Critical slowing down for predicting controller induced loss of control in quadrotors

What the math gives to ML

The paper offers a model-free early-warning mechanism for controller-induced loss of control: critical slowing down causes recovery from perturbations to become slower before a stability boundary is crossed. This appears statistically as increasing variance, increasing lag-one autocorrelation, and a declining recovery rate in rolling time series, without requiring labeled loss-of-control trajectories or a plant model. The strongest neural-network transfer is an online monitor for recurrent policies, state-space models, neural controllers, or optimizer trajectories that detects approach to instability and triggers damping, step-size reduction, or safe fallback before divergence. The key falsifiable signature is that the estimated AR(1) coefficient approaches one and the variance grows according to the inverse restoring rate.