Runtime Assurance

Runtime assurance studies how to check and, when needed, modify or block a system’s actions while it is running, especially when a learned component cannot be trusted to behave safely in every situation. Common approaches include Simplex architectures with verified backup controllers, control-barrier-function and rule-based shields, and monitors that use contracts, reachability, or statistical risk bounds to trigger recovery or abstention. These methods are tested in settings such as robot manipulation, autonomous driving, spacecraft control, power grids, and tool-using agents, where they must balance safety with task success, latency, and operator control. A central challenge is making guarantees hold under partial observation, disturbances, and changing conditions without making interventions so conservative that the system becomes ineffective; many results still depend on stated assumptions or bounded evaluations.