End-to-end safety guarantees for controllers using statistical perceptual abstractions

Develop control synthesis frameworks that provide provable end-to-end safety guarantees for autonomous systems whose perception and state estimation are abstracted statistically (e.g., via conformal prediction), ensuring that safety holds from raw sensor measurements through estimation to actuation under measurement noise and unknown dynamics.

Background

Section 4.4 discusses statistical perceptual abstractions for learned perception and state estimation, constructed via conformal prediction over $\epsilon$-nets and continuity assumptions on sensors and estimators.

While initial progress integrates such abstractions into robust control, guaranteeing end-to-end safety—from sensors to control outputs—across the full autonomy stack remains unresolved.