ParallelCBF: Scalable CBF Architectures

• ParallelCBF is a family of control barrier function techniques that decompose safety certification into independent, tractable subproblems using parallel strategies.
• Variants include neural CBF verification with linear bound propagation, simultaneous enforcement of multiple barrier constraints, and parallel finite-horizon rollouts for safe control.
• Empirical results demonstrate significant speed-ups and scalability in systems like Cart-Pole, quadrotors, and robotics, ensuring practical safety guarantees under diverse conditions.

ParallelCBF is a label used in recent arXiv literature for several distinct, though related, control-barrier-function-centered constructions. Across these works, the common motif is not a single standardized algorithm but the use of parallelism to make safety filtering, verification, or deployment tractable: parallel verification over state-space partitions, parallel enforcement of multiple barrier constraints, parallel rollout evaluation over fallback-policy libraries, and tensor-parallel training pipelines with hard-gate safety filters. In that sense, ParallelCBF denotes a family of CBF methodologies rather than a unique formalism (Vertovec et al., 9 Nov 2025).

1. Terminological scope and recurring design pattern

The label appears in at least six technically different senses in the cited literature. The following summary organizes those senses without collapsing their distinct assumptions or guarantees.

Usage of “ParallelCBF”	Core mechanism	Representative source
Neural CBF verification	Linear bound propagation, McCormick relaxation, simplicial refinement in parallel	(Vertovec et al., 9 Nov 2025)
Safety between parallel boundaries	Separate constant-sum CBFs for each boundary, extended to higher relative degree by backstepping	(Kim et al., 21 May 2025)
Multiple output box constraints	Multiple ECBFs with vector relative degree and a closed-form controller	(Cohen et al., 4 Sep 2025)
Multiple obstacle constraints	Multiple high-order CBFs stacked into one QP	(Aali et al., 2022)
Runtime fallback-policy safety filter	Parallel finite-horizon rollouts over a policy library, then one QP	(Kim et al., 15 May 2026)
Tensor-parallel RL framework	Vectorized environments, hard-gate CBF filters, sharded BC-to-RL, auditability APIs	(Lu et al., 15 May 2026)

A common misconception is to treat ParallelCBF as a single canonical method. The literature does not support that interpretation. Instead, the term is used for different constructions that share a CBF substrate and a parallelization strategy. This suggests that the unifying concept is architectural: each variant decomposes safety certification into independently evaluable subproblems and then recombines them through affine constraints, rollout selection, or batched execution.

2. ParallelCBF as scalable verification of neural control barrier functions

In "Scalable Verification of Neural Control Barrier Functions Using Linear Bound Propagation" (Vertovec et al., 9 Nov 2025), ParallelCBF is a verification framework for neural-based control barrier functions in control-affine systems

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

with $x\in\mathcal X\subset\mathbb R^n$, $$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$, and continuously differentiable $f,g$. A candidate CBF is a differentiable neural network $h(x)$ whose $0$-superlevel set

$C=\{x\mid h(x)\ge0\}$

must satisfy $C\subseteq S$ and, for all $x\in C$,

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

The framework replaces expensive non-linear SMT/MIP reasoning with sufficient linear conditions. For network values, standard LBP yields affine upper and lower bounds over a convex region $x\in\mathcal X\subset\mathbb R^n$0: $x\in\mathcal X\subset\mathbb R^n$1 For gradients, the method extends LBP to the Jacobian by writing $x\in\mathcal X\subset\mathbb R^n$2 as a product of layer Jacobians, relaxing each $x\in\mathcal X\subset\mathbb R^n$3 via affine bounds on $x\in\mathcal X\subset\mathbb R^n$4, and using McCormick relaxation for the bilinear terms in each two-layer product. This yields region-wise linear bounds

$x\in\mathcal X\subset\mathbb R^n$5

Once affine bounds on $x\in\mathcal X\subset\mathbb R^n$6 and on $x\in\mathcal X\subset\mathbb R^n$7 are available, the control term $x\in\mathcal X\subset\mathbb R^n$8 is again relaxed with McCormick envelopes, producing affine lower and upper bounds of the form

$x\in\mathcal X\subset\mathbb R^n$9

and similarly for an upper bound. The dynamics $$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$0 are bounded on each region by first-order certified Taylor expansions around the simplex center, with remainder enclosure via Lipschitz/Hessian bounds or Bernstein-coefficient bounds: $$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$1

These ingredients are combined into an affine lower bound on the CBF condition,

$$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$2

and encoded as the region-wise certification formula

$$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$3

A dual formula $$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$4 is constructed by upper-bounding the Lie terms, so any satisfying $$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$5 is a counterexample.

The refinement strategy partitions the global domain into a simplicial mesh $$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$6. For each simplex, $$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$7 and $$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$8 are evaluated independently. Certified simplices are accepted; simplices with a counterexample terminate verification; inconclusive simplices are split along their longest edge into two smaller simplices. Crucially, each region’s bound propagation and SAT checks are independent, so groups of simplices can be batched on the GPU, and no inter-region synchronization is required beyond managing the work queue. The reported complexity is $$u\in\mathcal U=[\underline u,\bar u]\subset\mathbb R^m$$9, with per-region bound propagation scaling $f,g$0 per layer plus $f,g$1 for McCormick (Vertovec et al., 9 Nov 2025).

The reported benchmarks are Barrier 2 and 3 from Jiang et al. with $f,g$2 networks, Darboux with $f,g$3, 2D-Control with $f,g$4, and Cart-Pole with $f,g$5. Regions examined ranged from $f,g$6 up to $f,g$7 for Cart-Pole. Verification times were $f,g$8 s for Barrier 2 versus dReal/SAT $f,g$9 s, $h(x)$0 s for Barrier 3 versus $h(x)$1 s, $h(x)$2 s for Barrier 4, $h(x)$3 s for Darboux with SMT timing out at $h(x)$4 h, $h(x)$5 s for 2D-Control, and $h(x)$6 s for Cart-Pole. In all cases the method achieved $h(x)$7 certification of the safe region (Vertovec et al., 9 Nov 2025).

3. Parallel CBF constructions for multiple or symmetric state constraints

A second line of work uses “parallel” to mean the simultaneous enforcement of several barrier inequalities rather than parallel hardware execution. In "Constant-Sum High-Order Barrier Functions for Safety Between Parallel Boundaries" (Kim et al., 21 May 2025), the safe set is defined by two smooth “parallel” surfaces $h(x)$8 and $h(x)$9,

$0$0

with the constant-sum condition

$0$1

The paper identifies why a single CBF for both boundaries can fail: for

$0$2

the gradient vanishes at the mid-plane $0$3, so $0$4 there and no control can enforce the barrier condition.

For the relative-degree-one case, the separate CBF conditions

$0$5

reduce, using $0$6, to a two-sided inequality

$0$7

where

$0$8

The associated QP is always feasible, and Theorem 3.2 gives a closed-form solution. For higher relative degree, a backstepping chain

$0$9

transforms the problem to a final relative-degree-one pair $C=\{x\mid h(x)\ge0\}$0. Theorem 4.1 states that, under the stated differentiability and initialization assumptions, applying the same relative-degree-one filter to $C=\{x\mid h(x)\ge0\}$1 guarantees $C=\{x\mid h(x)\ge0\}$2 and $C=\{x\mid h(x)\ge0\}$3 for all $C=\{x\mid h(x)\ge0\}$4 (Kim et al., 21 May 2025).

In "Compatibility of Multiple Control Barrier Functions for Constrained Nonlinear Systems" (Cohen et al., 4 Sep 2025), “Parallel CBF” denotes a closed-form controller for box constraints on vector-valued outputs $C=\{x\mid h(x)\ge0\}$5. For each coordinate, one defines

$C=\{x\mid h(x)\ge0\}$6

giving $C=\{x\mid h(x)\ge0\}$7 ECBFs. Vector relative degree guarantees that the affine safety constraints are compatible. With

$C=\{x\mid h(x)\ge0\}$8

the QP decouples and admits the explicit solution

$C=\{x\mid h(x)\ge0\}$9

where $C\subseteq S$0 and $C\subseteq S$1. The resulting controller is unique and locally Lipschitz in $C\subseteq S$2, and the paper additionally characterizes degradation of nominal tracking objectives via an ISS-type bound on the tracking error with respect to the CBF-intervention magnitude (Cohen et al., 4 Sep 2025).

A related earlier usage appears in "Multiple Control Barrier Functions: An Application to Reactive Obstacle Avoidance for a Multi-steering Tractor-trailer System" (Aali et al., 2022). There, multiple high-order CBFs are stacked into a single QP,

$C\subseteq S$3

with one affine constraint per barrier. The application uses one obstacle barrier per body per obstacle for a tractor-trailer, with tractor barriers of relative degree $C\subseteq S$4 and trailer barriers of relative degree $C\subseteq S$5. At each control tick, the method measures the state, solves an LTV-MPC to obtain $C\subseteq S$6, builds $C\subseteq S$7, and then solves

$C\subseteq S$8

subject to all CBF constraints and actuator limits. The paper reports QP solve times below $C\subseteq S$9 s with up to eight obstacles on a standard laptop, with all $x\in C$0 maintained at all times (Aali et al., 2022).

Taken together, these works show that one important meaning of ParallelCBF is structural rather than computational: multiple safety conditions are enforced in parallel, either because the safe set itself has a product or strip structure, or because independent barrier inequalities admit a compatible or stacked representation.

4. Policy-library CBF and parallel finite-horizon rollouts

In "Policy Library CBF: Finite-Horizon Safety at Runtime via Parallel Rollouts" (Kim et al., 15 May 2026), ParallelCBF appears as PL-CBF, a runtime safety filter built around a finite library of fallback policies

$x\in C$1

For a horizon $x\in C$2, each policy induces a finite-horizon safety margin

$x\in C$3

If $x\in C$4, the policy remains in the safe set $x\in C$5 over $x\in C$6. The framework evaluates all $x\in C$7 fallback policies in parallel by simulating

$x\in C$8

for each $x\in C$9, recording the barrier values, and collecting

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

These $$\sup_{u\in\mathcal U}\bigl[L_f h(x) + L_g h(x)\,u + \alpha\bigl(h(x)\bigr)\bigr] \ge 0.$$1 rollouts are completely independent and run in $$\sup_{u\in\mathcal U}\bigl[L_f h(x) + L_g h(x)\,u + \alpha\bigl(h(x)\bigr)\bigr] \ge 0.$$2 time on CPU or in one batched GPU kernel.

For each safe policy, the instantaneous admissible control set is

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

The method selects

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

and then solves the final QP

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

Theorem 1 states that if there exists an admissible policy $$\sup_{u\in\mathcal U}\bigl[L_f h(x) + L_g h(x)\,u + \alpha\bigl(h(x)\bigr)\bigr] \ge 0.$$6 with clearance

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

and the library precision satisfies $$\sup_{u\in\mathcal U}\bigl[L_f h(x) + L_g h(x)\,u + \alpha\bigl(h(x)\bigr)\bigr] \ge 0.$$8, then PL-CBF will find a library member whose rollout stays in the safe set over $$\sup_{u\in\mathcal U}\bigl[L_f h(x) + L_g h(x)\,u + \alpha\bigl(h(x)\bigr)\bigr] \ge 0.$$9. The guarantee is explicitly finite-horizon; repeated replanning is required at perception updates (Kim et al., 15 May 2026).

The reported runtimes are millisecond-scale. On a MacBook Air with Apple M4, the planar double-integrator with $x\in\mathcal X\subset\mathbb R^n$00 policies required approximately $x\in\mathcal X\subset\mathbb R^n$01–$x\in\mathcal X\subset\mathbb R^n$02 ms per step, highway driving with $x\in\mathcal X\subset\mathbb R^n$03 required approximately $x\in\mathcal X\subset\mathbb R^n$04 ms, and the $x\in\mathcal X\subset\mathbb R^n$05-state quadrotor with $x\in\mathcal X\subset\mathbb R^n$06 evasive policies, $x\in\mathcal X\subset\mathbb R^n$07, achieved $x\in\mathcal X\subset\mathbb R^n$08 failures in $x\in\mathcal X\subset\mathbb R^n$09 trials at $x\in\mathcal X\subset\mathbb R^n$10 ms per step. In the same quadrotor study, single-policy baselines failed $x\in\mathcal X\subset\mathbb R^n$11–$x\in\mathcal X\subset\mathbb R^n$12 at $x\in\mathcal X\subset\mathbb R^n$13–$x\in\mathcal X\subset\mathbb R^n$14 ms per step (Kim et al., 15 May 2026).

This usage is conceptually distinct from multiple-constraint Parallel CBF constructions. Here “parallel” refers to parallel policy evaluation over a discrete fallback set, and the barrier object is a finite-horizon surrogate $x\in\mathcal X\subset\mathbb R^n$15 induced by closed-loop rollouts rather than a fixed analytic $x\in\mathcal X\subset\mathbb R^n$16 alone.

5. parallelcbf as a tensor-parallel reinforcement-learning framework

In "parallelcbf: A composable safety-filter and auditability framework for tensor-parallel reinforcement learning" (Lu et al., 15 May 2026), ParallelCBF is a software framework rather than a single control law. ParallelCBF v0.1.0 is released under Apache 2.0 and organized into four abstract-base-class layers: environments, safety filters, algorithms, and operational auditability. The environment layer centers on SafeEnv, which extends Gymnasium.Env with safety_state(), safety_metrics(), and hard_constraint_violations(). The safety layer defines a SafetyFilter ABC and a DualBarrierCBF that uses two hard-gate CBFs around each obstacle, solved via a closed-form QP per environment. The algorithm layer follows a Stable-Baselines3-style interface, and the ops layer includes PreRegistration, DefaultWatchdogRegistry, FailureForensics, AtomicCheckpoint, and DatasetAudit (Lu et al., 15 May 2026).

The hard-gate dual-barrier formulation is defined for relative position $x\in\mathcal X\subset\mathbb R^n$17, velocity $x\in\mathcal X\subset\mathbb R^n$18, effective obstacle radius $x\in\mathcal X\subset\mathbb R^n$19, and predictive margin $x\in\mathcal X\subset\mathbb R^n$20 by

$x\in\mathcal X\subset\mathbb R^n$21

The enforced safety constraints per environment are

$x\in\mathcal X\subset\mathbb R^n$22

Nominal action $x\in\mathcal X\subset\mathbb R^n$23 is projected by solving

$x\in\mathcal X\subset\mathbb R^n$24

subject to the two barrier inequalities with slack variables $x\in\mathcal X\subset\mathbb R^n$25, which hard-gate filters set to zero (Lu et al., 15 May 2026).

The framework emphasizes auditability as a first-class systems property. PreRegistration commits convergence criteria, curriculum, and reward specifications to a SHA-256 hash. WatchdogRegistry derives threshold checks from the committed specification and raises a halt event when triggered. FailureForensics stores a rolling buffer of recent metrics, gradient norms, and activation statistics for JSON dumping on halt. AtomicCheckpoint implements a crash-safe “.tmp $x\in\mathcal X\subset\mathbb R^n$26 fsync $x\in\mathcal X\subset\mathbb R^n$27 rename” protocol, and DatasetAudit provides Pydantic-validated audits of scene-type distribution, action ranges, and BPTT integrity (Lu et al., 15 May 2026).

The reported reproducibility and dataset statistics are unusually specific. The framework includes $x\in\mathcal X\subset\mathbb R^n$28 pytest cases—vectorized, Hypothesis-driven, and parameterized—and all $x\in\mathcal X\subset\mathbb R^n$29 tests complete in $x\in\mathcal X\subset\mathbb R^n$30 s on a single GitHub Actions CPU core. It also includes a $x\in\mathcal X\subset\mathbb R^n$31-episode behavior-cloning data-collection campaign. The curriculum mix is

$x\in\mathcal X\subset\mathbb R^n$32

with expected aggregate yield $x\in\mathcal X\subset\mathbb R^n$33 from a $x\in\mathcal X\subset\mathbb R^n$34-episode dry run and observed $x\in\mathcal X\subset\mathbb R^n$35 on $x\in\mathcal X\subset\mathbb R^n$36 attempts. The per-bucket yields are $x\in\mathcal X\subset\mathbb R^n$37 for open-space, $x\in\mathcal X\subset\mathbb R^n$38 for single static, $x\in\mathcal X\subset\mathbb R^n$39 for multi-obstacle, and $x\in\mathcal X\subset\mathbb R^n$40 for dynamic. The dataset SHA-256 is 50e59d2f7886dcc00f3c2405ae449cdbaa879b1b5be79b7dcd2993ce65bb5145 (Lu et al., 15 May 2026).

A representative execution described in the source halted a downstream training stage because pre-registered convergence criteria were not met: a BC pre-train stage logged final loss $x\in\mathcal X\subset\mathbb R^n$41 but failed the pre-registered success-rate $x\in\mathcal X\subset\mathbb R^n$42. The watchdog fired, FailureForensics dumped the last $x\in\mathcal X\subset\mathbb R^n$43 steps to JSON, AtomicCheckpoint rolled back any half-written model, and the pipeline stopped before critic warm-up or PPO (Lu et al., 15 May 2026).

6. Theoretical commonalities, guarantees, and limits

Despite the terminological heterogeneity, the cited ParallelCBF variants share a common control-theoretic substrate. All are formulated for control-affine systems

$x\in\mathcal X\subset\mathbb R^n$44

all rely on a barrier-certified safe set of the form $x\in\mathcal X\subset\mathbb R^n$45 or an intersection of such sets, and all enforce safety through affine-in-$x\in\mathcal X\subset\mathbb R^n$46 inequalities or conservative affine surrogates. The distinction lies in where the parallel decomposition occurs: state-space regions in neural verification, multiple barriers in strip or box constraints, obstacle-body pairs in high-order stacking, policy rollouts in finite-horizon runtime safety, or vectorized environment instances in RL infrastructure (Vertovec et al., 9 Nov 2025).

The guarantees also differ materially. The neural-verification framework provides conservatively sufficient certification and can also construct a counterexample certificate through $x\in\mathcal X\subset\mathbb R^n$47 (Vertovec et al., 9 Nov 2025). Constant-sum barrier constructions and multiple-ECBF frameworks establish forward invariance of the intended safe set under stated relative-degree and initialization assumptions (Kim et al., 21 May 2025). The vector-relative-degree “Parallel CBF” gives compatibility, a closed-form controller, and local Lipschitz continuity (Cohen et al., 4 Sep 2025). The tractor-trailer formulation guarantees simultaneous enforcement of all stacked high-order barriers under feasibility and rank assumptions (Aali et al., 2022). PL-CBF provides finite-horizon safety, not infinite-horizon invariance, with completeness conditioned on policy-library coverage $x\in\mathcal X\subset\mathbb R^n$48 (Kim et al., 15 May 2026). The tensor-parallel RL framework does not claim a new invariance theorem beyond its dual-barrier filter, but instead elevates safety invariance tests and operational reproducibility to framework-level APIs (Lu et al., 15 May 2026).

A second common misconception is to interpret “parallel” as always meaning GPU acceleration. The literature shows four different meanings. In (Vertovec et al., 9 Nov 2025), it means independent simplicial-region certification with GPU batching. In (Kim et al., 21 May 2025, Cohen et al., 4 Sep 2025), and (Aali et al., 2022), it means parallel barrier constraints enforced simultaneously. In (Kim et al., 15 May 2026), it means independent finite-horizon rollouts across a policy library. In (Lu et al., 15 May 2026), it means tensor-parallel simulation and sharded BC-to-RL pipelines. This suggests that ParallelCBF is best understood as a family resemblance term: the central object is the CBF, while the adjective “parallel” identifies the decomposition strategy used to preserve tractability.

The limitations are similarly context-specific. Neural verification remains conservative because certification is based on affine lower bounds and refinement depth limits (Vertovec et al., 9 Nov 2025). Constant-sum constructions are specialized to strips between parallel boundaries and currently list bounded-input, mixed-relative-degree, and moving-boundary cases as extensions under study (Kim et al., 21 May 2025). The vector-relative-degree box-constraint formulation covers box constraints on outputs but does not explicitly handle input bounds (Cohen et al., 4 Sep 2025). The tractor-trailer approach inherits feasibility dependence on actuator limits and obstacle geometry (Aali et al., 2022). PL-CBF depends on accurate rollout simulation, a finite planning horizon, and library coverage, with model mismatch and perception uncertainty not explicitly handled (Kim et al., 15 May 2026). The RL framework is presently limited to a pure-Python, CPU PyTorch reference CBF, a 2D toy environment, and a RandomActionAlgorithm, with GPU-resident kernels and Isaac Lab adapters planned for later versions (Lu et al., 15 May 2026).