
Bridging Rigorous Methods with Machine Learning
Chuchu Fan, an Associate Professor at MIT's AeroAstro department and Laboratory for Information and Decision Systems, presents this Stanford robotics seminar on making learned controllers safe for real-world deployment. She lays out why learned policies resist safety guarantees and why rare failure scenarios are hard to find efficiently, then describes a framework that closes the loop between synthesizing safe policies and falsifying them through targeted testing. Fan shows how neural certificates can be framed as value functions within reinforcement learning, so stability and safety become built into the policy's structure rather than an added penalty term. She also covers an adaptive test-case generation method using Bayesian active learning to surface failure scenarios that are both revealing and broadly useful for coverage, including how the approach can predict failures from limited hardware demonstrations and support evolving, system-level experimental design.