Robot Beings

Abstract

Being a robot in a human habitat requires dealing with cluttered, unconstrained and dynamically changing environments. Most research on autonomous mobile robots assumes a static world. At best, dynamic aspects of the world are to be avoided. We report on a robot, Seymour, which is designed to interact with people while operating in a crowded office environment. Seymour cannot be dynamically told what to do. Rather, like children and dogs, he does what is in his nature (which is determined by programs residing onboard in EPROMs on power up). He pursues his own activities while responding to the presence and actions of nearby people. Seymour bristles with sensors. But rather than fuse the data from his nine cameras and his pyroelectric array into a world model, he will have many independent perceptual systems which are individually and intimately tied into behavior-generating networks of simple computational elements. Each perceptual subsystem extracts only those aspects of the world which are relevant to the particular task for which it is tuned. Fusion happens closrr 1.0 t.hr motor levrl I.lian t.lir srnsor lrvrl. Srymonr usrs t.hr modified subsnmption architecture which is a methodology for implementing complex agents as an incrementally evolved network of augmented finite state machines. Our approach in building Seymour and other robots has been inspired in many ways by biological systems and research. In particular, we have adopted an evolutionary method of building complex autonomous agents, where the components are simple distributed computational elements. This gives us strong advantages in dealing with the complexity of the environment. We are not particularly interested however, in simply reproducing the complexity of Nature’s solutions. In fact, we maintain that biological inspirations can be taken too far. In particular introspection to determine how perception or even reasoning works is bound to fail.