A Colony Architecture for an Artificial Creature
MASSACHUSETTS INST OF TECH CAMBRIDGE ARTIFICIAL INTELLIGENCE LAB
Pagination or Media Count:
In this report we describe a distributed control system for a mobile robot which operates in an unmodified office environment occupied by moving people. The robots controller is composed of over 40 separate processes which run on a loosely connected network of 24 processors. Together this ensemble helps the robot locate empty soda cans, collect them with its arm, and bring them back home. A multi-agent system such as this has many advantages over classic monolithic controllers. For instance, it can be developed in stages, each new layer building on the last. It can also be split among several processors or agents, so as new capabilities are required new hardware can be easily added. Furthermore, its performance degrades gracefully - if a single agent fails the robot continues to function, albeit at a lower level of competence. However, in order to achieve these goals the system must be decomposed following certain guidelines. First, the internal workings of each agent should be isolated from all other agents. This improves the modularity of the system and helps prevent implementation dependencies. Second, all decisions should be based on spatially and temporally local information. This keeps the robot from relying on incorrect models of its dynamically changing world and allows it to operate with incomplete sensory input. Unfortunately, these restrictions make it nearly impossible to use conventional techniques to perform tasks requiring spatial reasoning. The can collection task is particularly difficult because it requires three different types of spatial knowledge. The robot must be able to navigate through its environment, recognize the shape of a can, and determine how to move its arm for grasping.