Learning in a multiresolutional conceptual framework

Abstract

A real-time system for the control of an autonomous vehicle consisting of a nested hierarchy of control modules is discussed. The proposed intelligent controller has nonhomogeneous knowledge representation and a neural-network-based decision-making system operating in real time. The focus is on the Pilot module, which provides the real-time guidance of the system. It is responsible for the generation and tracking of dynamically feasible trajectories which follow the planned path given by the upper level (Navigator) and avoid local obstacles. Control of a complex system (mobile robot) is facilitated by the use of a feedforward neural network. How such an approach addresses constant response time of decision-making (control) and online learning and adaptability is discussed. Dealing with constraints is done via a multiresolutional system of dynamic avoidance regions, which are analogous to the concept of potential field by require much simpler representation and computational procedures.<>