Towards Robust and Safe Decision-Making Under Uncertainty: A Model-Centric View

Abstract

Intelligent autonomous systems are increasingly deployed in safety-critical domains operating in open, complex, and potentially adversarial environments. Ensuring the trustworthiness of their decision-making requires addressing fundamental challenges, such as uncertainty in agent-environment interactions, high-dimensional state spaces, and partial observability. In this paper, we present a unifying perspective on our ongoing work toward safe, robust decision-making in autonomous systems. We argue that adequate symbolic models (whether designed or learned from observations) are essential for achieving trust by design. Building on this insight, we outline a neuro-symbolic decision-making framework that integrates model checking, controller synthesis, and reinforcement learning through a continuous loop of model construction and data-driven refinement, verification, and testing, both at development time and runtime. Finally, we propose a research roadmap toward hybrid verification approaches that combine learning and formal methods to enable adaptive, scalable, and trustworthy autonomous systems utilising agentic AI. This vision highlights the role of model learning and model-based verification and synthesis for bridging the gap between decision-level guarantees and efficient data-driven learning.