Autonomous systems rely heavily on various sensors combined with advanced perception algorithms to create an accurate understanding of the current operating environment. In order to reach a safety-critical control decision, the information of these sensors is fused step-by-step until a unified environment model is created. During this fusion, sensor values may disagree, leading to the question which influence this has on uncertainty and risk. Even though calculation methods exist for determining the risk stemming from sensor failures when integrated into one of the well-established fault-tolerance architectures, such as a 2003 architecture, precisely determining how risk is affected by various sensor fusion methods remains unanswered. To bring clarity to this topic, this work developed the fundamental theory that allows sensor fusion methods to be quantitatively considered in a risk assessment process. Moreover, this work identified the importance of propagating uncertainty information throughout a system’s architecture and differentiating between expected and safe values in order to reduce risk and increase utility. To demonstrate this, a variant of the Doer-Checker architecture, the so called Counter-Player architecture, is extended to be able to differentiate between safe and expected values, hence allowing the system to increase utility while ensuring safety.
Philipp Schleiss (Wed,) studied this question.