Learning Symbols for Hierarchical Control from Interaction: Controllability Control
Gregor H. W. Gebhardt · 2011
One of the most challenging tasks in robotics is the perception and compact representation of situations in unstructured, natural environments. The severity of this task lies also in the high dimensionality of these highly complex environments. The representation should for instance classify situations in a way that is relevant for action selection and allows for hierarchical decision making. The downscaling of high dimensional states to lower dimensional, potentially symbolic representations is an important step to solve planning and controlling problems within these environments. This thesis uses the concept of controllability states as symbolic representations for the states of an environment. A controllability state specifies, which degrees of freedom are controllable in a given state. The core contribution of this thesis is to show how such symbolic representations of controllability can be extracted based on the data an agent collects during the interaction with a continuous environment. The developed method is demonstrated in a 2-dimensional environment. Based on the learned symbolic representations and dynamic models associated to each symbol I derive the Controllability Control method: a controller that is able to provide a sequence of control instructions, that finally lead to a desired controllability state and hence to the controllability of one or more specific degrees of freedom. The Controllability Control model has successfully been tested for the task of moving an object, that is initially not controllable to a desired position.