A framework for integrating perception, action, and trial-and-error learning

Meliora is a system being developed at the University of Rochester that learns to assemble objects from random piles of blocks. It is a testbed for exploring issues in Integrated Architectures. It is based on a control architecture that is reactive, adaptive, has an active -but limited- sensory system, has a limited internal representation, and does not depend on a priori domain knowledge. In these working notes, we describe Meliora's control architecture, discuss its features and limitations, and describe how it can be integrated with more "intelligent" mechanisms that improve its performance. The notes are organized around the question set laid out by the Program Committee.