Representation of Local Space in Perception/Action Systems: Behaving Appropriately in Difficult Situations

There have been two main approaches to determining action for autonomous agents: classical planning and reactive planning. Classical planners can construct plans which take into account complex interactions between the various actions the agent may take, but is computationally expensive, and requires complete knowledge of a static environment. In contrast, reactive planning systems simply map sensor inputs to actions. These mappings may be done in constant time, and the perception-based nature of reactive systems enables the agent to cope with dynamic and uncertain environments. However, reactive planning abandons the machinery needed to contend with complex situations.
This dissertation presents a new paradigm for interaction with complex, dynamic, three-dimensional environments which builds on the reactive approach. The centerpiece of this paradigm is the effective field of view, as implemented by marker-based representations of the local environment. The effective field of view is an extension of the standard field of view of a sensor, via representation of past sensor inputs. The effective field of view endows the agent with more information regarding the environment than the direct sensor inputs alone. By judiciously extracting and representing information for inclusion in the effective field of view based on the relevance of the information to a task, i.e., by marking the useful information, the competence of an autonomous agent is increased beyond that achievable by agents constructed using the pure classical or pure reactive approaches.
The effective field of view paradigm is demonstrated via an agent that interacts with a dynamic, three-dimensional, hostile virtual environment using visual perception alone. This agent is modelled after an herbivore which must collect food while avoiding obstacles and a predator. The addition of marker-based local-space representations to expand the effective field of view is shown to measurably increase the performance of such an agent.
The representations used to expand the field of view are amenable for use with advanced classical planners which relax the complete information assumptions required by older planners. This dissertation sets the roundwork for the construction of agents which capitalize on the strengths of both classical and reactive planning paradigms.