Recurrent Control Barrier Functions

• Recurrent Control Barrier Functions are safety certificates that require trajectories to return to a designated safe set within a finite time horizon.
• They are implemented via direct recurrent formulations or layered frameworks that leverage reduced-order models and nonparametric methods for efficient safety verification.
• The approach replaces pointwise invariance with finite-time recurrence, allowing transient departures while ensuring long-term safety under varying control conditions.

Recurrent Control Barrier Functions (RCBFs) are barrier-based safety certificates that replace the standard requirement of forward invariance by a recurrence condition on trajectories. In the recurrent formulations introduced in 2025, a trajectory may temporarily leave an auxiliary barrier superlevel set, provided it returns within a finite horizon $\tau$; safety is then recovered either by combining recurrence with reachable-set exclusion (Liu et al., 2 Oct 2025) or by coupling a reduced-order-model control barrier function with a recurrent tracking certificate for a full-order model (Liu et al., 1 Oct 2025). The term is distinct from earlier uses of the same acronym for reciprocal or robust control barrier functions.

1. Terminology, scope, and acronym ambiguity

In the control-barrier literature, the acronym RCBF is overloaded. The recurrent usage is recent and should be separated from the reciprocal and robust usages that already existed in the CBF literature.

Meaning of RCBF	Representative papers	Core safety idea
Recurrent Control Barrier Function	(Liu et al., 1 Oct 2025, Liu et al., 2 Oct 2025)	Finite-time return or $\tau$-recurrence
Reciprocal Control Barrier Function	(Ames et al., 2016, So et al., 2023)	Barrier blows up at the boundary
Robust Control Barrier Function	(Spiller et al., 24 Mar 2025, Breeden et al., 2021)	Invariance under disturbances and input constraints

The standard background is the continuous-time CBF framework for a safe set $\mathcal C=\{x:h(x)\ge 0\}$, where forward invariance is enforced through pointwise control inequalities such as

$L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$

or, in reciprocal form, through a singular barrier on $\operatorname{Int}(\mathcal C)$ (Ames et al., 2016). Recurrent CBFs depart from that logic by weakening the certificate itself: the set associated with the recurrent barrier need not be invariant, and the barrier quantity need not satisfy a pointwise differential inequality at every instant.

This distinction matters because a recurrent CBF is not simply a standard CBF enforced repeatedly online. The repeated online QP logic of the classical CLF-CBF framework is a precursor, but it is not itself the recurrent formalism introduced under the name “Recurrent Control Barrier Function” in 2025 (Ames et al., 2016).

2. Finite-horizon recurrence as a barrier condition

One recurrent formulation considers the control system

$\dot{x}=F(x,u),$

with trajectories $\phi(t,x,u)$, unsafe set $\mathcal X_u$, and a continuous function $h:\mathbb R^n\to\mathbb R$. A compact set $S$ is control recurrent if, for every $\tau$0, there exists a control such that the trajectory returns to $\tau$1 infinitely often; it is control $\tau$2-recurrent if every return occurs within at most $\tau$3 time units (Liu et al., 2 Oct 2025).

The recurrent barrier condition is defined on

$\tau$4

by requiring that for every $\tau$5 there exists $\tau$6 such that

$\tau$7

where $\tau$8. A frequently used choice is

$\tau$9

The induced safety theorem has two parts. First, if $\mathcal C=\{x:h(x)\ge 0\}$0 is an RCBF, then $\mathcal C=\{x:h(x)\ge 0\}$1 is control $\mathcal C=\{x:h(x)\ge 0\}$2-recurrent. Second, if

$\mathcal C=\{x:h(x)\ge 0\}$3

where $\mathcal C=\{x:h(x)\ge 0\}$4 is the $\mathcal C=\{x:h(x)\ge 0\}$5-backward reachable tube of the unsafe set, then any control rendering the trajectory $\mathcal C=\{x:h(x)\ge 0\}$6-recurrent also guarantees

$\mathcal C=\{x:h(x)\ge 0\}$7

for every $\mathcal C=\{x:h(x)\ge 0\}$8 (Liu et al., 2 Oct 2025).

The conceptual shift is precise. A standard CBF says “once inside, never leave.” A recurrent CBF says “if a trajectory leaves, it must come back within $\mathcal C=\{x:h(x)\ge 0\}$9.” Safety is then obtained not from invariance of $L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$0 itself, but from the conjunction of finite-time return and exclusion of the unsafe set’s $L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$1-reachable region. This suggests a different geometric object: a recurrent-safe set rather than an invariant safe set.

3. Layered recurrent barriers from recurrent tracking functions

A second recurrent formulation arises in layered safety-critical control for a high-dimensional full-order model (FoM)

$L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$2

and a lower-dimensional reduced-order model (RoM)

$L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$3

The models are coupled by a projection $L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$4 and a layered controller

$L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$5

with RoM safety set

$L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$6

and lifted FoM safety set

$L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$7

The motivation is computational: synthesizing ordinary CBFs directly on the FoM is often computationally intractable, whereas synthesizing them on the RoM is much easier (Liu et al., 1 Oct 2025).

The RoM side remains classical. The barrier $L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$8 satisfies the standard CBF condition

$L_f h(x)+L_g h(x)u \ge -\alpha(h(x)),$9

and the main derivation specializes to the linear choice $\operatorname{Int}(\mathcal C)$0, $\operatorname{Int}(\mathcal C)$1. The novelty is the tracking layer. If $\operatorname{Int}(\mathcal C)$2 is a safe RoM reference, the projected FoM trajectory is $\operatorname{Int}(\mathcal C)$3, and the tracking error is

$\operatorname{Int}(\mathcal C)$4

The paper replaces monotone Lyapunov tracking with a Recurrent Tracking Function (RTF) $\operatorname{Int}(\mathcal C)$5. On a compact set $\operatorname{Int}(\mathcal C)$6, $\operatorname{Int}(\mathcal C)$7 must satisfy linear error bounds

$\operatorname{Int}(\mathcal C)$8

and a $\operatorname{Int}(\mathcal C)$9-exponential $\dot{x}=F(x,u),$0-recurrence condition

$\dot{x}=F(x,u),$1

This is weaker than a differential inequality such as $\dot{x}=F(x,u),$2: transient growth of the tracking measure is allowed, provided recurrent decrease occurs within every time window of length $\dot{x}=F(x,u),$3. Under this condition, the tracking-error derivative still obeys the exponential bound

$\dot{x}=F(x,u),$4

The recurrent barrier is then constructed as

$\dot{x}=F(x,u),$5

with superlevel set

$\dot{x}=F(x,u),$6

If the RoM barrier is a CBF with $\dot{x}=F(x,u),$7, the RTF has rate $\dot{x}=F(x,u),$8, and the stated assumptions hold, then $\dot{x}=F(x,u),$9 is a control $\phi(t,x,u)$0-recurrent set and $\phi(t,x,u)$1 is an RCBF on $\phi(t,x,u)$2 (Liu et al., 1 Oct 2025).

The crucial point is that $\phi(t,x,u)$3 is only recurrent, but the FoM safety variable can still be forward invariant. Along the projected FoM trajectory,

$\phi(t,x,u)$4

Using the exponential tracking bound and the initial margin condition

$\phi(t,x,u)$5

the paper proves

$\phi(t,x,u)$6

provided $\phi(t,x,u)$7. The separation condition $\phi(t,x,u)$8 is therefore the key timescale requirement: the tracking correction must dominate the rate at which the RoM barrier budget can decay. The guaranteed safe set is not all of $\phi(t,x,u)$9, but the subset whose initial barrier value is large enough relative to the initial tracking error.

The same paper also gives a disturbance extension. If the tracking layer satisfies the ISS-type estimate

$\mathcal X_u$0

then the practical RTF condition is shifted by

$\mathcal X_u$1

and the robust recurrent barrier becomes

$\mathcal X_u$2

Safety is retained if the initial condition lies in the correspondingly shrunk set $\mathcal X_u$3 (Liu et al., 1 Oct 2025).

4. Signed-distance RCBFs and nonparametric safety verification

A major theoretical result of the direct recurrence-based formulation is that, under mild assumptions, the signed-distance function can itself serve as an RCBF. For a closed set $\mathcal X_u$4, the signed distance is

$\mathcal X_u$5

and the candidate barrier is

$\mathcal X_u$6

If a classical CBF $\mathcal X_u$7 satisfies the sector containment condition

$\mathcal X_u$8

over $\mathcal X_u$9, and if $h:\mathbb R^n\to\mathbb R$0 is closed with

$h:\mathbb R^n\to\mathbb R$1

then $h:\mathbb R^n\to\mathbb R$2 is an RCBF over $h:\mathbb R^n\to\mathbb R$3 for suitable $h:\mathbb R^n\to\mathbb R$4, $h:\mathbb R^n\to\mathbb R$5, and an explicit lower bound on $h:\mathbb R^n\to\mathbb R$6 (Liu et al., 2 Oct 2025).

This turns barrier construction into set identification. Instead of synthesizing a smooth certificate directly, one can identify a set $h:\mathbb R^n\to\mathbb R$7 and define the barrier as its signed distance. The paper uses this observation to propose a data-driven nonparametric method based on adaptive cell decomposition. The state space is partitioned into disjoint cells

$h:\mathbb R^n\to\mathbb R$8

with tentative safe cells $h:\mathbb R^n\to\mathbb R$9, unsafe cells $S$0, and candidate set

$S$1

The certification logic is trajectory-based. For two initial states $S$2 and $S$3 in the same radius-$S$4 ball and under the same control, the trajectory deviation implies

$S$5

This bound yields robust sufficient tests for whether an entire cell is outside or inside the $S$6-backward reachable tube of the unsafe set, and whether the RCBF recurrence condition holds or fails throughout the cell. Unresolved cells are split into smaller balls of radius $S$7 using the stencil

$S$8

Because each cell can be processed independently, the method is described as massively parallelizable (Liu et al., 2 Oct 2025).

The same paper reports a 3D evasion example with $S$9 s and $\tau$00 control samples per cell. At tested resolutions $\tau$01, HJ reachability captured $\tau$02, $\tau$03, $\tau$04, and $\tau$05 of the unsafe zone, respectively, while the recurrent-set method captured $\tau$06 at all tested resolutions. The reported computation times were $\tau$07, $\tau$08, $\tau$09, and $\tau$10 s for HJ, and $\tau$11, $\tau$12, $\tau$13, and $\tau$14 s for the recurrent-set method (Liu et al., 2 Oct 2025).

The layered RTF-based paper gives a different implementation profile. In a 2D double-integrator FoM with a first-order RoM, the RoM safe velocity is produced by the QP

$\tau$15

subject to

$\tau$16

and the FoM controller is

$\tau$17

Using $\tau$18, $\tau$19, and an exponentially convergent tracking error with $\tau$20 and $\tau$21 when the CBF constraint is inactive, the study compares $\tau$22, $\tau$23, and $\tau$24. Safety can fail when the RoM CBF is invalid, and it also cannot be guaranteed when the initial state lies outside $\tau$25; only when the RoM CBF is valid and the initial condition lies in $\tau$26 does the FoM remain safe for all time (Liu et al., 1 Oct 2025).

5. Relation to standard, reciprocal, robust, and learning-based barrier methods

The foundational difference between recurrent and classical CBFs is structural. Classical zeroing and reciprocal CBFs certify forward invariance of a set or of its interior through pointwise inequalities on the infinitesimal dynamics (Ames et al., 2016). Recurrent CBFs replace that with a finite-horizon recurrence inequality. In the direct recurrent formulation, the set $\tau$27 need not be invariant; in the layered formulation, the auxiliary superlevel set $\tau$28 need not be forward invariant either, even though it is used to prove forward invariance of the original FoM safety set (Liu et al., 1 Oct 2025).

The distinction from reciprocal CBFs is especially important because Ames, Xu, Grizzle, and Tabuada use “RCBF” to mean Reciprocal Control Barrier Function, and later stochastic work uses the same acronym in the reciprocal sense (Ames et al., 2016). The distinction from robust CBFs is equally important: papers on multiple robust CBFs for bounding-box constraints and on high-relative-degree satellite safety also use “RCBF” to mean Robust Control Barrier Function, with the central issue being feasibility under disturbances and input constraints rather than recurrence (Spiller et al., 24 Mar 2025, Breeden et al., 2021).

Several adjacent learning-based works are relevant but not terminologically identical. One paper learns a feasibility boundary for CBF/HOCBF QPs and improves it with a recurrent training algorithm, but it does not define a recurrent control barrier function as a new barrier object (Xiao et al., 2023). Another learns discrete-time CPA barrier functions from one-step data via the recursive condition

$\tau$29

which is a one-step invariance recurrence rather than the continuous-time $\tau$30-return notion used by recurrent CBFs (Strong et al., 25 Nov 2025). A plausible implication is that the recurrent CBF idea is part of a broader shift from purely differential barrier certificates toward trajectory-level and data-driven safety conditions, but the formal definitions are not interchangeable.

6. Assumptions, conservatism, and unresolved issues

The recurrent formulations obtain their flexibility by trading pointwise invariance conditions for stronger auxiliary assumptions elsewhere. In the layered RTF construction, one must verify forward completeness and local Lipschitz continuity, the relative-degree compatibility relation

$\tau$31

the bounded barrier-gradient condition

$\tau$32

a valid RoM CBF, a valid RTF with constants $\tau$33, and the strict separation condition $\tau$34. Safety is guaranteed only for the margin-restricted subset

$\tau$35

not for all $\tau$36. Under disturbances, the initial margin must be enlarged further to $\tau$37 (Liu et al., 1 Oct 2025).

In the signed-distance and nonparametric formulation, the safety guarantee depends on excluding the $\tau$38-backward reachable tube of the unsafe set and on verifying recurrence over cells using a Lipschitz bound $\tau$39. The paper explicitly notes that quantifying the approximation gap between the computed and true safe set remains future work, and it does not provide a formal sampling-complexity or error law. Sparse sampling or overly conservative Lipschitz bounds enlarge the robust margins $\tau$40, which can make certification conservative. The method still faces refinement growth in high dimensions, although more gracefully and more parallelizably than global PDE or SOS approaches (Liu et al., 2 Oct 2025).

The choice of $\tau$41 is itself a conservatism–cost parameter. In the nonparametric recurrent-set method, as $\tau$42, recurrent sets approach invariant sets; smaller $\tau$43 reduces the volume gap relative to the HJ/BRT benchmark, but computation time increases sharply, approximately exponentially (Liu et al., 2 Oct 2025). This suggests that recurrent CBFs are not a replacement for invariant-set methods in all regimes. Rather, they supply a different safety geometry—finite-time return, not perpetual stay-inside—that can be advantageous when invariant barrier synthesis is computationally prohibitive or when transient departures of auxiliary certificates are acceptable.

In that sense, recurrent control barrier functions define a distinct safety paradigm within barrier-function theory. They preserve the barrier viewpoint, but shift the certifying mechanism from instantaneous inward-pointing inequalities to trajectory recurrence over finite horizons.