Modern autonomous systems are composed of interacting components for control, scheduling, and learning, each operating under practical limitations such as resource constraints, timing uncertainties, or perception errors. Traditional design and verification approaches aim for component-level perfection, an assumption that is increasingly untenable for complex systems. This dissertation develops methods for quantitative, system-level safety that explicitly account for imperfect components and characterize how their combined effects impact closed-loop behavior. Using quantitative safety metrics such as deviation from ideal behavior and reachable-set size, this dissertation considers four major classes of imperfect components: resource-constrained computational platforms, systems with probabilistic timing behavior, deep neural network–based perception systems, and imperfect tools and validation processes themselves when designing autonomous systems. Taken together, they provide a basis for systematically designing safe autonomous systems without requiring their individual components to be perfect.
Shengjie Xu (2026) studied this question.