Oracle-Guided Differentiable Safeguards for Safe Flow-Matching Robot Policies

A new method allows robot foundation models to learn safe policies using only binary feasibility oracles for physical constraints. The safeguard uses binary queries to correct unsafe updates during inference and extracts local boundary information to provide training signals without requiring differentiable constraint models.