The Animate Agent Architecture

An autonomous, intelligent agent requires a variety of capabilities. It must be able to sense the world and control its actions in real-time. It must also be able to execute complex, multi-step plans, confirm that plan steps have had their intended effects, and deal with problems and opportunities as they arise. An agent's ability to carry out complex plans in dynamic environments hinges on knowing large numbers of different methods for carrying out a task in different situations and for detecting problems and correcting them.

Modularity and reusability are especially important aspects of any plan library the agent uses to encode these methods. An agent must be able to pursue a variety of different goals at different times and each goal will require different plans. Even a single plan will consist of multiple steps that each corresponds to a new subgoal for the agent. Many of these subgoals will correspond to conceptually identical tasks like picking an object up or moving to a specific location. If it is possible to reuse the same reactive plans for these subgoals, even when they are used as steps in different higher-level plans, the agent's behavior will be much easier to program, much more reliable, and the plans themselves will be easier to adapt and learn.

The Animate Agent Project at the University of Chicago is building an agent control system that integrates reactive plan execution, behavioral control, and active vision within a single software framework. The architecture splits the agent's representation of actions into two levels: one consisting of continuous, real-time active vision and control processes that can be combined in different ways at run-time to create different behavior, and another consisting of hierarchical methods of discrete steps for achieving goals and coping with contingencies in the world. Modularity is emphasized in both levels so that sensing and action processes can be combined in multiple ways and task methods can share common subtasks.

Within the Animate Agent Architecture there are several areas where standardization efforts are relevant. The architecture itself is an example of the 3T architectural design in use at Johnson Space Center and a variety of other institutions involved in mobile robot work. In addition, the interface between the discrete and continuous control layers relies on a relatively universal representation that applies in a wide variety of domains. Finally, the vision processing system we have developed for building active vision routines at run-time is specifically designed to encapsulate vision algorithms in a standard way.


R. James Firby

Artificial Intelligence Laboratory
Computer Science Department
University of Chicago
1100 East 58th Street
Chicago, IL 60637

firby@cs.uchicago.edu