DrYulong Gao

My research asks how we can trust autonomous systems that learn. Modern autonomy increasingly relies on components such as perception networks, generative planners, world models, whose behaviour is learned rather than designed, and for which classical guarantees do not apply. The Safe Autonomy and Intelligence Lab (SAIL) develops the mathematics needed to close that gap: control theory, formal verification and probabilistic reasoning brought together so that safety can be certified for systems whose inner workings are learned, and so that those certificates remain meaningful when the system is deployed in the physical world.

In detail, we are interested in the following research directions. 

• Control under uncertainty and partial observation. Distributional reachability, distributional and risk-sensitive optimal control, and POMDP synthesis, for systems where uncertainty must be quantified rather than bounded away.

• Set-based safety certificates for learning-enabled control. Invariant covers, probabilistic controlled invariant sets and temporal logic trees give computable representations of what a system can guarantee. We use these as safety contracts around synthesized controllers. 
• Safety over learned representations. When planning happens in a latent space, safety must be defined there too. We study when a guarantee certified over a learned model transfers to physical safety, and what fidelity of the representation this requires.