A neuro-computational mobile robot architecture for mapping localization and navigation
Max Donath, Vassilios Morellas · 1995
This dissertation has focused on the development of a mobile robot architecture, which allows robots to behaviorally mimic naturally intelligent systems. The specific problem which is addressed is robot navigation in unknown, unstructured and dynamically changing environments. Our approach assumes no a priori information: the robot perceives the environment solely by sensing its environment as it moves. The body of the thesis consists of three major parts: (1) Development of a procedure which allows a robot to develop local maps of different working environments in real time, by processing incoming range data, (2) Formation of view invariant representations (landmark representations) and (3) Fuzzy classification of the landmark representations. Formation of these maps is obtained through the use of a Self-Organizing Mapping (SOM) neural network. These maps represent a robot's own local view of its surrounding environment and represent discretized representations of the continuous world space. View invariant representations are extracted by processing SOMs using a logarithmic scaling procedure which is motivated by physiological observations. Finally, classification of landmark representations is obtained through an Adaptive Resonance Theory (ART), self-organizing neural network which incorporates fuzzy learning. Classification of view invariant representations is used to solve the problem of robot localization. The methodology was tested and demonstrated by using an actual mobile robot. Overall, this research integrates ideas from two different approaches which dominate the field of mobile robotics: (i) Artificial Intelligence (AI) and (ii) behavior-based robotics. Our results enhance the behavior-based paradigm by introducing behaviors which are biologically inspired. This last element allows us to integrate the previous two approaches, by showing that symbolic and non-symbolic structures which are developed by the robotic agents themselves can coexist in the same architecture allowing robots to be adaptable, autonomous and intelligent.