Tag
This paper introduces certified safety curation for safe offline reinforcement learning, using a filter-then-clone pipeline with distribution-free guarantees to select safe trajectories without needing cost functions on every transition.