Robust Safety Controllers (R-SC)

• Robust Safety Controllers are strategies that guarantee safety in dynamical systems by ensuring forward invariance of designated safe sets under uncertainties.
• They integrate control barrier functions, optimization methods, and observer-based techniques to manage disturbances, modeling errors, and parameter variations.
• These controllers are applied in high-stakes environments like autonomous vehicles and multi-agent systems, providing real-time safety and robust performance.

A robust safety controller (R-SC) enforces safety invariance properties for dynamical systems subject to exogenous uncertainties, disturbances, modeling errors, partial observability, or parameter variation. In advanced frameworks, R-SC synthesis leverages structured system knowledge, data-driven insights, barrier function techniques, and control-theoretic optimization to guarantee forward invariance of pre-specified safe sets. This article surveys the mathematical foundations, synthesis methodologies, and contemporary implementations of robust safety controllers, focusing on rigorous guarantees and algorithmic design under uncertainty.

1. Robust Safety Fundamentals and System Modeling

Robust safety controllers operate on control-affine or polynomial dynamical systems described generically as

$\dot{x} = f(x) + g(x)u + d(x,t),$

where $d(x,t)$ captures exogenous model perturbations, bounded disturbances, structured parametric uncertainty, multi-modal or distributionally ambiguous error, or unknown dynamics reconstructed by data. Safety is specified as forward invariance of a safe set $S$, with $S = \{x \in \mathbb R^n : h(x) \ge 0\}$, where $h$ is a smooth control barrier function (CBF), or, for more complex contexts, a composite robust-adaptive or delay-aware functional (Nanayakkara et al., 24 Aug 2025, Akbarzadeh et al., 28 Jan 2026, Liu et al., 2023).

Uncertainties handled in R-SC frameworks may include:

• Additive or multiplicative disturbances with known or unknown bounds (Dong et al., 18 Mar 2025, Zhao et al., 2023, Nguyen et al., 2020).
• State-dependent bounded sets $\Sigma_f(x), \Sigma_g(x)$ for drift/control channels (Wei et al., 2022, Wei et al., 2023).
• Gaussian or multi-modal uncertainties, including random latent modes (Wei et al., 2023).
• Distributional ambiguity captured via Wasserstein balls over unknown probability laws (Mestres et al., 2023).
• Unknown physical parameters, estimated adaptively (Liu et al., 2023).
• Partially observed or estimated states, incurring set-valued error (Agrawal et al., 2022, Nanayakkara et al., 24 Aug 2025).

System representations extend to discrete-time, input-affine polynomial, and hybrid systems for modern safety-critical applications (Akbarzadeh et al., 2024, Akbarzadeh et al., 28 Jan 2026, Shmarov et al., 2017).

2. Control Barrier Function Theory and Robust Extensions

The core tool for R-SC is the robust control barrier function (R-CBF), a function $h(x)$ such that the set $S = \{x : h(x) \ge 0\}$ is forward invariant for all admissible uncertainties and disturbances. For control-affine systems, robust CBF constraints can be formulated as

$$\inf_{\Delta \in \mathcal D} [L_f h(x) + L_g h(x) u + \nabla h(x) \cdot \Delta] \ge -\alpha(h(x)),$$

where $\Delta$ denotes the uncertainty and $\alpha$ is a class-$\mathcal K$ function. For data-driven and polynomial settings, sum-of-squares (SOS) certificates enable robust CBF feasibility directly from input-state trajectories (Akbarzadeh et al., 2024, Ashoori et al., 2 Aug 2025, Akbarzadeh et al., 28 Jan 2026).

Key robustification concepts:

• Uncertainty-margined CBF: Strengthening the CBF constraint by absorbing the worst-case error, as in $$L_f h(x) + L_g h(x)u + \alpha(h(x)) \ge \rho(\|L_g h(x)\|)$$, with $\rho$ a robustness function independent of the uncertainty bound (Nanayakkara et al., 24 Aug 2025).
• Volume CBF (VCBF): Ensuring persistent feasibility of the safety QP under multiple constraints, by enforcing invariance of the control set's volume via a CBF acting on feasible set measures (Dong et al., 18 Mar 2025).
• Adaptive and learning-based CBFs: Including parameter adaptation within the barrier condition, creating robust-adaptive CBFs that guarantee safety for all unknown parameters in a set (Liu et al., 2023).

3. Synthesis Procedures: Optimization and Observer Integration

Robust safety controllers are synthesized through convex optimization programs (often quadratic or second-order cone programs) that enforce CBF and system constraints robustly. The general R-SC synthesis workflow includes:

• Safety index/safety filter design: Parameterize a safety index $\phi(x)$ and enforce, via a QP or convex semi-infinite program, the robust constraint $\dot\phi(x,u) \le -\gamma(\phi(x))$, evaluated for all uncertainties (Wei et al., 2022, Wei et al., 2023).
• Disturbance/uncertainty observer integration: Employ nonlinear observers (e.g., RISE-based, high-gain, or backup-flow-based) to estimate disturbances, injecting observer error bounds into safety constraints to reduce conservatism and ensure robust invariance (Daş et al., 2022, Dong et al., 18 Mar 2025, Wijk et al., 19 Mar 2025).
• Model-free data-driven synthesis: Replace unavailable model terms with polynomial or function approximators informed by persistently exciting data, enforcing robust CBF decrease conditions via SOS or convex optimization, under rank/coverage guarantees (Akbarzadeh et al., 2024, Ashoori et al., 2 Aug 2025, Akbarzadeh et al., 28 Jan 2026).
• Control barrier/backup composition: For systems with strict input or state constraints, guarantee forward invariance via compositional and decentralized set-theoretic synthesis, leveraging robust controlled invariant (RCI) sets and compositional set intersections (Liu et al., 2021, Kaynama et al., 2013).
• Preference-based learning: Jointly optimize safety filters and performance by embedding robust CBF QPs within a human-in-the-loop learning paradigm, adjusting tuning parameters based on preference labels while maintaining barrier robustness (Cosner et al., 2021).

Offline computational methods include ellipsoidal reachability and SOS polynomial optimization for high-dimensional linear or polynomial systems, while online synthesis exploits efficient QP or SOCP solvers enhanced with real-time feasibility screening (Kaynama et al., 2013, Mestres et al., 2023).

4. Guarantees and Theoretical Properties

Formal guarantees for robust safety controllers—under broad model classes and robustified constraints—are well established:

• Forward invariance / safety: For any initial condition in the robust safe set, the R-SC enforces $x(t)\in S$ for all $t$, for all admissible uncertainties/disturbances (Nanayakkara et al., 24 Aug 2025, Agrawal et al., 2022, Liu et al., 2023, Akbarzadeh et al., 2024).
• Convergence/boundedness: If the system uncertainty/disturbance level remains below design thresholds, the original intended safe set is invariant; for larger uncertainties, an inflated superlevel set is invariant (Nanayakkara et al., 24 Aug 2025, Dong et al., 18 Mar 2025).
• Recursive feasibility/persistent realizability: Provided robust controlled invariance and properly enforced volume/measure constraints, safety QPs or CSIPs never lose feasibility along system trajectories, avoiding deadlocks or unnecessary conservatism even under tight input bounds or multi-modal uncertainties (Wei et al., 2022, Wei et al., 2023, Dong et al., 18 Mar 2025).
• Regularity of law: R-SC mappings $x\mapsto u^*(x)$ are point-Lipschitz under strict feasibility of the underlying optimization, ensuring closed-loop existence and avoiding chattering or controller discontinuities (Mestres et al., 2023).
• Performance tradeoff: Integrated learning allows user-preferred selection of CBF robustness parameters without violating rigorous input-to-state or set-invariance margins (Cosner et al., 2021).

5. Implementation Aspects and Case Studies

Robust safety controllers have been validated in a diversity of platforms and scenarios:

• Multi-agent and platooning systems: Decentralized synthesis of RCI sets and compositional controllers for interconnected vehicle systems with polytopic coordination constraints and disturbances (Liu et al., 2021).
• Robust safe control under measurement/process noise: RISE and high-gain observer-based robustness for obstacle-avoidance with quadrotors and blimps, demonstrating smallest-conservativeness among compared methods (Dong et al., 18 Mar 2025, Agrawal et al., 2022).
• Data-driven and model-free settings: SOS-based design of R-CBCs in input-affine polynomial systems of 3–7 states using only finite data trajectories, with rigorous horizon safety guarantees (Akbarzadeh et al., 2024, Ashoori et al., 2 Aug 2025, Akbarzadeh et al., 28 Jan 2026).
• Adaptive, parameter-uncertain systems: Polynomial raCBF synthesis for systems with unknown constant parameters, showing up to 55% performance improvement over worst-case robust CBFs while maintaining 100% safety across MC trials (Liu et al., 2023).
• MPC and backup architectures: Tube-MPC R-SC for autonomous vehicles, ensuring backup control and recursive feasibility under bounded disturbances across 100+ randomized obstacle scenarios (Nezami et al., 2022).
• High-dimensional LTI flight envelope protection: Hybrid automaton R-SC guaranteeing 12-D quadrotor safety over long horizons subject to actuator saturation and environmental disturbances (Kaynama et al., 2013).

Quantitative metrics include safety margin/loss, feasibility violation rates, controller solve-times, and performance efficiency relative to baseline or non-robust CBF/CLF approaches.

6. Extensions: Multi-modal, Distributional, and Delay-aware Robustness

Advanced robust safety controller frameworks address:

• Multi-modal and stochastic uncertainty: Explicitly modeling latent modes and multi-modal process/actuation distributions, enabling least-conservative chance-constrained safety indices, and persistent realizability using empirical Bayesian posterior guarantees (Wei et al., 2023).
• Distributional robustness: Wasserstein-ambiguous DRO-SOCP formulations for R-SC, providing (CVaR) risk constraints and fast on-line feasibility certification via Schur complement and eigenvalue screening (Mestres et al., 2023).
• Delayed dynamics: Krasovskii-type CBCs aggregating delayed state histories, solvable directly from input-state data, with infinite-horizon invariance for time-invariant delayed, uncertain, and unknown polynomial systems (Akbarzadeh et al., 28 Jan 2026, Zhao et al., 2023).
• Backup strategies: DO-bCBF frameworks that use observer-based backup trajectory flows, guaranteeing robustness under input constraints and time-varying disturbances (Wijk et al., 19 Mar 2025).

A summary table of core R-SC synthesis principles:

Principle	Key Guarantee	Representative Approach/Paper
Robust CBF margin	Forward invariance	(Nanayakkara et al., 24 Aug 2025, Nguyen et al., 2020)
Observer-integrated disturbance rejection	Real-time safety, less conservatism	(Dong et al., 18 Mar 2025, Daş et al., 2022)
Volume/feasible-set constraint	Persistent feasibility	(Dong et al., 18 Mar 2025, Wei et al., 2022)
Data-driven SOS synthesis	Model-free robustness	(Akbarzadeh et al., 2024, Ashoori et al., 2 Aug 2025)
Distributional/conic program robustness	Sampled ambiguity, optimality	(Mestres et al., 2023, Wei et al., 2023)
Delay/Krasovskii aggregation	Infinite-horizon safety	(Akbarzadeh et al., 28 Jan 2026, Zhao et al., 2023)

7. Outlook and Open Challenges

Current robust safety controller methods exhibit rigorous guarantees and validated real-time performance on challenging nonlinear, high-dimensional, and partially observable systems. However, ongoing challenges and research frontiers include:

• Scalability and sparsity for SOS-based design in high-dimensional systems (Akbarzadeh et al., 28 Jan 2026);
• Real-time adaptive update, online learning, and robustification for nonstationary, nonpolynomial, or adversarially non-Gaussian uncertainties (Liu et al., 2023, Wei et al., 2023);
• Transient safety under distributed, possibly unknown delays or networked communication constraints;
• Joint design of safety, performance, stabilizing, and preference-based control for multi-objective, human-in-the-loop systems (Cosner et al., 2021);
• Seamless integration of data-driven synthesis, model validation, and on-policy adaptation for black-box or legacy systems (Akbarzadeh et al., 2024, Ashoori et al., 2 Aug 2025).

The robust safety controller paradigm thus unifies barrier-functional, observer-based, data-driven, and learning-theoretic safety assurance in a mathematically rigorous and computationally tractable framework suitable for present and future safety-critical systems.