Robust Control Barrier Functions (R-CBF)

• Robust Control Barrier Functions (R-CBF) are safety-critical constructs that extend standard barrier certificates to ensure set invariance under disturbances, uncertainties, and model errors.
• They employ worst-case optimization, adaptive margins, and online estimation to modify nominal barrier conditions into robust constraints for control-affine systems.
• R-CBFs have been applied in motion planning, vehicle control, and robotics, balancing safety, feasibility, and conservatism in challenging uncertain environments.

Robust Control Barrier Functions (R-CBFs) are extensions of control barrier functions that preserve safety when the nominal assumptions behind standard barrier certificates are violated by disturbances, model mismatch, input uncertainty, state-estimation error, environmental uncertainty, or time variation in the safe set. In the nominal control-affine setting

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

with safe set $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$, a standard CBF or zeroing CBF enforces forward invariance through inequalities of the form

$L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$

R-CBF formulations retain this invariance logic but replace the nominal derivative constraint by a worst-case, margin-based, adaptive, or value-function-based condition so that safety is preserved exactly or for a disturbance-dependent relaxation or inflation of the safe set (Xu et al., 2016, Cohen et al., 2022, Nanayakkara et al., 24 Aug 2025, Knoedler et al., 2024).

1. Nominal barrier foundations and robust safety semantics

The modern R-CBF literature is anchored in the zeroing formulation of barrier functions. For autonomous dynamics $\dot{x}=f(x)$, a continuously differentiable $h$ is a zeroing barrier function if there exists a locally Lipschitz extended class-$\mathcal{K}$ function $\alpha$ such that

$L_f h(x)\ge -\alpha(h(x)).$

For control-affine systems, the controlled analogue is the zeroing control barrier function (ZCBF),

$$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$

with admissible safe inputs

$$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$

A locally Lipschitz feedback $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$0 renders $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$1 forward invariant (Xu et al., 2016).

A central robustness insight is that a zeroing barrier induces a Lyapunov-like function for the set $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$2. Under vanishing disturbances, $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$3 remains asymptotically stable; under bounded non-vanishing disturbances, the invariant object becomes a disturbance-dependent relaxed set $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$4, which is locally asymptotically stable rather than the original safe set itself (Xu et al., 2016). This practical-safety interpretation remains a recurring template in later R-CBF work: robustness is often expressed as exact invariance for sufficiently small uncertainty and invariance of an inflated or relaxed set otherwise (Nanayakkara et al., 24 Aug 2025).

The zeroing formulation is also important because it vanishes on $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$5, unlike reciprocal barrier functions that diverge at the boundary. The robustness paper of Ames and coauthors explicitly argues that this boundary behavior is advantageous under perturbations, since reciprocal barriers can require unbounded control when the disturbance makes the invariance condition infeasible (Xu et al., 2016). Later work nevertheless revisits reciprocal structure in modified form, most notably the reciprocal resistance-based construction, which inserts a term $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$6 into a zeroing-style inequality to create an internal buffer zone near the boundary without requiring explicit disturbance bounds in the controller design (Wang et al., 25 Jul 2025).

2. Major robust formulations and uncertainty models

R-CBF is not a single mathematical object but a family of robustifications matched to different uncertainty channels. The common structure is a nominal barrier derivative augmented by either a worst-case optimization, an explicit robustness margin, or an estimated uncertainty term.

Uncertainty source	Representative robust mechanism	Representative paper
Parametric uncertainty in $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$7	$\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$8	(Cohen et al., 2022)
State-estimation error	$\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$9	(Nanayakkara et al., 24 Aug 2025)
Sector-bounded input uncertainty	worst-case term proportional to $L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$0	(Buch et al., 2021)
Environment-state uncertainty	worst-case errors in $L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$1, $L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$2, and $L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$3	(Hamdipoor et al., 2023)
Unknown disturbance/model error	online estimate $L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$4 plus residual bound $L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$5	(Daş et al., 2023)

For parametric uncertainty in both drift and actuation,

$L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$6

the duality-based formulation defines a robust control barrier function by

$L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$7

Because the worst-case term is bilinear in $L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$8 and $L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0.$9, the paper reformulates the inner minimization over the polyhedral uncertainty set $\dot{x}=f(x)$0 using LP duality, yielding a QP with auxiliary dual variables rather than vertex enumeration (Cohen et al., 2022).

For uncertain state information, the robustification can be expressed directly in the barrier inequality. The state-uncertainty paper defines $\dot{x}=f(x)$1 as an R-CBF if there exists $\dot{x}=f(x)$2 and a robustness function $\dot{x}=f(x)$3 such that

$\dot{x}=f(x)$4

This yields two regimes. If the actuation disturbance induced by estimation error is small enough, the original safe set $\dot{x}=f(x)$5 is forward invariant and asymptotically stable. If the uncertainty is larger, an inflated set

$\dot{x}=f(x)$6

is forward invariant and asymptotically stable instead (Nanayakkara et al., 24 Aug 2025).

For sector-bounded uncertainties at the plant input, the uncertainty is modeled by an unknown memoryless or time-varying map $\dot{x}=f(x)$7 lying in a sector $\dot{x}=f(x)$8. After loop shifting, the robust barrier condition becomes a norm-bounded worst-case inequality, and the associated min-norm safety filter can be recast as a second-order cone program. This is a distinctive R-CBF setting because the uncertainty is multiplicative in the commanded control norm rather than an exogenous additive disturbance (Buch et al., 2021).

For uncertainty that enters through the environment rather than the plant dynamics, environmentally robust CBFs treat the environment state $\dot{x}=f(x)$9 as uncertain and explicitly bound the induced errors in the barrier value, gradient, and time derivative. The robust residual term takes the form

$h$0

which produces an SOCP-based safe controller and, in a scalar-input reformulation, a QP-based minimally modifying filter (Hamdipoor et al., 2023).

A separate line of work reduces conservatism by estimating the uncertainty online. For

$h$1

the estimator-based method constructs $h$2 together with explicit bounds on the estimation error $h$3, and then either compensates the matched uncertainty directly or inserts $h$4 and $h$5 into a robustified CBF or HOCBF constraint (Daş et al., 2023).

3. High-relative-degree, exponential, and multiple-constraint R-CBFs

Many safety constraints have relative degree greater than one, so the control does not appear in $h$6. This is the setting of exponential CBFs, high-order CBFs, CBF backstepping, and several specialized robust constructions.

In the motion-planning formulation for a bicycle model, the basic obstacle-avoidance barrier is

$h$7

Because $h$8 has relative degree two, the paper uses an exponential CBF recursion,

$h$9

and then robustifies it under additive and multiplicative model error terms $\mathcal{K}$0 and $\mathcal{K}$1. The resulting worst-case robust constraints are

$\mathcal{K}$2

which are evaluated during RRT expansion rather than by a separate optimization-heavy controller (Manjunath et al., 2020).

High-relative-degree robustness also motivates smooth approximations. For disturbed dynamics

$\mathcal{K}$3

the standard robust term $\mathcal{K}$4 is nonsmooth at $\mathcal{K}$5. The smooth RCBF paper replaces it by

$\mathcal{K}$6

yielding the smooth robust CBF condition

$\mathcal{K}$7

This smooth surrogate is then embedded into a recursive CBF backstepping construction for unicycle obstacle avoidance with unknown moving obstacles treated as disturbances (Kim et al., 2024).

A different high-relative-degree program appears in the satellite-trajectory literature. There, the barrier $\mathcal{K}$8 is not the original constraint $\mathcal{K}$9 but an inner-safe-set construction whose zero sublevel set is a subset of the states that can actually be kept safe under bounded input and disturbance authority. The paper develops three robust constructions: a constant-control-authority relative-degree-2 form,

$\alpha$0

a variable-authority relative-degree-2 form based on a decreasing $\alpha$1, and a predictive construction

$\alpha$2

for general relative degree $\alpha$3 (Breeden et al., 2021).

When several barrier constraints must hold simultaneously, mutual feasibility becomes a separate issue. For bounding-box constraints on a second-order system with bounded input and disturbance, the paper based on the theorem of Breeden and Panagou constructs two robust barriers $\alpha$4 and $\alpha$5 and shows that the common feasible input interval can be empty unless the gains are selected carefully. A sufficient condition for non-emptiness of the combined feasible set is

$\alpha$6

under the corresponding control-authority assumptions (Spiller et al., 24 Mar 2025).

4. Synthesis, optimization, and embedded planning

Nominal CBF synthesis is classically realized by a minimum-norm QP. For a CBF alone,

$\alpha$7

Under local Lipschitz assumptions and a uniform relative-degree-one condition $\alpha$8, the resulting feedback is locally Lipschitz. The same regularity extends to combined CLF-CBF QPs with a relaxed CLF constraint and a hard CBF constraint (Xu et al., 2016).

Robust synthesis preserves this optimization template but alters the constraint geometry. The duality-based parametric-uncertainty method produces a QP in $\alpha$9 whose constraints exactly encode the worst-case uncertainty over a polyhedral parameter set through dual variables, thereby avoiding a $L_f h(x)\ge -\alpha(h(x)).$0 vertex check (Cohen et al., 2022). Sector-bounded input uncertainty and the original environmentally robust formulation instead yield SOCPs because the robust barrier constraint contains norms of affine functions of the control (Buch et al., 2021, Hamdipoor et al., 2023).

Several papers retain a QP safety filter structure by moving the robustness burden outside the online decision variable. The estimator-based method inserts $L_f h(x)\ge -\alpha(h(x)).$1 and a residual bound into the barrier inequality (Daş et al., 2023). The state-uncertainty R-CBF constructs a minimally invasive controller

$L_f h(x)\ge -\alpha(h(x)).$2

subject to the stronger inequality

$L_f h(x)\ge -\alpha(h(x)).$3

and thereby requires no prior knowledge of the state-estimation error magnitude (Nanayakkara et al., 24 Aug 2025).

The online-adaptation formulation for safe navigation under state uncertainty modifies the robustification parameters $L_f h(x)\ge -\alpha(h(x)).$4 at run time rather than fixing them globally. Its robust CBF constraint is

$L_f h(x)\ge -\alpha(h(x)).$5

where $L_f h(x)\ge -\alpha(h(x)).$6 are selected online by derivative-free optimization using local samples around the current state estimate (Das et al., 26 Aug 2025).

Robust barrier constraints can also be embedded directly in planning rather than in feedback filtering. The robust RRT-KBF planner samples a kinodynamically feasible control, computes $L_f h(x)\ge -\alpha(h(x)).$7, and rejects the sampled tree extension unless the robust inequalities $L_f h(x)\ge -\alpha(h(x)).$8 and $L_f h(x)\ge -\alpha(h(x)).$9 hold. In this architecture, the R-CBF is a feasibility filter during RRT expansion, not an external controller (Manjunath et al., 2020).

5. Verification, learning, value functions, and discrete-time extensions

A major development after the original disturbance-robust CBF papers is the use of value functions, verification, and learning to construct robust certificates rather than hand-designing $$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$0.

The policy-value approach replaces an analytic barrier by the worst future constraint violation under a policy. In the robust case,

$$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$1

and the practical approximation is

$$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$2

This Robust Policy CBF (RPCBF) is used inside a standard safety-filter QP after estimating its gradient, with cubic splines improving the approximation of the continuous-time supremum and the associated gradient quality (Knoedler et al., 2024).

In discrete time, a robust DTCBF for

$$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$3

must ordinarily satisfy

$$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$4

The counterexample-guided synthesis paper replaces this infinite family of constraints by the Lipschitz-based sufficient condition

$$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$5

which is directly usable in an online least-deviation controller and can be verified over continuous regions by iterative domain splitting and counterexample generation (Shakhesi et al., 16 Jun 2025).

For control-affine polynomial systems with bounded additive uncertainty and polynomial input constraints, verification and synthesis can be posed as multilevel polynomial optimization problems. By using the KKT conditions of the inner control optimization and a closed-form reduction of the worst-case uncertainty, the verification problem becomes a single-level polynomial optimization problem, and Lasserre’s hierarchy yields asymptotically convergent semidefinite relaxations (Kang et al., 2023).

At the opposite end of the spectrum, robust discrete-time CBFs can be tied to the exact maximal robust safe set through dynamic programming. The adversarial-RL framework shows that the safety value function solving the Isaacs equation,

$$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$6

is a valid robust discrete-time CBF on the maximal robust safe set, and that a lifted state-action-disturbance quantity

$$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$7

supports a robust Q-CBF safety filter even for black-box nonlinear systems (Oh et al., 14 Apr 2026).

Terminological overlap is substantial in this area. The recurrent CBF framework is explicitly not a disturbance-robust R-CBF in the standard sense; it uses finite-time recurrence rather than invariance, and its robustness is a verification robustness with respect to sampling neighborhoods and Lipschitz trajectory tubes (Liu et al., 2 Oct 2025). Likewise, the composite soft-minimum construction uses “R-CBF” in the sense of a relaxed barrier inequality with an auxiliary relaxation variable $$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$8, rather than disturbance robustness (Rabiee et al., 2024).

6. Applications, empirical behavior, and recurring trade-offs

R-CBF methods have been evaluated in motion planning, vehicle control, aerial robotics, legged locomotion, and robotic manipulation. Across these domains, the recurring empirical pattern is a trade-off between safety margin, feasibility, and conservatism.

In robust RRT-KBF planning on the Cogniteam Hamster V7 robot car, the robust barrier formulation explicitly handled obstacle position error of $$\sup_{u\in U}\big[L_f h(x)+L_g h(x)u+\alpha(h(x))\big]\ge 0,$$9 and obstacle radius error of $$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$0. Over 100 runs, typical runtimes were approximately $$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$1–$$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$2 for plain RRT, $$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$3–$$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$4 for RRT-CBF QP, and $$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$5–$$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$6 for RRT-KBF; under larger robust uncertainty bounds, robust RRT-KBF increased to around $$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$7–$$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$8 while maintaining safe clearances that nominal KBF planning could lose (Manjunath et al., 2020).

RPCBF was tested on high-relative-degree, box-input-constrained systems including the double integrator, Segway, AutoRally, and a Crazyflie quadcopter. In AutoRally experiments integrated with Shield-MPPI, the reported collision/crash metrics were $$K_{\mathrm{cbf}}(x)=\{u\in U:\,L_f h(x)+L_g h(x)u+\alpha(h(x))\ge 0\}.$$9 for MPPI, $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$00 for SMPPI-HOCBF, $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$01 for SMPPI-PCBF, and $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$02 for SMPPI-RPCBF. On quadcopter hardware, RPCBF computed online with $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$03 and $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$04 disturbance samples maintained safety under model mismatch, while PCBF collided when it assumed the nominal model was exact (Knoedler et al., 2024).

The safe-navigation work under state uncertainty demonstrates the performance cost of fixed robust margins. On a tracked GVR-Bot with OpenVINS state estimates and covariance-derived uncertainty bounds, non-robust R-CBF-QP with $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$05 became unsafe and the robot tipped over, fixed-gain R-CBF-QP remained safe but conservative, and online-adapted R-CBF-QP remained safe while achieving the lowest reported $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$06 among safe controllers on hardware (Das et al., 26 Aug 2025).

The reachability-exact robust Q-CBF framework reports strong results in high dimension. On a disturbed inverted pendulum, the learned robust Q-CBF nearly recovered the maximal robust safe set and achieved $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$07 empirical safety under stress testing. On a 36-D Unitree Go2 quadruped in MuJoCo, the unfiltered task policy achieved a $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$08 safe rate over 50 trials, least-restrictive safety filtering achieved $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$09, and the neural robust Q-CBF achieved $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$10 (Oh et al., 14 Apr 2026).

Robust adaptive time-varying CBFs have also been deployed in robotic surface treatment, where time-varying force bounds enforce material-removal quality. In that setting, set-membership identification reduced conservatism by about $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$11 in simulation and $\mathcal{C}=\{x\in\mathbb{R}^n:h(x)\ge 0\}$12 in experiments, while the baseline TVCBF violated the force constraint under uncertainty and disturbance (Kim et al., 17 Jun 2025).

The principal misconceptions surrounding R-CBFs stem from conflating these distinct robustness mechanisms. An R-CBF may mean worst-case invariance under bounded disturbances, invariance under unknown but bounded state-estimation error, adaptive robustness to parametric uncertainty, rollout-based robustness to sampled disturbance trajectories, or merely a relaxed barrier inequality. The literature also makes clear that robustness is not free: larger disturbance bounds, larger smoothness surrogates, or poorly tuned robustification gains shrink feasible control sets, enlarge internal buffer zones, and can create infeasibility when multiple barriers must hold simultaneously (Spiller et al., 24 Mar 2025, Kim et al., 2024, Das et al., 26 Aug 2025). This suggests that R-CBF research is as much about feasibility management and conservatism reduction as it is about set invariance itself.