Pierre-Yves Oudeyer: learning through curiosity

Before the task comes the question

A robot can be given a destination, a movement to imitate or a score to maximise. Pierre-Yves Oudeyer has spent much of his career investigating an earlier problem: how could it decide what is worth trying? The question shifts attention from the performance of a finished system to the experiences through which its abilities develop. It also makes room for an unusual meeting of disciplines, bringing computational models into conversation with developmental psychology and the study of language.

Oudeyer studied theoretical computer science at the École normale supérieure de Lyon between 1997 and 2000. He completed his doctorate at Université Paris VI in 2003, worked as a senior researcher at Sony Computer Science Laboratories in Paris, and joined Inria in 2008. His curriculum vitae dates his appointment as an Inria research director to 2012. He now heads the Flowers team, a role confirmed in an Inria Alumni announcement for November 2026.

From the emergence of speech to learning itself

His early work asked how organised communication might arise from interacting individuals. In The Self-Organization of Speech Sounds, published in 2005, populations of computational agents developed shared, structured sound systems through the coupling of perception and production. The model explored how collective regularities could emerge without an agent being handed a complete speech code.

This was a proposed mechanism, rather than evidence that human language had followed precisely that historical route. Its wider significance for his trajectory was the attention it gave to development: complex behaviour could be investigated as something that forms through interactions. An institutional profile published by ENS de Lyon traces his passage from research on the origins of language towards the mechanisms that help learners discover and acquire new skills.

Making curiosity a research question

With Frédéric Kaplan, Oudeyer helped clarify what researchers mean when they describe a machine as intrinsically motivated. Their 2007 survey distinguishes several computational approaches instead of treating curiosity as a single ingredient. A system might seek unfamiliar situations, prediction errors, or evidence that its abilities are improving. Those choices produce different behaviour.

The distinction between difficulty and progress is particularly useful. An experience can remain unpredictable without offering anything learnable; another can become less rewarding for exploration once it has been mastered. A mechanism sensitive to learning progress can redirect attention towards activities in which improvement is still possible. Here, curiosity names an operational rule for choosing experience. It does not establish that a machine feels the desire to know, and the paper places these ideas within a longer history of research rather than claiming a solitary invention.

Robots that choose what to practise

Work with Adrien Baranes made that principle concrete. Their 2013 study of intrinsically motivated goal exploration examined how a robot could select goals and learn the movements needed to reach them. Instead of searching only through combinations of motor commands, the system explored a space of possible outcomes and used changes in competence to guide practice. This gave learning a developmental structure: what the robot attempted could change as its capabilities changed.

A later JMLR article by Sébastien Forestier, Rémy Portelas, Yoan Mollard and Oudeyer extended the study of goal exploration and automatic curricula across simulated settings and a physical Poppy robot experiment. In the latter, a robot used joysticks to influence another robot and objects in its environment. The results concerned deliberately constructed tasks; their interpretation depends on the supplied perceptual machinery and goal spaces. The researchers had not removed human design from learning, but had moved some decisions about practice into the learning process.

A wider conversation about intelligence

This programme does not require dismissing language models. Oudeyer was among the authors of Grounding Large Language Models in Interactive Environments with Online Reinforcement Learning, presented at ICML in 2023. Led by Thomas Carta with Clément Romac, Thomas Wolf, Sylvain Lamprier and Olivier Sigaud, the study investigated adapting a language model through interaction in a text-based environment. It addressed the connection between words and consequences, while remaining distinct from an experiment with a physical robot.

The same interest in choosing informative experiments reaches into artificial life. In work on Flow-Lenia, Thomas Michel, Marko Cvjetko, Gautier Hamon, Oudeyer and Clément Moulin-Frier used curiosity-driven exploration to search for diverse dynamics in simulated continuous cellular automata. The expanded 2026 version follows an ALife 2025 contribution. Its moving, interacting patterns are computational phenomena; their resemblance to ecosystems is a means of investigation, not evidence that biological life has been created.

A 2026 preprint on hardware interference carries exploration into a different domain. Ludovic Matar, Clément Moulin-Frier, Christophe Marabotto, Eric Jenn and Oudeyer studied behavioural coverage in a simplified dual-core simulator. The distinction matters: finding varied simulated behaviours is an exploratory step, not the certification of a physical processor or a complete explanation of its failure mechanisms.

A career in seven moments

1. 1997–2000 · A theoretical foundationStudies computer science at ENS de Lyon.
2. 2003 · A doctorate on the emergence of speechCompletes his PhD at Université Paris VI.
3. 2005 · Shared sounds from interacting agentsPublishes his computational study of speech self-organisation.
4. 2007 · Defining intrinsic motivationWith Frédéric Kaplan, surveys computational approaches to curiosity.
5. 2008 · Joining InriaBegins the next institutional stage of his research career.
6. 2018 · Institutional recognitionReceives the Inria–Académie des sciences Young Researcher Prize.
7. 2022–2026 · From curricula to exploratory scienceCollaborative studies examine goal exploration, language grounding and the discovery of simulated dynamics.

Oudeyer’s contribution gives a distinctive shape to the question of intelligence. Learning is not only a matter of supplying more examples or improving an answer. It also depends on how a learner finds its next experience. His work makes that choice an object of scientific study, with experiments that expose both the possibilities of autonomous exploration and the structures on which it still relies.