Designing behavior networks for autonomous robotic systems
Sunil Cherian · 1995
Design paradigms for achieving intelligent behavior in autonomous robotic systems are undergoing a shift from symbolic reasoning paradigms to highly distributed, reactive strategies that fall under the rubric of behavior-based design. The unifying theme behind behavior-based robotics is the view that intelligent behavior in machines is the outcome of the global dynamics of collections of loosely interconnected task-achieving competence modules. In this dissertation two control architectures are developed with different methodologies for interconnecting behavior modules. These control architectures are validated within a computational environment by using them to control simulated creatures called animats. Analysis of the structure and dynamics of these control algorithms leads to the view that the collection of interconnected task-achieving behaviors, or behavior network, is best viewed as a complex dynamical system composed of many interacting components. This view leads to new insights about basic architectural requirements, system organization, and analysis methods. A review of some of the properties of naturally occurring, autonomous, complex systems suggests new directions of inquiry for behavior-based robotic systems. In particular, the concept of a doubly determinate representation hierarchy offers a novel framework for investigating bi-directional control relations between two adjacent layers in the hierarchy. Of two such layers, the lower layer constitutes a micro-description of the system and the higher layer constitutes a macro-description of the same system. The higher layer achieves its macro-description within a representation space of lower dimensionality than that employed by the lower layer. This concept of double determinacy and its relevance to behavior-based robot control is investigated in this dissertation. In order to pursue this investigation, the concept of a collection of $\Delta$ machines, forming a $\Delta$-network, is introduced. $\Delta$-networks constitute a mechanism for formalizing the autonomous nature of each task-achieving behavior and their causal interactions with each other in a behavior network. The analysis of the state-space dynamics of a simple behavior network, considered as an instance of a $\Delta$-network, provides insight into the characteristic dynamics of behavior networks. The development of control architectures and the analysis of behavior networks presented in this dissertation is aimed at extending our understanding of the synthesis of coherent global behavior in robots from collections of low level task-achieving behaviors.