Offline Learning of Neural Control Barrier Functions with Out-of-Distribution Detection

Abstract

:Control Barrier Functions (CBFs) have emerged as a powerful tool for ensuring safety in 
robotic systems. Traditional learning-based CBF methods rely on extensive state-space sampling or 
online system interactions, making them impractical for high-dimensional applications. This paper 
presents a novel Offline Learning Framework for CBFs, integrating Out-of-Distribution (OOD) 
Detection to enhance safety and generalization. Our approach learns CBFs from a fixed, sparsely 
labeled dataset, utilizing OOD-aware annotation techniques to propagate safety constraints efficiently. 
We evaluate the method on real-world robotic platforms, demonstrating superior safety performance in 
dynamic environments compared to existing methods.