Projekt

The main goal of the project is to build a theory of child-like curiosity. Children are arguably the only known system that demonstrably and reproducibly develops into intelligent agents through playful exploration. The project team aims to build machines that do the same. This involves studying how children explore their environment during free play, extracting the algorithms they apply, and using these models to build robots that effectively explore similar environments. Computationally mirroring this development requires a formal understanding of curiosity - the ability to explore environments in the absence of rewards. Studying curiosity demands an inter-disciplinary approach, where developmental psychologists, cognitive scientists, and roboticists work together to understand the human ability to be curious and build algorithms that mirror this ability. The proposed project will coalesce around three objectives. In the first objective, curiosity will be studied by letting children play freely while tracking their actions. In the second objective, the best model to describe children's behavior will be identified, by building cognitive algorithms of child-like curiosity. Finally, in the third objective, robots that can efficiently solve similar tasks will be build. The ultimate goal is to build more powerful robots that play like children, thereby moving towards a science of curiosity.

• Dr. Eric Schulz

  Max-Planck-Institut für biologische
  Kybernetik
  Computational Principles of Intelligence Lab
  Tübingen
  

• Prof. Dr. Georg Martius

  Universität Tübingen
  Distributed Intelligence
  Tübingen
  

• Prof. Dr. Dr. Azzurra Ruggeri, D. Phil

  Technische Universität München
  School of Social Sciences and Technology
  München
  

• Curious Exploration via Structured World Models Yields Zero-Shot Object Manipulation

• Modeling Human Exploration Through Resource-Rational Reinforcement Learning