Spatial Learning and Representation in Animats

Abstract

Animat AI has generally emphasised learning of a dispositional, or task-specific nature over that of a representational or task-independent kind. However, many animals are capable of both forms of learning, and, in particular, exploit representational learning to construct spatial knowledge that allows efficient and flexible navigation behaviour. The focus on building versatile mobile robots may therefore force the development of representational learning systems in animat AI. This paper considers the navigation problem and argues against the view that qualitative spatial representations, encoding principally topological relations, may necessarily be simpler to construct, store, or use than more quantitative models. It further argues against constructing a unified or global representations of space suggesting instead that knowledge should be distributed between multiple, partial, local models encoding complimentary constraints which can be combined at run-time to address a specific navigation task.