Safe training is not the same as safe deployment. Reinforcement learning policies may satisfy safety constraints during training yet violate them when the deployment environment shifts. Whether a robot encounters changing surface friction or a diabetes-management policy faces unseen patient dynamics, this safety generalization gap limits the reliable deployment of autonomous systems.
In this talk, I present my research toward safe decision-making under distribution shift through two complementary approaches: runtime safety shields that intervene when learned policies become unreliable, and safe in-context reinforcement learning methods that adapt to unseen task and safety configurations from interaction history, without retraining.
I demonstrate these approaches across robotic control and diabetes management simulations, and conclude by discussing future directions for scaling safe autonomy to richer specifications, including temporal logic and multi-agent coordination.
