Researchers created robots with curiosity similar to that of toddlers, resulting in them acquiring language at twice the speed.

      Grant a robot curiosity, and it begins to learn language (and misbehave) much like a toddler.

      Researchers at the Okinawa Institute of Science and Technology (OIST) have been exploring how children rapidly acquire language for decades, and a recent study may have uncovered a key factor: curiosity.

      The team created a virtual robot equipped with a brain-like neural network and let it explore a simulated 3D environment filled with shapes, colors, and simple commands such as “push left magenta dumbbell.”

      Some robots were only rewarded for accurately completing tasks, while others received additional rewards for their curiosity, essentially giving them a small internal reward whenever they encountered something that challenged their existing understanding.

      What role does curiosity play in a robot's language learning?

      The curious robots significantly outperformed their uninterested counterparts, achieving a meaningful understanding of language in about half the time, according to the study published in Science Advances.

      The study's author, Theodore Tinker, likened this to trying white chocolate for the first time even though one enjoys dark chocolate; taking that risk expands your overall knowledge of chocolate.

      The experiment became even more intriguing midway through the training process. The curious robots began to knock things over and experiment with unintended actions, essentially engaging in play. This behavior was not programmed; it emerged spontaneously.

      Do robots truly make the same errors as toddlers?

      The robots also exhibited a familiar pattern seen in how children learn language. Children often initially use certain verb forms correctly, but then apply grammatical rules too broadly, making mistakes with verbs they had previously mastered, before ultimately identifying exceptions and correcting themselves. The robots displayed a similar U-shaped decline in performance.

      This stands in contrast to how current chatbots learn. Large language models, like ChatGPT, are trained on extensive datasets and generate the statistically probable next word. In contrast, this robot’s neural network functions more akin to ours, focusing on accuracy while striving to maintain its beliefs, only altering them when confronted with enough surprise to warrant adjustment.

      While this does not imply that robots comprehend language as humans do, it suggests that curiosity combined with diverse experiences may significantly contribute to how toddlers decode language with minimal foundational knowledge.