“Safety Verification for Robots Controlled with Learned Models”
Location: 177 Huntington Ave, conference room 503
Abstract: Neural network controllers and world models are increasingly used in autonomous robots that operate in the real world. Ensuring the safety of such systems under uncertainty is challenging. This talk will cover some of my lab’s recent work on formalizing these problems and developing algorithms to provide safety assurances. In particular, I will discuss forward and backward reachability techniques, methods for verifying that a team of robots won’t collide with one another, and a new way to reason about controllers that map directly from camera image inputs to commanded actions (“pixels-to-torques”).
Bio: Michael Everett, Assistant Professor in ECE. His research lies at the intersection of robotics, deep learning, and control theory, with the goal of developing certifiable learning machines.