Uncertainty-Aware Safe Control for Autonomous Mobile Systems: Integrating Model-Based Control with Learning-Based Uncertainty Models

Industrial-scale autonomous mobile systems (AMS), such as automated vehicles and mobile robots, transform industries and transportation by enhancing advanced features related to safety, sustainability, and energy efficiency. Trust worthiness, as a non-negotiable system requirement, demands safe operation to ensure user confidence and societal acceptance. Autonomy relies on various sensors, which serve as the ’eyes’ of AMS, supplying critical signals to perception and decision-making functions. However, in real operation situations, sensor performance can be degraded due to intrinsic issues such as unexpected noise, unmanaged physical damage, unavoidable component aging, or external environmental impacts such as unexpected weather conditions. Meanwhile, the perception functions receiving such sensor signals can produce non-optimal outputs not only because of flawed inputs, but also due to the probabilistic nature of algorithms used for tasks like localization. In effect, such vulnerabilities in sensory and perceptual functions introduce unexpected nondeterminism into an AMS control system. This thesis focuses on developing an uncertainty-aware safe control strategy for safety-critical AMS accounting for uncertainties stemming from the sensor components and perception functions. It aims to leverage machine learning, dynamic optimization, and control theory to facilitate the safe operation of AMS. It proposes a framework that builds on: (1) model predictive control (MPC) for con troller design; (2) control barrier functions (CBFs) for safety filter design; and (3) an uncertainty model for treating situation variability in safety constraint design. The contributions of this thesis are: (1) the development of a fault injection platform for generating data under various operational conditions, which serves as the core stage for subsequent developments; (2) an integration of multifunction control based on safe operational metrics; (3) the construction of dynamic safety constraints using CBFs according to safe operational requirements; (4) the development of a condition monitoring service for computing health and risk indices and for modeling perception uncertainties; and (5) the development of safety constraints given by the uncertainty models, and the indices. The approach advances safety-critical control design by integrating MPC–CBF methods with learned uncertainty models. Support Vector Regression and Long Short-Term Memory methods are employed to capture perception uncertainty under varying weather conditions, as well as uncertainty prediction over time. To incorporate the uncertainty prediction model into the optimization problem, the Learning Parametrized Convex Function method is used to construct a convex uncertainty model. In addition, statistical algorithms are employed to capture uncertainty models influenced by component aging and degradation. The proposed approaches are validated through comprehensive simulation studies.