ÄûÃʵ¼º½

Autotelic Curiosity - The Co-Evolution of Goals and Behaviors

Cédric Colas 
MIT, Inria

Abstract: Humans continually innovate, escaping dead ends, making serendipitous discoveries, and building on each other's ideas across generations. Current AI frameworks, which treat goals as given, struggle to account for this open-endedness. I argue that a key missing ingredient is autotelicity: the intrinsic drive to generate, select, and pursue one's own goals. Humans are autotelic. Our goals and behaviors co-evolve throughout our lifetimes and across cultures, creating diverse learning landscapes that enable transfer, exaptation, and cumulative innovation. In this talk, I present a research program for building autotelic AI agents, spanning intrinsic motivation for goal selection, language models for goal generation, and the cultural dimension, where open-ended innovation emerges from the collective co-evolution of goals and solutions.

Bio: Cédric Colas is a Marie Curie postdoctoral fellow working between MIT and Inria. He introduced the autotelic agent framework, studying how curiosity and intrinsic motivation can drive agents to generate, select, and pursue their own goals. By making goals endogenous, autotelic agents can organize their own learning trajectories and build open-ended repertoires of skills through the co-evolution of goals and behaviors. His recent work extends this to the collective level, investigating how language, social interaction, and cultural transmission shape goal dynamics and might sustain open-ended innovation. His research bridges reinforcement learning, cognitive science, and cultural evolution. Contact: cedric.colas@inria.fr Website:

This guest talk is hosted by the , Department of Computer Science.

An opportunity to book 1:1 meetings with Cédric will be provided as part of the talk. If you wish to book a 1:1 but won’t be able to attend the talk, please contact christian.guckelsberger@aalto.fi.

Cédric’s visit is supported by the Aalto Science Institute (AScI) Visiting Scientist Programme.