Uncertainty Estimators for Robust Backup Control Barrier Functions
DAVID E. J. VAN WIJK 1, ERSIN DAŞ 2, ANIL ALAN 3, SAMUEL COOGAN 4, TAMAS G. MOLNAR 5, JOEL W. BURDICK 1, MANORANJAN MAJJI 6, KERIANNE L. HOBBS 7 · 2025
Read Uncertainty Estimators for Robust Backup Control Barrier Functions by DAVID E. J. VAN WIJK 1, ERSIN DAŞ 2, ANIL ALAN 3, SAMUEL COOGAN 4, TAMAS G. MOLNAR 5, JOEL W. BURDICK 1, MANORANJAN MAJJI 6, KERIANNE L. HOBBS 7 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-15734.txt.
Introduction
David van Wijk and co-authors make a safety method for machines with limited actuation work under uncertainty about how the machine behaves. Backup barrier methods keep a system safe by always holding a fallback plan in reserve. But if the model of the system is wrong, the predicted path of that fallback is wrong too, and the safety promise breaks.
Their fix estimates the model error online and rebuilds the predicted flow around that estimate, refining it as data arrive. The resulting controllers provably keep the system safe despite unknown disturbances. The approach is aimed at safety-critical machines, from robots to flight hardware, where a wrong model must not mean a crash.
Catalog record
| CREATOR | DAVID E. J. VAN WIJK 1, ERSIN DAŞ 2, ANIL ALAN 3, SAMUEL COOGAN 4, TAMAS G. MOLNAR 5, JOEL W. BURDICK 1, MANORANJAN MAJJI 6, KERIANNE L. HOBBS 7 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 62 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2503-15734
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Uncertainty Estimators for Robust Backup Control Barrier Functions DAVID E. J. VAN WIJK 1, ERSIN DAŞ 2, ANIL ALAN 3, SAMUEL COOGAN 4, TAMAS G. MOLNAR 5, JOEL W. BURDICK 1, MANORANJAN MAJJI 6, KERIANNE L. HOBBS 7 Abstract Designing safe controllers is crucial and notoriously challenging for input-constrained safety-critical control systems. Backup control barrier functions offer an approach for the construction of safe controllers online by considering the flow of the system under a backup controller. However, in the presence of model uncertainties, the flow cannot be accurately computed, making this method insufficient for safety assurance. To tackle this shortcoming, we integrate backup control barrier functions with uncertainty estimators and calculate the flow under a reconstruction of the model uncertainty while refining this estimate over time. We prove that the controllers resulting from the proposed Uncertainty Estimator Backup Control Barrier Function (UE-bCBF) approach guarantee safety, are robust to unknown disturbances, and satisfy input constraints. \pubyear 2026 \sptitle Article Category keywords Constrained control, Lyapunov methods, Robust control † † volume: 00 † † issue: XX † † paper-type: 1234567 † † received: August 2026 † † corresponding: CORRESPONDING AUTHOR: D. E. J. van Wijk (e-mail: [email protected] ) † † authornote: This work was supported by in part by DARPA under the LINC program and by the Technology Innovation Institute (TII). I Introduction Controllers that satisfy safety constraints are of paramount importance for many autonomous systems. Control barrier functions (CBFs) [ 1 ] offer a simple and effective approach for safety-critical control by providing sufficient conditions for forward invariance of safe sets. However, designing such safe sets for input-constrained systems remains a challenge, especially for high-dimensional dynamics and complex state constraints. Furthermore, CBFs rely on an accurate model of the system dynamics, but such models are rarely without errors in real-world applications. In this paper, we seek to address both of these challenges concurrently. The safety-critical control literature is rich with attempts to accommodate model mismatches. Robust methods [ 2 , 3 , 4 ] address disturbances typically through worst-case analysis that provides safety guarantees using an additional robustifying term. Input-to-state safety [ 5 ] addresses input disturbances [ 6 ] and can be made less conservative via a tunable robustness parameter [ 7 ] . Uncertainty estimators such as disturbance observers have been used to reconstruct a representation of the disturbance signal [ 8 , 9 , 10 ] . Adaptive methods, which are effective in handling parametric uncertainty, can also reduce conservatism [ 11 ] . Data-driven [ 12 ] and learning-based [ 13 ] approaches show promise in handling uncertainty in dynamics or states for real-world systems. While these approaches present viable solutions to addressing model uncertainties, they assume that such a safe set can be found explicitly that will lead to the satisfaction of input constraints—a strong assumption for most systems. To design safe sets in which every state has a safe control action (i.e., controlled invariant sets), we leverage backup control barrier functions (bCBFs) [ 14 , 15 ] which construct controlled invariant sets online by examining the predicted state evolution under a backup control policy. While this approach guarantees safety with input constraints, it is sensitive to model uncertainties because it requires forward integrating the uncertain model. To remedy this, our previous work [ 16 ] derived conditions for online controlled invariance in the presence of disturbances. We used an upper bound on the state evolution uncertainty through a worst-case analysis, resulting in conservative safety constraints in some cases. In this work, we introduce an approach to online controlled invariance in the presence of disturbances, and reduce conservatism via uncertainty estimators. Our main contribution is uncertainty estimator backup CBFs —a novel class of CBFs for the safety-critical control of input-constrained uncertain systems. The proposed method uses state predictions under the reconstruction of the disturbance to define a subset inside a controlled invariant safe set of the disturbed system. We derive forward invariance conditions for such a subset, and we show that these are made less conservative over time using the uncertainty estimator. We use these conditions to design robust safety-critical controllers, which account for the evolution of the uncertainty estimator error and for the sensitivity of state predictions to the estimated disturbance, instead of merely adding an uncertainty estimator to existing methods like [ 16 ] . This generalizes the proposed approach to be utilized across a large range of possible uncertainty estimators and predictors. We prove that the proposed approach guarantees safety for systems with limited control authority even in the presence of unknown, bounded disturbances. The remainder of the paper is organized as follows. Section II reviews CBFs and bCBFs. In Section III , we formalize uncertainty estimators in the context of implicit set safety, and present the proposed uncertainty estimator backup CBF approach. In Section IV , we cover the results of our numerical simulations, and we conclude with Section V . II Preliminaries II-A Control Barrier Functions Consider a nonlinear control affine system of the form \displaystyle\dot{\bm{x}}=\bm{f}(\bm{x})+\bm{g}(\bm{x})\bm{u},\mkern 8.0mu\bm{x}\in\mathcal{X}\subset\mathbb{R}^{n},\mkern 8.0mu\bm{u}\in\mathcal{U}\subseteq\mathbb{R}^{m}, (1) with continuously differentiable functions {\bm{f}\!:\!\mathcal{X}\!\rightarrow\!\mathbb{R}^{n}} and {\bm{g}\!:\!\mathcal{X}\!\rightarrow\!\mathbb{R}^{n\times m}} . We assume that the admissible input set \mathcal{U} is an m -dimensional convex polytope. For an initial condition {\bm{x}_{0}\triangleq\bm{x}(0)\in\mathcal{X}} , if \bm{u} is given by a locally Lipschitz controller {\bm{k}:\mathcal{X}\rightarrow\mathcal{U}} , {\bm{u}=\bm{k}(\bm{x})} , the closed-loop system has a unique solution. A notion of safety is defined by membership to a set \mathcal{C}_{\rm S} , and safe controllers render this safe set forward invariant. A set {\mathcal{C}\subset\mathbb{R}^{n}} is forward invariant along the closed-loop system if {\bm{x}(0)\in\mathcal{C}\!\implies\!\bm{x}(t)\in\mathcal{C},} for all {t>0} . Consider the safe set \mathcal{C}_{\rm S} as the 0-superlevel set of a continuously differentiable function {h:\mathcal{X}\rightarrow\mathbb{R}} with {\mathcal{C}_{\rm S}\triangleq\{\bm{x}\in\mathcal{X}:h(\bm{x})\geq 0\}} , where the gradient of h along the boundary of \mathcal{C}_{\rm S} remains nonzero. A function {h:\mathcal{X}\rightarrow\mathbb{R}} is a CBF [ 1 ] for ( 1 ) on \mathcal{C}_{\rm S} if there exists a class- \mathcal{K}_{\infty} function 1 1 1 The function {\alpha:\mathbb{R}_{\geq 0}\rightarrow\mathbb{R}_{\geq 0}} is of class- \mathcal{K}_{\infty} if it is continuous, strictly increasing, \alpha(0)=0 , and \text{lim}_{x\rightarrow\infty}\mkern 2.0mu\alpha(x)=\infty . \alpha such that for all {\bm{x}\in\mathcal{C}_{\rm S}} \sup_{\bm{u}\in\mathcal{U}}\dot{h}(\bm{x},\bm{u})=\nabla h(\bm{x})\big(\bm{f}(\bm{x})+\bm{g}(\bm{x})\bm{u}\big)>-\alpha(h(\bm{x})). Theorem 1 ( [ 1 ] ) . If h is a CBF for ( 1 ) on \mathcal{C}_{\rm S} , then any locally Lipschitz controller {\bm{k}:\mathcal{X}\rightarrow\mathcal{U}} , {\bm{u}=\bm{k}(\bm{x})} satisfying \displaystyle\nabla h(\bm{x})\big(\bm{f}(\bm{x})+\bm{g}(\bm{x})\bm{u}\big)\geq-\alpha(h(\bm{x})), (2) for all {\bm{x}\in\mathcal{C}_{\rm S}} renders the set \mathcal{C}_{\rm S} forward invariant. One can ensure the safety of ( 1 ) for a primary controller {\bm{k}_{\rm p}\!:\!\mathcal{X}\!\to\!\mkern 2.0mu\mathcal{U}} by solving the following quadratic program (QP): \displaystyle\bm{k}_{\rm safe}(\bm{x})=\underset{\bm{u}\in\mathcal{U}}{\text{arg\,min}}\mkern 9.0mu \displaystyle\left\|\bm{k}_{\rm p}(\bm{x})-\bm{u}\right\|^{2} (CBF-QP) s.t. \displaystyle\nabla h(\bm{x})\big(\bm{f}(\bm{x})+\bm{g}(\bm{x})\bm{u}\big)\geq-\alpha(h(\bm{x})). Ensuring the feasibility of the ( CBF-QP ) can be challenging, especially for high-dimensional systems with input bounds. This motivates a related tool known as backup CBFs. II-B Backup Control Barrier Functions Backup CBFs [ 14 , 15 ] construct controlled invariant sets online for feasibility guarantees with input bounds. A set {\mathcal{C}\subset\mathbb{R}^{n}} is controlled invariant if there exists a controller {\bm{k}:\mathcal{X}\rightarrow\mathcal{U}} , {\bm{u}=\bm{k}(\bm{x})} rendering \mathcal{C} forward invariant for ( 1 ). As in Section II-A , assume that safety is defined by a set \mathcal{C}_{\rm S} which is not necessarily controlled invariant. Now, suppose there exists a backup set and corresponding backup controller with the following