Building Trustworthy Physical AI Systems: From Theory to Large-Scale Deployment

Over the past decade, artificial intelligence has transitioned from laboratory demonstrations to real-world deployment in safety-critical physical systems. This transformation presents unprecedented opportunities alongside significant challenges in ensuring safety, generalizability, and trust at scale. In this talk, I will present the evolution of our lab's research in developing trustworthy physical AI systems, from foundational theoretical frameworks to large-scale industrial deployment. I will begin with rare-event safety: deep importance sampling and safe reinforcement and imitation learning methods that handle sparse failures in safety-critical autonomy. I will then highlight generalizable safety, including prompt-based transfer, task discovery, and morphology generalization so systems adapt reliably with minimal demonstrations.