Affordances as a Framework for Robot Control

Abstract

The concept of affordances, with its emphasis on the interaction between the organism and the environment, is highly relevant for robotics. In this paper, we present a new formalization of affordances that provides a framework for robot control, learning and planning. We argue that affordances, as relations within the robotenvironment system, can be learned by the robot from its interactions with the environment. The interaction is represented as a nested triple of the form (effect, (entity, behavior)) indicating that the behavior applied in an environment perceived as the entity, would produce a perceivable effect. The robot accumulates such triples of raw sensory-motor data from different interactions and learns the affordance relations through the formation of equivalence classes. We present three studies that implement certain aspects of the formalism on a mobile robot moving in an environment that contains different types of objects. These studies respectively explore (1) the perceptual learning of affordances, (2) the development of goal-directed behaviors from a set of primitive ones through the learning of affordances, and (3) the use of learned affordance relations in planning.