Events

ECE Seminar: Guiding Reinforcement Learning with Formal Methods for Reliable and Efficient Real-world Systems

Speaker: Hanna Krasowski

Location: 370 Jay Street, Room 825

Date: Monday, November 3, 2025

Machine learning models achieve remarkable performance when large amounts of training data are available. However, data is often limited for many real-world systems such as biological systems and space exploration robots. Additionally, many real-world systems are safety-critical, and learning-based models usually do not exhibit safety guarantees, especially for out-of-distribution data.

In this talk, I will introduce provably safe reinforcement learning (RL) approaches that ensure hard safety guarantees for RL both during training and deployment by leveraging formal methods. In particular, we will discuss a safeguard for nonlinear system dynamics in dynamic environment that ensures collision avoidance, and continuous action masking that leverages domain knowledge encoded in convex action sets to focus RL exploration.

About Speaker
Hanna Krasowski is a postdoc in the Arcak lab at the EECS department of the University of California Berkeley and affiliated with the Berkeley Artificial Intelligence Research Lab. Her research combines formal methods and machine learning to develop safe and data-efficient models 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 working in the Cyber-Physical Systems group and was a visiting researcher in the AMBER lab at Caltech in 2022. She is the principal developer of CommonOcean a benchmarking and software suite for maritime navigation. Hanna was selected as an RSS Pioneer and EECS Rising Star 2025.