Incorporating Spatial Representations at Multiple Levels of Abstraction in a Replicated Multilayered Architecture for Robot Control

Abstract

This paper discusses how spatial information at multiple levels of abstraction is used in ATLAS, a multilayered replicated architecture for robot control. ATLAS was specifically designed to address the real-time requirements of robot systems executing complex tasks in real-world environments. The architecture is multilayered in the sense that robot system control is performed at multiple levels of abstraction and at varying control rates, and it is also replicated, meaning that the fundamental activities of sensing, decisionmaking, and execution occur at each layer of the architecture. ATLAS differs significantly from many other robot control structures in that the layers of the architecture are not ordered along increasing levels of abstraction of the information being processed, but are rather organized as a multirate control system. Consequently, the faster control layers operate closer to the physical robot system and the slower control layers operate on top of faster ones. This arrangement allows ATLAS to respond to urgent control demands in real-time, while maintaining the ability to approach complex tasks in a planned and deliberative manner. Since each ATLAS layer requires access to sensor-based information, we outline how different levels of abstraction of spatial information are used in the architecture, and discuss how these representations are automatically generated from sensor data. These levels of abstraction of spatial information range from raw sensor data to sensor-close Occupancy Grid models, geometric representations and topological models. Each representation is used in the ATLAS layers in which it fits most naturally, both in terms of the information needs of those layers and their cycle time requirements. The paper concludes with an application of these concepts to a mobile robot executing repetitive navigational tasks. We show that, by monitoring the accumulation of information by the robot, ATLAS can switch between its active layers, allowing the robot to improve in the execution of its task over time and thereby to exhibit a skill acquisition behaviour.