An integrated architecture for learning and planning in robotic domains

Abstract

Our system, GINKO, is fully implemented and integrates aspects of Learning, Planning, Execution, Perception, and Robotics In short, from Robotics we take our domains and we borrow the notion of planning in configuration space. We use Machine Learning techniques to classify regions of configuration space according to their qualitative behaviors. This corresponds to learning the conditional effects of robot operators. Our planner generates an abstract plan of transitions between regions of qualitative behavior in the configuration space. A concrete plan is generated by augmenting the abstract plan with execution monitors that sense critical aspects of the situations to ensure that a particular plan is being followed and that the regions from which the plan was derived are accurately characterized.