Fuzzy behavior coordination for robot learning from demonstration

This paper proposes a fuzzy framework for supervised training of hierarchical, reactive robotic behaviors by demonstration. The novel approach constitutes a method for the identification of behavioral response and activation rules in the context of partially observable behaviors, perceptual aliasing and motor action ambiguity that are typical for teleoperated robotic control. The behavior representation is based on the fusion of preferences for actions rather than fusion of actions. It is this property that makes learning robotic behaviors and their coordination from demonstration feasible, because behavioral preferences for actions can be matched with actually demonstrated actions. A genetic fuzzy system is employed for rule identification and adaptation. An evolutionary algorithm identifies those rules that match the observed state action examples extracted from demonstration. Behavioral rules are evaluated according to their consistency and specialization with respect to the training set in the context of other concurrently active behaviors.