Abstract

Obstacle avoidance is a difficult task for autonomous robots. To overcome limitations of traditional computer vision systems, some robots have made use of efficient VLSI sensory-motor systems. However, because the processing in these systems is at the pixel level, it is difficult to achieve algorithms that can deal with real-world, unstructured environments. To make these VLSI sensory-motor systems more widely applicable, new architectures and strategies are needed. This thesis presents an architecture for a VLSI sensory-motor system designed for obstacle avoidance by a mobile robot in an unstructured environment. Drawing inspiration from biology and behavior-based robotics, the development of the architecture is guided by an emphasis on the requirements of an obstacle avoidance behavior for a mobile robot. The architecture incorporates features which enable it to deal with unstructured environments. A special foveation and weighting scheme are used to facilitate the detection of real-world objects. The sensory and motor maps of the system are aligned to relate features in the visual field to a left and a right control signal. The effectiveness of the architecture is demonstrated through computer simulation. A model of a sensory-motor system based on the architecture is used to create a realistic looking virtual environment simulator. Simulation results show that such a system is capable of efficient obstacle avoidance in an unstructured environment, while using only a small number of simple operations connected hierarchically, potentially leading to an implementation with small pixels