DYNAMIC FIELDS ENDOW BEHAVIOR-BASED ROBOTS WITH REPRESENTATIONS

The behavior-based approach to autonomous robotics has been quite successful at designing systems capable of simple reactive behaviors that are flexibly activated. Progress towards larger behavioral complexity and towards the equivalent of cognitive abilities has been hampered, however, by the very absence of representations in this approach. Also, a need for a theoretically analyzable general architecture for system integration has been recognized. We propose an architecture based on dynamic neural fields which aims to deal with both of these shortcomings. Sensory processing, representation, planning, and control are all addressed in the same terms by defining adequate levels of description. At each level a dynamic neural field with strong internal interactions is capable of generating behavior or its representation. Levels are mutually coupled by specifying for each other adequate representations. We demonstrate the architecture by solving the standard problem of target acquisition and obstacle avoidance for a vehicle moving in two dimensions. The limited viewing angle of a visual sensor is used to demonstrate in exemplary fashion the properties of subsymbolic memory as implemented by the dynamic fields. We show how representation alleviates the problem of spurious states in potential field approaches. Also, we demonstrate how memory enhances the system's capability to escape from closed-in situations (e.g., canyon boxes), and, more, generally, endows the system with cognitive features.