Previous: An Architecture for Autonomous Agents Up: A Computational Model of Color Perception and Color Naming Next: The GLAIR Architecture

Introduction and Overview

In the elephant paper [Brooks 1990] appearing in the proceedings of the predecessor of the current workshop, Brooks criticizes the ungroundedness of traditional symbolic AI systems, and proposes physically grounded systems as an alternative, particularly the subsumption architecture. Subsumption has been highly successful in generating a variety of interesting and seemingly intelligent behaviors in a variety of mobile robots. As such it has established itself as an influential approach to generating complex physical behavior in autonomous agents. In the current paper we explore the possibilities for integrating the old with the new, in an autonomous agent architecture that ranges from physical behavior generation inspired by subsumption to classical knowledge representation and reasoning, and a new proposed level in between the two. Although we are still struggling with many of the issues involved, we believe we can contribute to a solution for some of the problems for both classical systems and physically grounded systems mentioned in [Brooks 1990], in particular:

We agree with the requirement of physically implemented systems as the true test for any autonomous agent architecture, and to this end we are working on several different implementations. We will present both our general multi-level architecture for intelligent autonomous agents with integrated sensory and motor capabilities, GLAIR, and a physical implementation and two simulation studies of GLAIR-agents.

By an architecture we mean an organization of components of a system, what is integral to the system, and how the various components interact. Which components go into an architecture for an autonomous agent has traditionally depended to a large extent on whether we are building a physical system, understanding/modeling behaviors of an anthropomorphic agent, or integrating a select number of behaviors. The organization of an architecture may also be influenced by whether or not one adopts the modularity assumption of Fodor [Fodor 1983], or a connectionist point of view, e.g. [McClelland et al. 1986], or an anti-modularity assumption as in Brooks's subsumption architecture [Brooks 1985]. The modularity assumption supports (among other things) a division of the mind into a central system, i.e., cognitive processes such as learning, planning, and reasoning, and a peripheral system, i.e., sensory and motor processing [Chapman 1990]. Our architecture is characterized by a three-level organization into a Knowledge level (KL), a Perceptuo-Motor level (PML), and a Sensory-Actuator level (SAL). This organization is neither modular, anti-modular, hierarchical, anti-hierarchical, nor connectionist in the conventional sense. It integrates a traditional symbol system with a physically grounded system, i.e., a behavior-based architecture. The most important difference with a behavior-based architecture like Brooks's subsumption is the presence of three distinct levels with different representations and implementation mechanisms for each, particularly the presence of an explicit Knowledge level. Representation, reasoning (including planning), perception, and generation of behavior are distributed through all three levels. Our architecture is best described using a resolution pyramid metaphor as used in computer vision work [Ballard \& Brown 1982], rather than a central vs. peripheral metaphor.

Architectures for building physical systems, e.g., robotic architectures [Albus et al. 1981], tend to address the relationship between a physical entity, (e.g., a robot), sensors, effectors, and tasks to be accomplished. Since these physical systems are performance centered, they often lack general knowledge representation and reasoning techniques. These architectures tend to be primarily concerned with the body, that is, how to get the physical system to exhibit intelligent behavior through its physical activity. We say these systems are not concerned with consciousness. These architectures address what John Pollock calls Quick and Inflexible (Q&I) processes [Pollock 1989]. We define consciousness for a robotic agent operationally as being aware of one's environment, as evidenced by (1) having some internal states or representations that are causally connected to the environment through perception, (2) being able to reason explicitly about the environment, and (3) being able to communicate with an external agent about the environment.

Architectures for understanding/modeling behaviors of an anthropomorphic agent, e.g., cognitive architectures [Langley et al. 1991][Pollock 1989][Anderson 1983], tend to address the relationships that exist among the structure of memory, reasoning abilities, intelligent behavior, and mental states and experiences. These architectures often do not take the body into account. Instead they primarily focus on the mind and consciousness. Our architecture ranges from general knowledge representation and reasoning to body-dependent physical behavior, and the other way around.

We are interested in autonomous agents that are embedded in a dynamic environment. Such an agent needs to continually interact with and react to its environment and exhibit intelligent behavior through its physical activity. To be successful, the agent needs to reason about events and actions in the abstract as well as in concrete terms. This means combining situated activity with acts based on reasoning about goal-accomplishment, i.e., deliberative acting or planning. In the latter part of this paper, we will present a family of agents based on our architecture. These agents are designed with a robot in mind, but their structure is also akin to anthropomorphic agents. Figure schematically presents our architecture.

There are several features that contribute to the robustness of our architecture. We highlight them below (an in-depth discussion follows later):

lammens@cs.buffalo.edu