The Curious Robot

Curiosity is the strangest mental state. The mind usually craves certainty; being right feels nice; mystery is frustrating. But curiosity pushes back against these lesser wants, compelling us to seek out the unknown and unclear. To be curious is to feel the pleasure of learning, even when what we learn is that we’re wrong.

One of my favorite theories of creativity is the so-called “information gap” model, first developed by George Loewenstein of Carnegie-Mellon in the early 90s. According to Loewenstein, curiosity is what happens when we experience a gap “between what we know and what we want to know…It is the feeling of deprivation that results from an awareness of the gap.” As such, curiosity is a mostly aversive state, an intellectual itch begging to be scratched. It occurs when we know just enough to know how little we understand.

The abstract nature of curiosity – it’s a motivational state unlinked to any specific stimulus or reward – has made it difficult to study, especially in the lab. There is no test to measure curiosity, nor is there a way to assess its benefits in the real world. Curiosity seems important – “Curiosity is, in great and generous minds, the first passion and the last,” wrote Samuel Johnson – but at times this importance verges on the intangible.

Enter a new paper by the scientists Pierre-Yves Oudeyer and Linda Smith that explores curiosity in robots and its implications for human nature. The paper is based on a series of experiments led by Ouedeyer, Frederic Kaplan and colleagues in which a pair of adorable quadruped machines – they look like a dogs from the 22nd century – were set loose on an infant play mat. One of these robots is the “learner,” while the other is the “teacher.” Here’s a picture of the setup:

The learner robot begins with a set of “primitives,” or simple pre-programmed instincts. It can, for instance, turn its head, kick its legs and make sounds of various pitches. These primitives begin as just that: crude scripts of being, patterns of actions that are not very impressive. The robot looks about as useful as a newborn.

But these primitives have a magic trick: they are bootstrapped to a curious creature, as the robot has been programmed to seek out those experiences that are the most educational. Consider a simple leg movement. The robot begins by predicting what will happen after the movement. Will the toy move to the left? Will the teacher respond with a sound? Then, after the leg kick, the robot measures the gap between its predictions and reality. This feedback leads to a new set of predictions, which leads to another leg kick and another measurement of the gap. A shrinking gap is evidence of its learning.

Here’s where curiosity proves essential. As the scientists note, the robot is tuned to explore “activities where the estimated reward from learning progress is high,” where the gap between what it predicts and what actually happens decreases most quickly. Let’s say, for instance, that the robot has four possible activities to pursue, represented in the chart below:

A robot driven by curiosity will avoid activity 4 - too easy, no improvement - and also activity 1, which is too hard. Instead, it will first focus on activity 3, as investing in that experience leads to a sharp drop in prediction errors. Once that curve starts to flatten - the robot has begun learning at a slower rate - it will shift to activity 2, as that activity now generates the biggest educational reward.

This simple model of curiosity – it leads us to the biggest knowledge gaps that can be closed in the least amount of time - generates consistent patterns of development, at least among these robots. In Oudeyer's experiments, the curious machines typically followed the same sequence of pursuits. The first phase involved “unorganized body babbling,” which led to the exploration of each “motor primitive.” These primitives were then applied to the external environment, often with poor results: the robot might vocalize towards the elephant toy (which can’t talk back), or try to hit the teacher. The fourth phase featured more effective interactions, such as talking to the teacher robot (rather than hitting it), or grasping the elephant. “None of these specific objectives were pre-programmed,” write the scientists. “Instead, they self-organized through the dynamic interaction between curiosity-driven exploration, statistical inference, the properties of the body, and the properties of the environment.”

It’s an impressive achievement for a mindless machine. It’s also a clear demonstration of the power of curiosity, at least when unleashed in the right situation. As Oudeyer and Smith note, many species are locked in a brutal struggle to survive; they have to prioritize risk avoidance over unbridled interest. (Curiosity killed the cat and all that.) Humans, however, are “highly protected for a long period” in childhood, a condition of safety that allows us, at least in theory, to engage in reckless exploration of the world. Because we have such a secure beginning, our minds are free to enjoy learning with little consideration of its downside. Curiosity is the faith that education is all upside.*

The implication, of course, is that curiosity is a defining feature of human development, allowing us to develop “domain-specific” talents – speech, tool use, literacy, chess, etc. – that require huge investments of time and attention.* When it comes to complex skills, failure is often a prerequisite for success; we only learn how to get it right by getting it wrong again and again. Curiosity is what draws us to these useful errors. It’s the mental quirk that lets us enjoy the steepest learning curves, those moments when we become all too aware of the endless gaps in our knowledge. The point of curiosity is not to make those gaps disappear – it’s to help us realize they never will.

*A new paper by Christopher Hsee and Bowen Ruan in Psychological Science demonstrates that even curiosity can have negative consequences. Across a series of studies, they show that our "inherent desire" to resolve uncertainty can lead people to endure aversive stimuli, such as electric shocks, even when the curiosity comes with no apparent benefit. They refer to this as the Pandora Effect. I'd argue, however, that the occasional perversities of curiosity are far outweighed by the curse of being incurious, as that can lead to confirmation bias, overconfidence, filter bubbles and all sorts of errors with massive consequences, both at the individual and societal level.

Oudeyer, Pierre-Yves, and L. Smith. "How evolution may work through curiosity-driven developmental process." Topics in Cognitive Science.