Dissertation Talk: Design and Analysis of learning-based cyber-physical systems
Seminar: Dissertation Talk: EE | May 13 | 1:30-2:30 p.m. | 531 Cory Hall
Shromona Ghosh, University of California, Berkeley
We are finally at a point where we have the knowledge and resources to make safety-critical robotic systems a reality. Deploying such systems in the real world, however, requires addressing problems in detecting unsafe environments, reasoning with unknown or learnt components and providing strong safety assurances. In this talk, I will address two specific challenges in this domain: (1) developing techniques that can reason with machine learning (ML) components such as perception and decision making modules and make them more robust; and (2) bridging the gap between design time (which uses abstract system model) analysis and real world operation of such systems.
To address these challenges, I will first present VerifAI a toolbox for systematic testing and analysis of CPS systems in simulation. I will illustrate through video simulations how VerifAI can find unsafe environments when an autonomous car (AV) fitted with a perception system is navigating in urban areas. VerifAI is open-sourced and in early stages of adoption.
I will then present a specification centric simulation metric (SPEC), a model metric that computes the behavior mismatch of CPS and its model. Our framework utilizes SPEC to find safe environments and controllers for the real system. I will show how using SPEC we are able to design a lane keeping controller for a AV car with imperfect perception using only a simplified physics model with perfect state estimation.