Is Your Safe Controller Actually Safe? A Critical Review of CBF Tautologies and Hidden Assumptions

Taekyung Kim

REVIEW 2 major objections 5 minor 20 references

Is Your Safe Controller Actually Safe? A Critical Review of CBF Tautologies and Hidden Assumptions

T0 review · 2 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper argues that control-barrier safety guarantees are logically empty unless the barrier feasibility condition is verified under the true actuation limits, and that many robotic safety demonstrations rely on passively safe, driftless

desk verdict A correct and useful CBF tutorial with no new theory; the sampled-data gap is real but explicitly hedged. read the letter →

arxiv 2603.06954 v2 pith:67454OK5 submitted 2026-03-07 cs.RO cs.SYeess.SY

Taekyung Kim This is my paper · ORCID

classification cs.ROcs.SYeess.SY

keywords controlbarrierfunctionsforwardinvariancefeasibilityverificationinputconstraintspassivesafetydriftlesssystemsCBF-QPsafety-critical

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a control barrier function (CBF) safety guarantee is only meaningful if the feasibility condition – the existence of an admissible control satisfying the barrier inequality at every state in the claimed safe set – is actually verified. Without that verification, invoking the CBF theorem reduces to Assumption 1: the system is safe because a safe controller was assumed to exist. The paper also shows that many published 'safe' demonstrations use driftless, passively safe models where safety is structurally trivial, so the results do not transfer to inertial robots. A double-integrator study illustrates the gap: naive geometric constraints collide 87–93% of the time, while properly tuned CBF constraints can achieve zero collisions when feasibility holds. The paper concludes with a checklist for constructing realizable safety arguments.

What carries the argument

The central object is the CBF feasibility condition K_cbf(x;α) := {u ∈ U | L_f h(x) + L_g h(x) u ≥ −α(h(x))}. The paper's key move is to treat nonemptiness of this set on the whole claimed safe set as a hypothesis that must be explicitly verified rather than assumed; that verification is what separates a candidate barrier from a valid certificate. The companion object is the class of passively safe systems, defined by L_f h(x) ≥ 0 for all x in C, for which the zero input preserves safety and makes any CBF-style guarantee structurally trivial.

What would settle it

Run the double-integrator CBF-QP (v_max=5, α=5) with an event-driven ODE solver that detects exact boundary crossings, and compare the collision/infeasibility rates to the paper's Table III (74% collision / 7% infeasibility). If the rates are dramatically lower under exact continuous-time integration, the reported failures are an artifact of the Δt=0.01 s discretization rather than of the CBF condition; if they persist, the feasibility condition's role is confirmed.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central claim is that the standard CBF safety theorem (Theorem 2) is conditional on verifying that K_cbf(x;α) ≠ ∅ for all x in the claimed safe set under the true input bounds U. When this verification is skipped, the resulting 'guarantee' is a tautology: it assumes the existence of a controller that keeps the system safe and then concludes safety. The paper formalizes this by distinguishing candidate CBFs (any differentiable constraint function) from valid CBFs (those verified to satisfy the barrier inequality), and by defining passively safe systems whose unforced dynamics already ensure forward invariance. For such driftless systems, even naive geometric hard

Load-bearing premise

The paper's quantitative simulations run in discrete time (Δt=0.01 s) while the CBF theorem it relies on is continuous-time, and no proof is given that the sampled-data implementation preserves the continuous-time invariance.

Editorial extensions

If this is right

A CBF-QP controller is undefined at any state where K_cbf(x;α) is empty, so formal safety claims require proving nonempty feasibility on the entire safe set, not just at sampled states.
Demonstrations on single integrators or kinematic manipulators, which are driftless and passively safe, provide no evidence of safety for systems with inertia; the same naive constraint causes 87–93% collisions on a double integrator.
Class-K tuning is a trade-off: steep α allows faster approach but demands more control authority and can push required inputs outside U; shallow α preserves feasibility but enforces conservative behavior.
Domain restriction (e.g., bounding velocity or imposing v ≥ −√(2 u_max p)) can recover feasibility for inertial systems, so safety claims must state the exact domain on which they hold.
Safety-as-soft-penalty methods (cost penalties, neural-CBF penalties, RL reward shaping) produce safety-informed behavior, not forward invariance, and cannot be presented as safety guarantees.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference — the same feasibility check could serve as a screening test for any proposed barrier candidate: compute K_cbf(x;α) over a dense grid or via sum-of-squares relaxations; an empty cell invalidates the certificate before deployment.
Editorial inference — because the safety theorem is continuous-time while implementations are sampled-data (here Δt=0.01 s), a natural follow-up is to prove a sampled-data invariance result; without it, the empirical 0% collisions may not inherit the continuous-time guarantee.
Editorial inference — the passive-safety distinction suggests that learned or tuned safety filters should first be asked whether the nominal (zero-input) system is already safe; if so, the filter's value added is about performance, not safety.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

The paper is a clearly written tutorial with a correct central point: too many CBF-based papers quote Theorem 2 without checking the feasibility hypothesis, and that collapses into the tautology 'safe because we assumed a safe controller exists.' The candidate-versus-valid CBF distinction and the passive-safety examples (single integrator, kinematic manipulator) are useful and well presented. The double-integrator viability calculation is standard but appropriately placed.

The paper does not claim deep novelty; it is a review/tutorial. The simulations are illustrative: they show the naive constraint fails on the double integrator and that tuning class-K involves trade-offs. The reader flagged the discrete-time/continuous-time gap, and that is a real soft spot: the cited theorem is continuous-time, while the implementation uses Δt = 0.01 s with no sampled-data guarantee. But this is explicitly disclaimed ('zero collisions do not prove safety'), so it does not undermine the main argument. It does mean the numerical tables should be read as illustrative, not as formal verification. Another minor point: the paper leans on the author's own prior work, but none of those citations is load-bearing for the critique, so no issue there.

The biggest weakness is simply that there is nothing new for experts. The failure of geometric constraints for inertial systems is well known; the value is in the framing and the compilation of warnings into a checklist. The paper would be useful for practitioners, students, and reviewers who want to push back on over-claimed safety.

I would send this to peer review—it is the kind of tutorial that should be vetted for correctness (which it mostly is) and can serve as a reference for good practice.

Referee Report

2 major / 5 minor

Summary. The manuscript is a tutorial-style critical review of Control Barrier Function (CBF) safety guarantees. It argues that the standard CBF theorem (Theorem 2) is routinely invoked without verifying the key feasibility hypothesis K_cbf(x; α) ≠ ∅ on the claimed safe set under the true input bounds, which reduces the resulting 'guarantee' to the tautological statement of Theorem 1 (safe because a safe controller is assumed to exist). The paper distinguishes candidate from valid CBFs, discusses passive safety in driftless systems, uses a double integrator to illustrate that geometric safe sets may not be controlled-invariant, and presents simulation studies on a single integrator, a double integrator, and a kinematic manipulator. It concludes with practical guidelines and an open-source web demo.

Significance. If its message is taken up, the paper addresses a real and widespread source of over-claiming in safety-critical robotics: theorems are cited without checking their hypotheses. The central logical point is correct and clearly presented, and the explicit distinction between candidate and valid CBFs, the passive-safety examples, and the double-integrator viability condition are valuable pedagogical contributions. The paper is not a new theoretical result but could have substantial impact as a tutorial/review. The open-source code and interactive demonstration are explicit strengths that increase reproducibility and utility.

major comments (2)

[§IV.D, Example 2] The statement 'if one restricts the domain by enforcing v∈[vmin,vmax] ... then feasibility can be recovered' is not correct in general. For any vmin<0, the state (p,v)=(0,vmin) is in the restricted domain, but under |u|≤umax the position immediately becomes negative; no admissible control can prevent violation. The valid restriction must be state-dependent, e.g., v≥−√(2umax p) for v<0, or must set vmin≥0 with the safe set defined accordingly. Please revise the example and the domain-restriction guideline to make clear that the safe set C itself must be shrunk to a controlled-invariant subset, not merely the ambient domain D.
[§VII, Tables II–III and §VII.B.4] The simulations are discrete-time with Δt=0.01 s, while Theorem 2 is a continuous-time statement. The paper does hedge that zero collisions do not prove validity, but it also states that collisions show the controller is 'demonstrably invalid'. A sampled-data collision can occur even if the continuous-time CBF condition holds, owing to discretization and numerical QP solving. Please add an explicit sentence in §VII.A.1 or §VII.B.4 clarifying that the simulation results are illustrative of the discrete-time implementation and are not direct evidence about the continuous-time CBF property.

minor comments (5)

[§VII.A.1] State explicitly the input constraint set U for each system (e.g., |u|≤v_max for the single integrator, |u|≤a_max for the double integrator, |ω|≤ω_max for the manipulator). This is needed to interpret the feasibility and collision results in the tables.
[Table III] When the QP is infeasible, what control input is actually applied in the simulation (zero input, last feasible input, or something else)? The reported infeasibility rate and its effect on the collision rate depend on this choice.
[§II.B.2] The list of safety-informed fallacies is useful; consider adding one sentence noting that these methods may still be combined with a formal CBF certificate, but the certificate itself must be verified separately.
[§V.A, Definition 5] The term 'passively safe' is not standard in the CBF literature; consider adding a short remark relating it to the cited references [14], [15] so that readers can trace the terminology.
[§III, Example 1] Using the speed of light as the negative velocity is a vivid choice, but it distracts from the structural point. A conventional large velocity would make the same argument while keeping the reader focused on the control-theoretic issue.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity; the central CBF-feasibility critique is self-contained.

full rationale

The paper's core argument is a logical analysis of the CBF theorem: invoking Theorem 2 without verifying K_cbf(x;α)≠∅ under true U reduces the guarantee to Assumption 1. This is supported by the paper's own double-integrator calculation (p(t)=0−ct+1/2ut^2<0) and the explicit feasibility condition (3). Theorem 1 is intentionally vacuous ('Immediate from the definition of forward invariance') and is the object of critique, not a load-bearing derivation. The author's self-citations ([3], [11], [12], [13]) are used only for the candidate-CBF label, tuning trade-offs, domain-restriction examples, and a planner application; none of these carries the central claim. The externally cited CBF theorem [1] and Lipschitz reference [2] are independent support. The simulation section explicitly disclaims that zero collisions prove validity ('Our experiments do not prove that a given tuning is valid when collision count is zero'), so no fitted input is renamed as a prediction. Section VII's discrete-time implementation (Δt=0.01 s, cvxpy/Gurobi) is a genuine limitation relative to the continuous-time theorem, but the paper hedges this by presenting simulations as illustrations rather than formal proof. The self-citations are minor and non-load-bearing; the central derivation is independent and not circular.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim is a logical analysis of conditional safety theorems; it introduces no fitted parameters, no new axioms beyond standard CBF theory, and no invented entities. The only hand-chosen numbers are simulation settings (α, v_max). The implicit sampled-data assumption is the only additional premise the empirical support relies on.

free parameters (2)

Class-K gain α (and HOCBF gains α1=α2) = grid over {1.0, 2.0, 3.0, 4.0, 5.0}
Hand-tuned in the simulation to demonstrate the aggressiveness/conservatism trade-off (Section IV.C, Table III); not fitted to reproduce data.
Maximum velocity v_max (double integrator domain bound) = grid over {1.0, 2.0, 3.0, 4.0, 5.0} m/s
Domain restriction parameter used in the double-integrator simulations (Section IV.D); chosen to make the CBF condition feasible on a bounded domain, not fitted to data.

assumptions (3)

standard math Theorem 2 (Ames et al. [1]): if h is a CBF on D with C⊆D, then any locally Lipschitz controller u(x) ∈ K_cbf(x; α) renders C forward invariant.
The paper's critique is built around this theorem; it is cited, not proved, and treated as a correct background result.
domain assumption Locally Lipschitz f, g, and feedback policies ensure well-defined closed-loop solutions on X.
Used in the preliminaries and in the 'well-posedness is part of the safety claim' discussion (Section I and II.A.4).
ad hoc to paper The discrete-time simulation with Δt = 0.01 s and QP solving approximates the continuous-time closed-loop system well enough to draw qualitative conclusions.
Implicit in Section VII; the paper never justifies this approximation, which is the weakest assumption of the empirical section.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Is Your Safe Controller Actually Safe? A Critical Review of CBF Tautologies and Hidden Assumptions." pith.science (2026). https://pith.science/paper/67454OK5

@misc{pith2026260306954,
  author       = {Pith},
  title        = {Pith review of: Is Your Safe Controller Actually Safe? A Critical Review of CBF Tautologies and Hidden Assumptions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/67454OK5}},
  note         = {Machine review of arXiv:2603.06954}
}

read the original abstract

This tutorial provides a critical review of the practical application of Control Barrier Functions (CBFs) in robotic safety. While the theoretical foundations of CBFs are well-established, I identify a recurring gap between the mathematical assumption of a safe controller's existence and its constructive realization in systems with input constraints. I highlight the distinction between candidate and valid CBFs by analyzing the interplay of system dynamics, actuation limits, and class-K functions. I further show that some purported demonstrations of safe robot policies or controllers are limited to passively safe systems, such as single integrators or kinematic manipulators, where safety is already inherited from the underlying physics and even naive geometric hard constraints suffice to prevent collisions. By revisiting simple low-dimensional examples, I show when CBF formulations provide valid safety guarantees and when they fail due to common misuses. I then provide practical guidelines for constructing realizable safety arguments for systems without such passive safety. A crowd-navigation simulation study further illustrates that CBF-derived reward shaping in reinforcement learning can improve empirical behavior without establishing formal safety. The goal of this tutorial is to bridge the gap between theoretical guarantees and actual implementation, supported by an open-source interactive web demonstration that visualizes these concepts intuitively.

Figures

Figures reproduced from arXiv: 2603.06954 by the authors.

**Figure 1.** https://cbf.taekyung.me: interactive web demonstration of the CBF Playground. The demo contrasts passively safe (driftless) systems (single integrator, kinematic manipulator), for which naive geometric constraints already prevent collisions, and a system with inertia (double integrator), where feasibility depends on actuation limits, velocity bounds, and class-K tuning. Assumption 1 (Existence of a Safe Controller)… view at source ↗

**Figure 2.** Snapshots of simulation trials. Red circles are obstacles and the green cross is the goal. Interested readers may evaluate the simulation framework [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references

[1]

Control Barrier Functions: Theory and Applications,

A. D. Ames, S. Coogan, M. Egerstedt, G. Notomista, K. Sreenath, and P. Tabuada, “Control Barrier Functions: Theory and Applications,” in European Control Conference (ECC), 2019, pp. 3420–3431

2019
[2]

Reformulations of Quadratic Programs for Lipschitz Continuity,

D. R. Agrawal, H. Lee, and D. Panagou, “Reformulations of Quadratic Programs for Lipschitz Continuity,”IEEE Control Systems Letters, vol. 9, pp. 2603–2608, 2025

2025
[3]

How to Adapt Control Barrier Functions? A Learning-Based Approach with Applications to a VTOL Quadplane,

T. Kim, R. W. Beard, and D. Panagou, “How to Adapt Control Barrier Functions? A Learning-Based Approach with Applications to a VTOL Quadplane,” inIEEE Conference on Decision and Control (CDC), 2025, pp. 7050–7057

2025
[4]

Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning,

L. Brunke, M. Greeff, A. W. Hall, Z. Yuan, S. Zhou, J. Panerati, and A. P. Schoellig, “Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 5, pp. 411–444, 2022

2022
[5]

Borrelli, A

F. Borrelli, A. Bemporad, and M. Morari,Predictive Control for Linear and Hybrid Systems. Cambridge University Press, 2017

2017
[6]

A general Hamilton- Jacobi framework for non-linear state-constrained control problems,

A. Altarovici, O. Bokanowski, and H. Zidani, “A general Hamilton- Jacobi framework for non-linear state-constrained control problems,” ESAIM: Control, Optimisation and Calculus of Variations, vol. 19, no. 2, pp. 337–357, 2013

2013
[7]

Robust Control Barrier–Value Functions for Safety-Critical Control,

J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust Control Barrier–Value Functions for Safety-Critical Control,” inIEEE Conference on Decision and Control (CDC), 2021, pp. 6814– 6821

2021
[8]

Semidefinite programming relaxations for semialgebraic problems,

P. A. Parrilo, “Semidefinite programming relaxations for semialgebraic problems,”Mathematical Programming, vol. 96, no. 2, pp. 293–320, 2003

2003

Show all 20 references

[9]

Correctness Guarantees for the Composition of Lane Keeping and Adaptive Cruise Control,

X. Xu, J. W. Grizzle, P. Tabuada, and A. D. Ames, “Correctness Guarantees for the Composition of Lane Keeping and Adaptive Cruise Control,”IEEE Transactions on Automation Science and Engineering, vol. 15, no. 3, pp. 1216–1229, 2018

2018
[10]

Rate-Tunable Control Barrier Functions: Methods and Algorithms for Online Adaptation,

H. Parwana and D. Panagou, “Rate-Tunable Control Barrier Functions: Methods and Algorithms for Online Adaptation,” 2023

2023
[11]

Learning to Refine Input Constrained Control Barrier Functions via Uncertainty-Aware Online Parameter Adaptation,

T. Kim, R. I. Kee, and D. Panagou, “Learning to Refine Input Constrained Control Barrier Functions via Uncertainty-Aware Online Parameter Adaptation,” inIEEE International Conference on Robotics and Automation (ICRA), 2025, pp. 3868–3875

2025
[12]

Beyond Collision Cones: Dynamic Obstacle Avoidance for Nonholonomic Robots via Dynamic Parabolic Control Barrier Functions,

H. K. Park, T. Kim, and D. Panagou, “Beyond Collision Cones: Dynamic Obstacle Avoidance for Nonholonomic Robots via Dynamic Parabolic Control Barrier Functions,” inInternational Conference on Robotics and Automation (ICRA), 2026

2026
[13]

Visibility-Aware RRT* for Safety-Critical Navigation of Perception-Limited Robots in Unknown Environments,

T. Kim and D. Panagou, “Visibility-Aware RRT* for Safety-Critical Navigation of Perception-Limited Robots in Unknown Environments,” IEEE Robotics and Automation Letters, vol. 10, no. 5, pp. 4508–4515, 2025

2025
[14]

Safe Trajectories for Autonomous Ren- dezvous of Spacecraft,

L. Breger and J. P. How, “Safe Trajectories for Autonomous Ren- dezvous of Spacecraft,”Journal of Guidance, Control, and Dynamics, 2012

2012
[15]

A Sampling Based Approach to Spacecraft Autonomous Maneuvering with Safety Spec- ifications,

J. A. Starek, B. W. Barbee, and M. Pavone, “A Sampling Based Approach to Spacecraft Autonomous Maneuvering with Safety Spec- ifications,” inAnnual AAS Guidance & Control Conference, 2015

2015
[16]

Comparative Analysis of Control Barrier Functions and Artificial Potential Fields for Obstacle Avoidance,

A. Singletary, K. Klingebiel, J. Bourne, A. Browning, P. Tokumaru, and A. Ames, “Comparative Analysis of Control Barrier Functions and Artificial Potential Fields for Obstacle Avoidance,” inIEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 8129–8136

2021
[17]

Control Barrier Functions for Systems with High Relative Degree,

W. Xiao and C. Belta, “Control Barrier Functions for Systems with High Relative Degree,” inIEEE Conference on Decision and Control (CDC), 2019, pp. 474–479

2019
[18]

Safe Control Synthesis via Input Constrained Control Barrier Functions,

D. R. Agrawal and D. Panagou, “Safe Control Synthesis via Input Constrained Control Barrier Functions,” inIEEE Conference on De- cision and Control (CDC), 2021, pp. 6113–6118

2021
[19]

Discrete-Time Control Barrier Function: High-Order Case and Adaptive Case,

Y . Xiong, D.-H. Zhai, M. Tavakoli, and Y . Xia, “Discrete-Time Control Barrier Function: High-Order Case and Adaptive Case,”IEEE Transactions on Cybernetics, vol. 53, no. 5, pp. 3231–3239, 2023

2023
[20]

CVXPY: A Python-Embedded Modeling Language for Convex Optimization,

S. Diamond and S. Boyd, “CVXPY: A Python-Embedded Modeling Language for Convex Optimization,”Journal of Machine Learning Research, 2016

2016

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Control Barrier Functions: Theory and Applications,

A. D. Ames, S. Coogan, M. Egerstedt, G. Notomista, K. Sreenath, and P. Tabuada, “Control Barrier Functions: Theory and Applications,” in European Control Conference (ECC), 2019, pp. 3420–3431

2019

[2] [2]

Reformulations of Quadratic Programs for Lipschitz Continuity,

D. R. Agrawal, H. Lee, and D. Panagou, “Reformulations of Quadratic Programs for Lipschitz Continuity,”IEEE Control Systems Letters, vol. 9, pp. 2603–2608, 2025

2025

[3] [3]

How to Adapt Control Barrier Functions? A Learning-Based Approach with Applications to a VTOL Quadplane,

T. Kim, R. W. Beard, and D. Panagou, “How to Adapt Control Barrier Functions? A Learning-Based Approach with Applications to a VTOL Quadplane,” inIEEE Conference on Decision and Control (CDC), 2025, pp. 7050–7057

2025

[4] [4]

Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning,

L. Brunke, M. Greeff, A. W. Hall, Z. Yuan, S. Zhou, J. Panerati, and A. P. Schoellig, “Safe Learning in Robotics: From Learning-Based Control to Safe Reinforcement Learning,”Annual Review of Control, Robotics, and Autonomous Systems, vol. 5, pp. 411–444, 2022

2022

[5] [5]

Borrelli, A

F. Borrelli, A. Bemporad, and M. Morari,Predictive Control for Linear and Hybrid Systems. Cambridge University Press, 2017

2017

[6] [6]

A general Hamilton- Jacobi framework for non-linear state-constrained control problems,

A. Altarovici, O. Bokanowski, and H. Zidani, “A general Hamilton- Jacobi framework for non-linear state-constrained control problems,” ESAIM: Control, Optimisation and Calculus of Variations, vol. 19, no. 2, pp. 337–357, 2013

2013

[7] [7]

Robust Control Barrier–Value Functions for Safety-Critical Control,

J. J. Choi, D. Lee, K. Sreenath, C. J. Tomlin, and S. L. Herbert, “Robust Control Barrier–Value Functions for Safety-Critical Control,” inIEEE Conference on Decision and Control (CDC), 2021, pp. 6814– 6821

2021

[8] [8]

Semidefinite programming relaxations for semialgebraic problems,

P. A. Parrilo, “Semidefinite programming relaxations for semialgebraic problems,”Mathematical Programming, vol. 96, no. 2, pp. 293–320, 2003

2003