Towards Trustworthy
Physical AI
My research focuses on the development of algorithms that let autonomous systems cross from simulation into the real world with guarantees.
» As new legislation tightens the screws on AI in high-risk applications, we need tools that can provably certify that an AI-powered system is safe and trustworthy when deployed in the real world.
Research directions
A few of the research threads I'm interested in.
Bridging sim-to-real
A probabilistic simulation framework so controllers proven safe in simulation transfer to the real world with quantified guarantees.
Formal verification
Certifying black-box stochastic and learning-enabled systems from finite data — quantified safety where classical tools break down.
Data-driven abstractions
Scalable abstractions of uncertain stochastic systems — like binary-tree Gaussian processes — built for efficient model checking.
Selected projects
Each is a research line spanning several papers, tools, and collaborations.
SySCoRe
Correct-by-design control for uncertain stochastic systems — infinite-horizon specs, non-Gaussian noise, tensor-scaled synthesis.
5+ papers · HSCC · CDC · TACBinary-Tree Gaussian Processes
Data-driven abstractions that naturally partition the state space, simplifying error quantification for formal verification.
2 papers · IFAC ADHSRobust Multi-Objective Reasoning
Balancing competing objectives under uncertainty — from multi-objective physics-guided models of dynamical systems to spatiotemporal robustness of temporal-logic tasks.
2+ papers · CAV · IFAC ALCOSLet's build
trustworthy autonomy.
oschoen@ethz.ch
Open to collaborations, talks, and questions on formal verification, safe control, and learning-enabled systems.