Detaching the strings Practical algorithms for Learning from Demonstration

Learning from Demonstation (LfD) is a machine learning paradigm wherby agents are trained to execute tasks by observing demonstrations from other agents, most commonly humans. LfD methods are particularly appealing for tasks that cannot be easily defined either in the form of code or cost functions. Such tasks include robots acting in unstructured environments with implicit rules such as social interaction.
While powerful a significant amount of effort is required in order to deploy LfD algorithms in the real world. This thesis provides a number of algortihms that make LfD more applicable to such environments. It adresses problems encountered at all levels of the LfD pipeline for robotics, such as data collection and interpretation, path planning, local control, automatic task decomposition and meta-control.