Manifold-Based Sensorimotor Representations for Bootstrapping of Mobile Agents

Abstract

Subject of this thesis is the development of a domain-independent algorithm that allows an autonomous system to process sequences of the sensorimotor interaction with its environment and to assign a geometric interpretation to its motor capabilities. We utilize Lie groups, smooth manifolds endowed with a group structure, that allow for an elegant representation of geometric operations as a central foundation for such a sensorimotor representation. Expressing motor controls with respect to the manifold structure allows us to transform the sensorimotor interaction sequence into a specific set of data points. Finding a manifold and a transformation that minimizes an intrinsic conflict function corresponds to finding a topological structure that is the best fit for expressing the sensorimotor space the entity resides in. Experiments in a virtual environment are conducted that show the applicability of the approach with respect to different sensor and motor configurations.