Dynamics of behavior: Theory and applications for autonomous robot architectures

Abstract

Abstract Limitations both for the further development as well as for the actual technical application of autonomous robots arise from the lack of a unifying theoretical language. We propose three concepts for such a language: (1) Behaviors are represented by variables, specific constant values of which correspond to task demands; (2) Behaviors are generated as attractors of dynamical systems; (3) Neural field dynamics lift these dynamic principles to the representation of information. We show how these concepts can be used to design autonomous robots. Because behaviors are generated from attractor states of dynamical systems, design of a robot architecture addresses control-theoretic stability. Moreover, flexibility of the robot arises from bifurcations in the behavioral dynamics. Therefore techniques from the qualitative theory of dynamical systems can be used to design and tune autonomous robot architectures. We demonstrate these ideas in two implementations. In one case, visual sensory information is integrated to achieve target acquisition and obstacle avoidance in an autonomous vehicle minimizing the known problem of spurious states. In a second implementation of the same behavior, a neural dynamic field endows the system with a form of obstacle memory. A critical discussion of the approach highlights strengths and weaknesses and compares to other efforts in this direction.