Model-Free Based Computations of Recursive Control Barrier Function: Ultra-Local Model Approach

What the math gives to ML

The paper offers a transferable model-free safety mechanism: estimate an ultra-local input-output dynamics model online, quantify its prediction error with an uncertainty envelope, and insert that envelope directly into a robust control-barrier inequality. The key asset for neural networks is a safety wrapper for a learned controller or sequential predictor that does not require differentiating or identifying the full environment dynamics. For relative-degree-one safety outputs, the wrapper converts uncertainty into an explicit virtual-boundary shift, yielding a measurable conservatism-versus-safety tradeoff and a falsifiable violation threshold.