Reliable and safe autonomy for ground vehicles in unstructured environments

Abstract

This thesis is concerned with the algorithms and systems that are required to enable safe autonomous operation of an unmanned ground vehicle (UGV) in an unstructured and unknown environment; one in which there is no speci c infrastructure to assist the vehicle autonomy and complete a priori information is not available. Under these conditions it is necessary for an autonomous system to perceive the surrounding environment, in order to perform safe and reliable control actions with respect to the context of the vehicle, its task and the world. Speci cally, exteroceptive sensors measure physical properties of the world. This information is interpreted to extract a higher level perception, then mapped to provide a consistent spatial context. This map of perceived information forms an integral part of the autonomous UGV (AUGV) control system architecture, therefore any perception or mapping errors reduce the reliability and safety of the system. Currently, commercially viable autonomous systems achieve the requisite level of reliability and safety by using strong structure within their operational environment. This permits the use of powerful assumptions about the world, which greatly simplify the perception requirements. For example, in an urban context, things that look approximately like roads are roads. In an indoor environment, vertical structure must be avoided and everything else is traversable. By contrast, when this structure is not available, little can be assumed and the burden on perception is very large. In these cases, reliability and safety must currently be provided by a tightly integrated human supervisor. The major contribution of this thesis is to provide a holistic approach to identify and mitigate the primary sources of error in typical AUGV sensor feedback systems (comprising perception and mapping), to promote reliability and safety. This includes an analysis of the geometric and temporal errors that occur in the coordinate transformations that are required for mapping and methods to minimise these errors in real systems. Interpretive errors are also studied and methods to mitigate them are presented. These methods combine information theoretic measures with multiple sensor modalities, to improve perceptive classi cation and provide sensor redundancy. The work in this thesis is implemented and tested on a real AUGV system, but the methods do not rely on any particular aspects of this vehicle. They are all generally and widely applicable. This thesis provides a rm base at a low level, from which continued research in autonomous reliability and safety at ever higher levels can be performed.