Specification Guards for Neuro-Symbolic Autonomy

Abstract

Machine learning offers the potential to solve complex real-world tasks that elude traditional modeling, yet the solutions often lack the robustness required for safety-critical deployment and require high-quality data, which is expensive to generate. Conversely, model-based solutions provide rigorous in-domain guarantees but demand significant engineering effort and struggle in unstructured environments. To address this tension, my research introduces specification guards, which integrate scalable formal methods with machine learning to ensure safe and data-efficient autonomy. I demonstrate the real-world readiness of specification guards across diverse applications, including autonomous vehicles, biomolecular processes, and maritime navigation. In this talk, I present two primary research thrusts demonstrating how specification guards (1) provide rigorous behavioral assurances and (2) enable learning under data limitations. First, I introduce a generalizing specification guard for real-time collision avoidance capable of handling dynamic obstacles and uncertainty. I then present research integrating more complex specifications, transitioning from basic collision avoidance to general behavioral requirements. Second, I’ll highlight two approaches that make abstract domain knowledge computationally tractable and guide learning. I conclude by outlining a vision for reliable, rapidly deployable autonomy that leverages unstructured domain knowledge and improves soundly through real-world interaction.

Biography

Hanna Krasowski is a postdoc in the Arcak lab at the EECS department of UC Berkeley. Her research combines formal methods and machine learning to develop safe and data-efficient decision making for real-world systems with applications in maritime motion planning and biomolecular modeling. She earned her Ph.D. from the Technical University of Munich in 2024 and was a visiting researcher at Caltech in 2022. She is the principal developer of CommonOcean, a benchmarking and software suite for research on maritime navigation. Hanna was selected as an RSS Pioneer and EECS Rising Star 2025.