Robust Safe Control Under Uncertainty: Estimation, Contraction, and Regret Bounds

Abstract

Disturbance uncertainty requires safety filters based on control barrier functions (CBFs) to leave a robustness margin, which can alter the nominal control input and degrade closed-loop performance. This paper studies safety and the resulting performance loss for discrete-time nonlinear systems with additive disturbances and polytopic input constraints. We develop a causal disturbance observer and construct two robust barrier conditions. The first uses a time-dependent bound on the estimation error. The second uses a constant margin when the disturbance increments are uniformly bounded and moves the requirement on the initial error from the online filter to an explicit initialization condition. For affine barriers, both filters are quadratic programs. We bound their input deviations from an oracle filter with access to the current disturbance and derive sufficient conditions for uniform contraction of the oracle closed loop. The contraction analysis accounts for state-dependent actuation, unmatched disturbances, and active physical input limits. Combining the sensitivity and contraction estimates gives finite-time dynamic regret bounds that quantify the effects of observer convergence, disturbance variation, and robustness margins. With sublinear cumulative certified disturbance variation and observer innovation error, the bounds yield sublinear regret.