Dynamic field architecture for autonomous systems

Abstract

System integration is the central problem for the design of autonomous robots. While methods from the theory of dynamical systems are routinely used both for planning (in potential field approaches) and for control, we argue, that the processes of creating, updating, merging and deleting instances of behavior can likewise be addressed through concepts of dynamics. The key idea is to invoke the principle of neural representation on continuous topological spaces, which describe behavioral dimensions. Architectures based on dynamic neural fields can provide a common framework for all levels of sensory information processing, planning, and control. The crucial step is to consider the limit of strong intra field interaction, which leads to functionalities as varied as representation of memorized information and nonlinear control dynamics. We illustrate the architecture through a simple model system, which solves target acquisition, obstacle avoidance, and memorization of obstacle information.