@TechReport{Brooks89,
author = "Rodney A. Brooks",
title = "A Robot that Walks: Emergent Behaviors from a
Carefully Evolved Network",
institution = "MIT",
year = "1989",
number = "AI MEMO 1091",
}
In this paper Brooks describes his subsumption architecture. This architecture has been used and modified by several other people in the design of intelligent agents. It is a layered approach to building agents which Brooks feels in analogous to the development of intelligence in most animals. The expertise that animals have has been developed over a period of time. Each new generation of an animal has evolved from the previous generations. In this same way Brooks augments behavioural layers upon previously implemented ones. The experiments are done on a robot situated in the real world to maintain an sense of reality and to show the architecture is capable of handling real world problems.
The system is designed to incrementally improve the agents performance of tasks and to increase the range of tasks it can perform. This is done through the adding of layers which describe a behavioural pattern for the agent. The new layers are able to inhibit signals from other layers and put in new messages onto those signal lines. This allows the new behaviours to modify the existing behaviour layers.
Brooks describes the various steps and layers that were involved in the creation of a walking robot that is able to avoid obstacles and following moving objects. Each layer is simple, such as the lifting of a leg. And each layer builds on the previous ones, like when a leg is lifted it should be moved forward and all other ones on the ground moved backwards.
He is able to create a robust walking behaviour that is distributed with very little centralised control. This is a very different approach than traditional Artificial Intelligence has followed. Yet the resulting behaviour of the robot show an improvement over the more complex reasoning based systems. The architecture also shows that it is possible to create higher level behaviours through the interaction of simpler low level behaviours.