I work with the EU-Project "Extending Sensorimotor Contingencies to Cognition" (www.eSMCs.eu). The project pursues the idea of active perception: by actively interacting with the world and thereby generating sensory information, humans and animals can learn the regularities (sensorimotor contingencies) in the information flows which are structured by their embodiment: the morphology and physical properties of the body as well as the kind and placement of their sensors. Adopting this concept in robotics research is a promising path towards robots that are capable to deal with highly uncertain environments and therefore are more efficient in solving real-life tasks.
I use this concept as a premise for my research which includes the following topics:
- developmental robotics: Robots that learn in an incremental fashion. By starting with random motor actions the robot is exploring its capabilities, generating its own sensory input from which it can bootstrap the sensorimotor contingencies.
- information theory: by quantifying the amount of information that is contained in sensory channels, that one channel contains about another channel, or that flows from one channel to another, the robot can identify its sensorimotor network without any prior knowledge.
- reinforcement learning: while exploring the world and the own body, a robot tries to optimize a value system in order to achieve desired tasks (specified by the experimenter).
- learning machines such as reservoir computing, continuous-time neural networks or gaussian prosesses: to learn the consequences of its own actions, the robot can use biologically inspired machine learning methods, e.g. artificial neural networks that extract relevant features or that learn forward and nverse internal models of the robot.
Other research interests includes EEG-based brain-computer interfaces as well as models of human perception on behavioral and neuronal levels.
