From Safety Filtering to Safety Forecasting

Control Barrier Functions (CBFs) provide a powerful way to enforce safety constraints by filtering a nominal controller before unsafe actions are executed. However, the intervention often becomes most visible only when the system is already close to a safety boundary. In this project, we investigate a different question: can the safety filter itself tell us that maintaining safety is becoming difficult before the situation becomes critical? Our current observations suggest that the penalty and auxiliary signals used by Adaptive Control Barrier Functions (AdaCBFs) may contain early information about diminishing safety control authority. We study these internal dynamics as a potential predictive safety signal, compare them against conventional risk indicators such as barrier value and time-to-collision, and use them to trigger proactive changes in the nominal robot behavior. The final AdaCBF remains the hard safety layer, while the predictive signal informs higher-level decisions. We target human-aware multi-robot systems, where early warning can allow a robot team to slow down, yield, replan, or reconfigure before abrupt safety interventions are required.