is and what a cup is turns out to be enough. Today's AI, by contrast, needs to read billions of sentences just to manage the same trick.

A Japanese research team used robots to probe that gap. During training, the robots went through a stretch where their performance dipped before climbing back up — mirroring the stage when a child who has been saying a word correctly suddenly starts saying it wrong.

According to Science Advances, Vol. 12, No. 30, professors Tani Jun and Toya Kenji and researcher Theodore Tinker of the Okinawa Institute of Science and Technology Graduate University published findings on exactly this phenomenon.

The research team built a virtual robot resembling a truck-mounted crane inside a simulated environment. The robot moves according to scores — what actions earn points determines how it learns.

In a standard setup, there is only one scoring criterion: points are awarded only when the robot completes an assigned task. Through countless repetitions, the robot gradually refines its behavior toward whatever earns a reward.

The team added two more scoring conditions: one for when the robot's predictions turned out to be wrong, and another for when it moved its body in a novel way each time.

Before acting, the robot projects what it expects to see in the next moment — predicting, for instance, what scene will unfold if it extends its arm. Sometimes, when it actually moves, the result looks nothing like what it anticipated. That gap signals an experience the robot has never had before.

The team awarded points precisely at those moments of surprise. The more unexpected the outcome, the greater the reward — the equivalent, in human terms, of being the kind of person who keeps poking at something out of sheer curiosity. In effect, the robots were given a curiosity score.

After 60,000 training iterations, robots without the curiosity score achieved a success rate of 41 percent. Robots that received the curiosity score reached 85.1 percent. The same amount of training time produced a more than twofold gap in performance.

Midway through training, the curiosity-scored robots displayed an unusual behavior: they set aside their assigned tasks and wandered around as if playing. Their success rate at that point looked like aimless floundering, but the experience accumulated during that phase drove a sharp improvement in the second half.

The research team also ran a follow-up experiment with a trap built in.

Where the original setup rewarded the robot for doing what a command said, the new rules required it to do the opposite. The robot had to unlearn the rules it had internalized on its own in order to succeed.

After 60,000 training iterations, 10 robots achieved a success rate of 83 percent on commands they had been taught, 74 percent on commands they had not been taught, and 51 percent on the trap commands. Performance varied widely across individual robots, ranging from a low of 5 percent to a high of 90 percent.

What mattered was not the final score but the shape of the success-rate graph. Nine of the 10 robots traced a U-curve.

Performance started strong, then declined as training progressed, before recovering in the later stages and settling at a level higher than where it began. All 10 robots ultimately trended upward.

That U-shaped curve is a familiar pattern in child development. Children who have been correctly using an irregular form will, at a certain stage, start applying a regular rule to it instead — overgeneralizing a pattern they have just learned to cases where it does not apply. The same phase produces other similar errors across the board.

The same sequence showed up inside the robots' internal representations.

Each time the robot received a command, it placed that command in a position relative to others — grouping commands it judged to be similar close together and pushing those it judged to be different further apart. The process resembles laying cards out on a desk.

By examining which cluster a given command ended up near, the researchers could infer how the robot had interpreted its meaning. The team tracked how the robot arranged commands at each stage of training.

At 15,000 iterations, no clear command clusters had formed — the robot had not yet grasped any organizing rule.

By 30,000 iterations, distinct clusters had emerged: "look at" commands grouped with other "look at" commands, and "approach" commands with other "approach" commands. The problem was the trap command. "Look at the magenta pillar" had slotted in among the other "look at" commands.

By 60,000 iterations, that command had migrated to the "approach" cluster. The robot had recognized that despite the words "look at" in the sentence, the instruction actually meant to move closer.

The robots had followed the same learning sequence children go through.

However, drawing a direct equivalence between the robots and children would be a stretch. The experiment took place entirely inside a computer, and the robots' world was a miniature one — just six actions, six colors and five objects. The research team itself acknowledged as a limitation that 60,000 trials were needed to reach stable performance.

Meanwhile, the team identified two-way communication — in which a robot could ask for help on its own or adjust the difficulty of a task — as the next challenge to pursue.

DOI: 10.1126/sciadv.aee7533

Theodore J. Tinker et al., "Curiosity-driven development of action and language in robots through self-exploration." Sci. Adv. 12, eaee7533 (2026).