FIAS Scientific Report 2010 - Frankfurt Institute for Advanced Studies ...

4. Project “IM-CLeVeR”: Intrinsically Motivated Cumulative Learning Versatile
Robots
Collaborators: C. Dimitrakakis 1 , L. Lonini 1 , C. A. Rothkopf 1 , J. Triesch 1 , R. Pramod 1
1 Frankfurt Institute for Advanced Studies
IM-CLeVeR is an integrated project (IP) funded by the European Union and involves partners from multiple
European countries. The overall goal of the project is to better understand:
– The development of intrinsic motivation mechanisms for driving exploration and autonomous development.
– Cumulative learning of multiple skills, through the efficient building of new skills by reusing or refining old
skills.
IM-CLeVeR brings together Neuroscientists, Psychologists, Roboticists, and machine learning experts to tackle
these problems. There are diverse research efforts around these core areas, including exploration and exploitation
trade-offs, optimal planning, efficient representations, novelty detection and hierarchical world models, to
name but a few.
At FIAS, the research focuses in the following main areas:
1. Applied vision research. Using a mobile robot head, efficient algorithms for perceptual abstraction, object
recognition, tracking and autonomous learning are investigated. This effort also lays the groundwork for
models developed in other parts of the research to be implemented in the robot itself.
2. Hierarchical models for reinforcement learning, estimation and control. Complex inference and planning
problems can be more easily managed in a hierarchical framework. A number of statistical techniques
are investigated for efficient and accurate learning and planning.
3. Links between utility theory and apprenticeship learning. Apprenticeship learning is a potentially important
learning mode. The objective is not only to imitate what a demonstrator is doing, but to furthermore
infer what he is doing and improve upon it.
4. Links between models and animal learning. Investigate whether models exhibit emergent behaviour
reminiscent of animal learning, then design experiments to test this hypothesis.
5. Near-optimal Bayesian planning. Since Bayesian planning is intractable, investigate efficient approximate
algorithms that do as well as possible given a computational budget.
6. Utility theory models for curiosity. Characterise the value of information in an abstract classes of tasks,
thus obtaining a well-defined measure of curiosity. Derive approximate, but nearly optimal, “curiositybased”
algorithms based on this measure.
Related publications in 2010:
1) A. Alessi, L. Zollo, L. Lonini, R. De Falco, E. Guglielmelli, Incremental learning control of the DLR-HIT-
Hand II during interaction tasks, Engineering in Medicine and Biology Society (EMBC), 2010 Annual
International Conference of the IEEE. pp.3194–3197.
17