Experiences Looking into Niches
David Miller · 1995
I’ve worked on over a dozen different robot systems whose software control architecture ranged from nonexistent [4] to ponderous [2]. For the past many years, I along with several others, have been developing a layered architecture that joins reactive control with more traditional symbolic reasoning [6]. I believe that this architecture is the right way to go to create physical agents that can carry out a large variety of complex tasks in realistic environments. However, despite numerous publications with titles like An Implemented Architecture..., many of them coauthored by myself, the actual experience with robots using these layered architectures where the full system has been implemented on a physical robot has been extremely limited. Therefore the conclusions that can be gained for these experiences are, at this time, limited. Several of the other attendees are talking about these systems, and will cover them in detail. The previews of their submissions that I have seen, I in large part agree with, and therefore it seems that no purpose will be served by my adding yet another set of opinions based on such a small set of results. Fortunately, I believe there is a large class of problems that can be solved using physical agents that does not demand the versatility and cognitive capabilities we all hope to demonstrate using the symbolic to reactive layered architecture. These problems are simpler and better specified than the vague problem specifications one would give to Rosie the Robot or Lt. Cmdr. Data. Examples of the type of simpler problem I’m talking about would be keeping house floors clean, inspecting the space station, or delivering parts on the factory floor. I will refer to this class of tasks as niche tasks, and the robots that can solve these tasks as niche robots. Once again, niche robots are not the ultimate solution to intelligent robots. By their very definition their scope of applicability is limited. However, I and many other researchers have had considerable experience designing, building and testing niche robots. Therefore this experience can be more confidently codified, and the workshop questions more accurately answered with regard to these types of robots. While this at first might appear appear to be that of the drunk looking for his dropped keys under the lamp post because the lighting is better, I believe the situation is a little better in this case. In many ways the layered robots people are proposing and starting build resemble a colony of niche robots. While these more complex robots still have a single mechanical instantiation, the whole philosophy behind the strategic, and in some cases sequencing layers is to switch between different sets of reactive behaviors. Each of these behavior sets represents to some extent a niche robot a set of robust behaviors for accomplishing a task over a narrow domain. By better understanding the capabilities and limits of niche robots, we may better learn how to build truly intelligent agents. But while related evolutionarily, niche robots and layered robots are different species. As with all cross-species studies, conclusions based on experiments with one species should be transferred to another species only with the greatest of caution.