OIST researchers rewarded virtual robots for curiosity and watched play-like behavior, and even toddler-style grammar mistakes, emerge on their own.
Toddlers are not efficient learners in any conventional sense. They knock things over, repeat actions for no apparent reason and wander off task. Yet within a few years they master the staggering complexity of language. A new study from the Okinawa Institute of Science and Technology (OIST) suggests the wandering may be part of the point.
Writing in Science Advances, researchers in OIST’s Cognitive Neurorobotics Research Unit report that virtual robots rewarded for curiosity learned to follow language instructions far faster than robots without that drive, and along the way spontaneously developed behaviors strikingly reminiscent of children.
“The curious robots achieve what appears to be a genuine understanding of language in half the time that their indifferent cousins do,” said first author Theodore Tinker, a PhD student in the unit.
A brain-inspired network
Unlike the large language models behind today’s chatbots, which learn statistical patterns from vast troves of text, the OIST robots run on a predictive coding-inspired variational recurrent neural network, or PV-RNN. The architecture draws on influential theories of how the brain processes information, in which the mind constantly predicts its surroundings and updates its internal beliefs when those predictions fail.
The network works by minimizing what theorists call free energy, balancing two pressures: accuracy, or how well its predictions match reality, and complexity, or how much its internal beliefs must change to accommodate new experience. Tinker explained the trade-off with chocolate.
“Imagine you encounter a small, dark square. You eat it, and now you know what dark chocolate tastes like,” he said. “You then encounter another small, black square. Based on physical appearance and your internal beliefs, you predict it will taste like dark chocolate. If you’re correct, your prediction is accurate, and your internal beliefs stay intact, minimizing the complexity term. But if it was actually licorice, your prediction is inaccurate, and your internal beliefs have to change to account for your new experience — high complexity.”
Curiosity, in this framing, is a willingness to seek out the second kind of experience. “Maybe you really like dark chocolate but have never tried white. So, one day you try, even though it runs counter to your desire for accuracy and low complexity,” Tinker said. “You might still think dark is tastier, i.e., it provides a higher external reward than white, but you also get an internal rush for satisfying your curiosity. And, as a bonus, you’ve updated your understanding of what chocolate is like in general.”
Learning by playing
The researchers placed their robots in a physics-simulated environment and gave them instructions built from combinations of verbs, adjectives and objects, such as “be near green pillar” or “push left magenta dumbbell.” Training used reinforcement learning with two kinds of reward: an external one for completing the task and an internal one for curiosity, tied to the complexity term.
The curious robots not only learned faster but came up with novel solutions. Midway through training, they began knocking objects over, even as their task performance kept improving. Nothing in the tasks required it.
“And we were amazed by the play-like behavior and exception-handling performance that emerged independently during training,” Tinker said. “That’s my favorite part, because nothing about any of the tasks requires knocking over objects. We didn’t tell the robot to play around like a child. We told it to learn. And yet, apparently, one activity leads to the other — play-like behavior, reinforced by curiosity, helped the robots understand language faster.”
The poverty of the stimulus
The work speaks to one of linguistics’ most enduring puzzles. Noam Chomsky famously argued that children hear too little, and too messy, a sample of language to learn it purely from experience, a problem known as the poverty of the stimulus. How children nonetheless generalize so well remains contested.
The OIST team’s earlier work, published in 2025, found that a robot’s success on unseen tasks depended on the diversity of the sentence constructions it was trained on. The new study pushed that further. Curious robots trained on 48 combinations generalized to new ones at a rate of 25 percent. Expanding the training set to 180 combinations raised that to 85 percent.
Mistakes, just like a toddler’s
The most surprising result involved errors. Children learning English often go through a phase of saying “goed” instead of “went” or “runned” instead of “ran.” They first use the irregular forms correctly, then overgeneralize the regular “-ed” rule, then recover. Psychologists call this a U-shaped learning curve.
To see whether the robots would do something similar, the researchers swapped the meanings of two tasks, creating an exception to an otherwise consistent pattern. Performance on the exception first held, then dipped as the robots learned more tasks and overgeneralized, then rebounded once they absorbed the exception.
“The fact that the robots mirrored children’s exceptional-handling performance came as a complete surprise to us,” Tinker said, “because there was nothing in the model to specifically address overgeneralization or exception-handling.” The robots, he added, “seemed to pick this up on their own.”
A transparent window into learning
The researchers argue that networks like theirs make better scientific instruments than large language models, which contain trillions of parameters trained on data far removed from a child’s experience. A PV-RNN’s hidden state, which encodes its predictions based on past experience, is accessible, so the team can watch the robots’ plans unfold in real time and control what triggers their curiosity and how strongly.
Senior author Jun Tani, who heads the unit, was careful about the limits. “The robots are not thinking like children, and their situation is very abstract compared to the extremely dense language environments of children,” he said. “But this abstract environment allows us to study the impact of curiosity within a very convincing model of how humans learn to process language.”
“We’re very excited to see what other behaviors we can observe and study with this platform,” Tani said.
Notes
- Tinker, T., et al. “Curiosity-Driven Development of Action and Language in Robots Through Self-Exploration.” Science Advances, 23 July 2026. DOI: 10.1126/sciadv.aee7533.
- Okinawa Institute of Science and Technology (OIST). “Curious robots mimic how children can learn to understand language.” EurekAlert!, 23 July 2026. https://www.eurekalert.org/news-releases/1136155.