From Safe Perception to Trustworthy Autonomy: A Systems Engineering Approach with Digital Twins

Autonomous systems can only be as trustworthy as the perception pipelines that support their decisions. As AI and machine learning become central to perception in safety-critical domains, systems must remain dependable under uncertainty, faults, disturbances, and cyberattacks, while satisfying explicit risk constraints. This lecture frames trustworthy autonomy as justifiable autonomy: the ability to sustain dependable service and well-justified decisions as operating conditions evolve.

The talk argues that Model-Based Systems Engineering (MBSE) and Digital Twins provide a rigorous systems-level foundation for designing and assuring safe AI-enabled perception. Digital Twins are presented not merely as offline replicas, but as predictive run-time models that monitor the system and its environment, anticipate degradations, and support planning, adaptation, and safe reconfiguration through hierarchical feedback loops spanning sensors, functions, and system-of-systems levels.

These ideas are illustrated through a multi-sensor event-detection example, where redundant and heterogeneous sensing, sensor fusion, sensor trust and reputation mechanisms, and interpretable probabilistic models such as Dynamic Bayesian Networks improve robustness and transparency, and through the REXASI-PRO assistive “wheelchair-drone” platform for critical scenarios such as road crossing. The main message is that trustworthy autonomy is achieved not by AI alone, but by integrating AI with systems engineering, run-time assurance, and Digital Twins.