Title: Indistinguishability and Abstract Rendering for Certifiable Vision-Based Control
Abstract: This talk explores certifiable control and state estimation under coarse or partial observations—for example, when a robot detects a landmark only upon entering its field of view, or when an aerial vehicle perceives the world through pixelated camera images. I will first discuss the problem of localization with respect to unknown landmarks and characterize its fundamental limits in terms of indistinguishable states. I will then introduce the notion of abstract rendering, a method that computes the set of all images generated from uncertain camera poses in scenes represented by neural radiance fields or Gaussian splats. This enables formal certification of downstream modules for state estimation and control and forms the foundation of a computational library for the rigorous design and analysis of AI-enabled vision-based control systems.
Location & Time: CSL B02, October 22, 3-4PM
Reception in CSL 154 at 2:30PM