Cooperative Agent Architecture

摘要

Our goal is to develop an understanding of how an intelligent robotic agentshould represent and execute plans for everyday tasks that involve cooperative interaction with human beings. The work is being done in the larger context of the Animate Agent Project at the University of Chicago. The primary goal of this project is to implement intelligent robotic agents using a software architecture, vision processing system, and task control language, that support writing clear, concise plans for carrying out tasks in ordinary, unaltered environments. The system uses two primary sources of knowledge to describe ways to achieve goals: a library of modular skills and a library of higher level reactive plans called RAPs. Skills are meant to capture those aspects of sensing and action tasks that are best described in terms of continuous feedback loops. RAPs are meant to capture those aspects of task execution that are best described as symbolic plans. We believe that the key to building intelligent agents is understanding how to structure and represent behavior in a uniform manner so that the skills and plans used to carry out one goal can also be used to achieve other, similar goals. Such modularity will allow rapid reprogramming of an agent to carry out new tasks, addition of new capabilities that can be combined with the old, and plans that can be learned from a teacher or through experience. A major accomplishment this year was developing and refining the skills and RAPs for a trash collection task. This task was motivated by the AAAI Robot Competition and we have refined the plans involved to the point where the robot has thrown away hundreds of pieces of trash. P.4 (KAR).