Abstract:

Performance of control systems depend on the nature of available measurements. Perception of edges, keypoints, landmarks and other natural semantic features make certain coarse measurements available to control systems operating in complex environments.  This talk explores two problems related to control and estimation with such coarse measurements. First, I will introduce perception contracts—an approach for analyzing visual control systems that rely on Deep Neural Networks for state estimation. A perception contract provides an over-approximation of a state estimator while guaranteeing closed-loop system invariants. These contracts can be automatically synthesized using data and model-based analysis and have been applied to automated landing and lane-keeping systems. The second part of the talk will focus on algorithms for computing indistinguishable sets—sets of states that cannot be distinguished based on available coarse measurements. These sets help define the limits of state estimation and localization. The talk will touch on formal verification, information spaces, and related open problems.

Zoom Link (if unable to attend in-person): https://upenn.zoom.us/j/91934138074