Thema der Dissertation:
Automated Rule Formalization and Layered Knowledge Integration for Safety-Aware Autonomous Systems Bridging Formal Methods, Large Language Models, and Uncertainty-Aware Learning Thema der Disptutation:
Knowledge-Guided Autonomy: Formalization, Uncertainty, Prediction, and Planning
Automated Rule Formalization and Layered Knowledge Integration for Safety-Aware Autonomous Systems Bridging Formal Methods, Large Language Models, and Uncertainty-Aware Learning Thema der Disptutation:
Knowledge-Guided Autonomy: Formalization, Uncertainty, Prediction, and Planning
Abstract: Autonomous systems must operate safely in environments shaped by rules, uncertainty, and interaction, yet purely data-driven approaches remain limited in safety-critical settings. This talk examines how explicit knowledge can support learning and decision-making by transforming natural-language rules into machine-actionable representations for verification, uncertainty-aware prediction, and rule-guided planning.
Drawing on examples from autonomous driving and robotics, the talk argues that no single representation suffices across the full autonomy pipeline.
Formal specifications provide precision and verifiability, probabilistic models capture uncertainty, and learning-based components offer flexibility. A central theme is that uncertainty is not monolithic: aleatoric uncertainty reflects irreducible variability in agent behavior, while epistemic uncertainty stems from limited model knowledge and can be reduced with more data. Conflating the two leads to overconfident predictions precisely where caution matters most. The talk further shows how the same underlying rule knowledge can serve multiple roles: as a formal specification for verification, as procedural signals that shape learning, and as constraints guiding planning under multimodal uncertainty. Reliable autonomy emerges not from choosing between rules and learning, but from combining them according to the demands of each task.
Drawing on examples from autonomous driving and robotics, the talk argues that no single representation suffices across the full autonomy pipeline.
Formal specifications provide precision and verifiability, probabilistic models capture uncertainty, and learning-based components offer flexibility. A central theme is that uncertainty is not monolithic: aleatoric uncertainty reflects irreducible variability in agent behavior, while epistemic uncertainty stems from limited model knowledge and can be reduced with more data. Conflating the two leads to overconfident predictions precisely where caution matters most. The talk further shows how the same underlying rule knowledge can serve multiple roles: as a formal specification for verification, as procedural signals that shape learning, and as constraints guiding planning under multimodal uncertainty. Reliable autonomy emerges not from choosing between rules and learning, but from combining them according to the demands of each task.
Time & Location
Jul 07, 2026 | 10:00 AM
Seminarraum 140
(Fachbereich Mathematik und Informatik, Arnimallee 7, 14195 Berlin)