Pith. sign in

REVIEW 3 major objections 4 minor 37 references

Sample Efficient Certification of Discrete-Time Control Barrier Functions

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Verifying a discrete-time control barrier function can be done with progressively fewer samples toward the interior of the safe set.

desk verdict Real theoretical contribution, but the numerical certificate uses a lower bound as an upper bound—repairable, but the sample-efficiency claim needs rework. read the letter →

arxiv 2509.03899 v1 pith:73GUBEMQ submitted 2025-09-04 eess.SY cs.LGcs.SY

Sampath Kumar Mulagaleti , Andrea Del Prete This is my paper
classification eess.SYcs.LGcs.SY
keywords controlbarrierfunctionsdiscrete-timesystemsinvariantsetssafetyverificationLipschitzcontinuitysamplecomplexityscenario-basedneuralnetwork
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

The paper tries to establish that verifying a discrete-time control barrier function h over its whole controlled-invariant sublevel set S(0) does not require a uniformly dense sample cover. Its proof divides S(0) into level segments C_i = {h in [gamma_{i-1}, gamma_i]} and shows that an epsilon_i-net over each segment suffices, with the allowed epsilon_i growing as the segment sits deeper inside S(0). The explicit bound on epsilon_i depends on the Lipschitz constants of h and the closed-loop dynamics and on the segment's level value, so near the boundary, where h is close to zero, samples must be dense, while the interior can be certified with much coarser nets. The numerical example reports roughly ten to twenty times fewer samples than an equivalent one-shot verification. If this is right, learned barrier functions and their accompanying safe sets become practical to certify on modest hardware, which matters for safety guarantees in model-predictive and learning-based control.

What carries the argument

Level-segment epsilon-net certification. The central device is the partition of S(0) into level sets C_i = {x : gamma_{i-1} <= h(x) <= gamma_i} together with an epsilon_i-net D_i over each C_i. The proof mechanism is a Lipschitz propagation inequality that bounds h(f(x)) - (1-alpha_bar)h(x) at an arbitrary x by its value at the nearest sample plus terms proportional to epsilon_i, so a sample-level check plus a distance bound yields the pointwise certificate. The second device is the recurrence gamma_{i+1} = a gamma_i + b delta from Proposition 4, which chooses the level thresholds so that the segment widths track the admissible epsilon_i; its fixed point gamma_hat defines the invariant subse

What would settle it

Reproduce the numerical example, then compute the maximum ratio ||f(x1,u(x1))-f(x2,u(x2))||/||x1-x2|| over a much denser set of pairs inside the segment closest to h=0. If any ratio exceeds the reported Lf=1.4325, use that pair's neighborhood to locate an x whose nearest verified sample is within the epsilon_i used in Table I and evaluate r(x)+alpha_bar h(x); a positive value is a direct counterexample to the certified barrier inequality (20).

Watch

Extended reading notes

Core claim

Proposition 2 is the load-bearing result. Let alpha in [0,1] be the decay used in synthesis, alpha_bar in [alpha,1] the decay to be certified, and suppose each level segment C_i is sampled by a set D_i that is an epsilon_i-net. If every sample satisfies r(x) = h(f(x))-h(x) <= -alpha h(x) - delta, then the barrier inequality r(x) <= -alpha_bar h(x) holds for every x in S(0) whenever epsilon_i <= (delta + (alpha_bar-alpha)|gamma_i|)/(L_h L_f + (1-alpha_bar)L_h), with L_h and L_f upper bounds on the Lipschitz constants of h and the closed-loop map. Because |gamma_i| is largest for segments far from the boundary, the admissible epsilon_i grows toward the interior. The paper also converts this in

Load-bearing premise

The proof requires a genuine upper bound on the Lipschitz constant of the closed-loop dynamics f(x,u(x)), but Algorithm 1 only estimates a lower bound from finitely many sampled pairs; if the true constant is larger, the required sampling densities in (21) are too optimistic and the certificate may not hold.

Editorial extensions

If this is right

  • Under Proposition 2, a user who only needs safety rather than a specified decay rate can set alpha_bar = 1 and use the loosest epsilon_i bounds, making pure invariance certification the cheapest case.
  • Enforcing a positive margin delta buys a quantifiable inner invariant set S(gamma_hat), so the certificate includes an explicit safety buffer between the certified region and the boundary of S(0).
  • A failed verification returns a concrete violating state, so the certifier can be scripted as a loop: certify, append the violating sample to the training set, re-solve the penalized synthesis, and re-certify.
  • The reported 2D example needs roughly 1.1e5 to 2.7e5 samples where a one-shot cover needs 2.0e6 to 2.8e6, making verification computationally accessible for this class of problems.
  • The probabilistic bound (36) grows exponentially with state dimension, so the practical payoff is currently concentrated in low-dimensional settings unless tighter 0-1 risk bounds are found.

Reading between the lines

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

  • A natural extension not developed in the paper would replace the sampled lower-bound estimate of L_f with a certified upper bound, for example from interval propagation or Lipschitz-bounded network architectures; then the certificates would be logically unconditional rather than contingent on the estimate.
  • The same level-segment stratification could be used during synthesis to place training samples preferentially near the boundary, where the verification epsilon_i is smallest, instead of sampling uniformly over D.
  • The result suggests an adaptive refinement strategy for higher-dimensional problems: allocate samples only to segments where the current epsilon_i bound is binding, and keep coarse nets where |gamma_i| is large, rather than refining the whole grid at once.
  • One can test whether the approach extends to continuous-time CBFs by applying the same stratification to a sampled flow map of the closed-loop continuous system; the paper's bounds are explicitly discrete-time, so that extension would require a new derivation.
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.

Referee Report

3 major / 4 minor

Summary. The paper proposes a certification procedure for discrete-time control barrier functions (DT-CBFs) based on Lipschitz continuity and level-set decomposition. The core idea is that the sampling density required for verifying the barrier decrease condition can be coarser away from the boundary of the CBF's sublevel set. The authors formalize this through Proposition 2, which gives a bound on the sampling resolution ϵ_i in terms of the CBF level γ_i, and then derive probabilistic sample-complexity bounds (Propositions 7–9). A synthesis formulation via a penalized learning problem and a certification algorithm (Algorithm 1) are provided, with a numerical example demonstrating improved sample efficiency over a one-shot verification approach.

Significance. If the identified gaps are addressed, the paper makes a useful contribution: it gives a level-dependent discretization criterion for DT-CBF verification, which is a natural and nontrivial extension of grid-based and Lipschitz-based certification methods. The theoretical framework is clearly presented, with deterministic and probabilistic sample-complexity bounds, and the authors provide a code link for reproducibility. However, the numerical validation currently rests on an invalid Lipschitz estimate, and the probabilistic and set-certification claims are incomplete as written. The core Proposition 2 is mathematically plausible, but the numerical evidence for the central sample-efficiency claim is not yet justified.

major comments (3)
  1. [Section IV-C, Eq. (21)] Algorithm 1 Step 1 estimates 'a lower bound to Lf' by sampling pairs, but Proposition 2 requires Lf to be an upper bound on the true closed-loop Lipschitz constant. In Section V, the value Lf=1.4325 is then used directly in (21). Since a sampled lower bound cannot certify an upper bound, the resulting ε_i values may exceed the admissible resolution, and the implication from (18) to (20) does not follow. The certificate reported in Table I is therefore not justified. Please use a certified upper bound (e.g., via Lipschitz bounds for the neural-network controller and dynamics) or clearly state a confidence interval and account for the uncertainty in the final certificate.
  2. [Section V, Algorithm 1] The parameter θ is an input to Algorithm 1 and appears in κθ(Ni) in Proposition 7 and in the confidence statement of Proposition 9, but no value of θ is reported in the numerical example. Without θ, the reported sample counts Ntot in Table I cannot be reproduced and the confidence of the probabilistic certificate is unspecified. Please report the chosen θ (and the resulting confidence (1−θ)^q).
  3. [Section IV.A, Problem 2] Proposition 4 verifies invariance of S(γ̂), not of S(0). In the verification procedure, the authors set γ_q = γ̂ and certify the barrier condition only on S(γ̂). However, Problem 2 asks for conditions under which S(0) is a CI set, and the paper's abstract/conclusion claim certification of the 0-sublevel set. This is a mismatch between the verified object and the stated certificate. To resolve this, either verify that X_s ⊆ S(γ̂) and adapt the problem statement, or provide conditions under which certification on S(γ̂) implies the original S(0) claim.
minor comments (4)
  1. [Eq. (25)] The denominator in the definition of b appears to be missing an L_h: it should be L_h L_f + (1−ᾱ)L_h, consistent with the preceding derivation. Although L_h cancels in the final expression for γ̂, the displayed formula is incorrect.
  2. [Section V] Typo: 'warm-staded' should be 'warm-started'.
  3. [Proposition 9] The confidence statement '(1−θ)^q' implicitly assumes independence across the q segments; if only individual segment bounds are used, the union bound gives 1−qθ. Please state the independence assumption explicitly.
  4. [Notation] The symbol B is overloaded: it denotes a box in Proposition 5, the unit 2-norm ball in Proposition 6, and a set in the numerical example. Please use distinct symbols to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the verification bounds are derived from Lipschitz and concentration arguments, not from the target certificate.

full rationale

The derivation chain from synthesis constraints (11) to verification bounds (21), (31), (36) is self-contained and does not reduce to its own inputs. Proposition 2 is a Lipschitz propagation argument: condition (18) verified on an epsilon_i-net over each level segment C_i is extended to all of S(0) using upper bounds L_h, L_f and the level bound gamma_i; the bound (21) is exactly the sufficient condition that makes the propagated residual non-positive, not a restatement of (18). The sample-size formulas (31)-(36) come from an external empirical-Bernstein bound [23] and a measure-concentration lemma [7]; neither is a self-citation, and neither assumes the barrier inequality at unsampled points. The numerical comparison N_tot < N_base compares two sample counts computed from the same formula for q=1 vs q>1; this is a mathematical consequence of the derived bound, not a fit to the certificate. The only self-citation is [27] (including Del Prete), which is a related-work pointer and is not load-bearing. A non-circular correctness caveat should be flagged: Proposition 2 requires L_f to be an upper Lipschitz bound, while Section IV-C (Algorithm 1, Step 1) explicitly samples pairs to 'estimate a lower bound to L_f', and Section V uses that estimate directly in (21); if the true closed-loop Lipschitz constant is larger, the numerical certificate in Table I would not be justified. This is an input-assumption violation, not a circular reduction, so it does not raise the circularity score. No circular step is present.

Assumptions & free parameters 6 free parameters · 7 assumptions · 0 invented entities

The derivation relies on standard Lipschitz and concentration arguments. The main uncertified inputs are the closed-loop Lipschitz constant Lf, the user-chosen alpha, alpha_bar, and delta, and the unverified containment of the safe set in S(gamma_hat). No new entities are introduced.

free parameters (6)
  • alpha (synthesis decay rate) = 0.01
    User-chosen scalar controlling the required decay in (7) and (11d). It affects the synthesis and the sample-complexity formula but is not fitted.
  • delta (verification slack) = 0.01
    User-chosen slack in (18); determines zeta in (21) and the invariant subset S(gamma_hat) in (24).
  • alpha_bar (target verification decay rate) = 0.4, 0.6, 0.8 in Table I
    User-chosen decay rate for certification; larger alpha_bar allows coarser sampling and a smaller certified invariant subset.
  • Lf (closed-loop Lipschitz estimate) = 1.4325
    Estimated from sampled state pairs as a lower bound, but used as an upper bound in (21). This is the main unverified number in the certificate.
  • Lh (barrier Lipschitz estimate) = 1.6854
    Estimated for the neural barrier h; it must be an upper bound for (21), but no certified computation is described.
  • theta (confidence level) = not reported
    Proposition 9 requires a confidence 1 minus theta for the probabilistic certificate, but the numerical section does not state theta.
assumptions (7)
  • domain assumption f(x,u) and h(x) are Lipschitz continuous on the relevant sets with constants Lf and Lh
    Used throughout Section IV, for example in the proof of Proposition 2 to bound h(f(x)) minus h(f(xbar)).
  • domain assumption The samples Di are i.i.d. draws from the uniform distribution on Ci
    Required for the 0-1 risk bound in Proposition 7 and the volume argument in Proposition 8.
  • standard math The empirical Bernstein bound of Maurer-Pontil [23, Theorem 4] is valid in the stated form
    Used directly as (31) to upper-bound the 0-1 risk.
  • standard math Lemma C.1 of Boffi et al. [7] bounding the in-radius of a measure-kappa set is valid
    Used as Proposition 8 to convert a measure bound into a ball-radius bound.
  • domain assumption The QP-based controller u(x; theta_u) is continuous and Lipschitz in x, so the closed-loop system is Lipschitz
    Needed for Lf to exist and for the implicit function theorem differentiability statement in Section III.
  • ad hoc to paper The estimated Lf is an upper bound on the true closed-loop Lipschitz constant
    The paper only computes a lower bound via samples in Section IV-C, yet (21) and the numerical certificate require an upper bound.
  • ad hoc to paper The safe set Xs is contained in S(gamma_hat), so that certifying S(gamma_hat) suffices for safety
    Algorithm 1 sets gamma_q equal to gamma_hat in Section IV-C and certifies only S(gamma_hat); containment of Xs is not checked in the numerical example.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Sample Efficient Certification of Discrete-Time Control Barrier Functions." pith.science (2026). https://pith.science/paper/73GUBEMQ

@misc{pith2026250903899,
  author       = {Pith},
  title        = {Pith review of: Sample Efficient Certification of Discrete-Time Control Barrier Functions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/73GUBEMQ}},
  note         = {Machine review of arXiv:2509.03899}
}
read the original abstract

Control Invariant (CI) sets are instrumental in certifying the safety of dynamical systems. Control Barrier Functions (CBFs) are effective tools to compute such sets, since the zero sublevel sets of CBFs are CI sets. However, computing CBFs generally involves addressing a complex robust optimization problem, which can be intractable. Scenario-based methods have been proposed to simplify this computation. Then, one needs to verify if the CBF actually satisfies the robust constraints. We present an approach to perform this verification that relies on Lipschitz arguments, and forms the basis of a certification algorithm designed for sample efficiency. Through a numerical example, we validated the efficiency of the proposed procedure.

Figures

Figures reproduced from arXiv: 2509.03899 by the authors.

Figure 2
Figure 2. Variation of number of samples required for [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 33 canonical work pages

  1. [1]

    Discrete control barrier functions for safety-critical control of discrete systems with application to bipedal robot navigation

    Ayush Agrawal and Koushil Sreenath. Discrete control barrier functions for safety-critical control of discrete systems with application to bipedal robot navigation. In Robotics: Science and Systems, volume 13, pages 1–10. Cambridge, MA, USA, 2017

    work page 2017

  2. [2]

    Burdick, and Aaron D

    Mohamadreza Ahmadi, Andrew Singletary, Joel W. Burdick, and Aaron D. Ames. Safe policy synthesis in multi-agent pomdps via discrete-time barrier functions. In 2019 IEEE 58th Conference on Decision and Control (CDC), pages 4797–4803, 2019

    work page 2019

  3. [3]

    Lus F. C. Alberto, Tas R. Calliero, and Andre C. P. Martins. An invariance principle for nonlinear discrete autonomous dynamical systems. IEEE Transactions on Automatic Control, 52(4):692– 697, 2007

    work page 2007

  4. [4]

    Ames, Samuel Coogan, Magnus Egerstedt, Gennaro Notomista, Koushil Sreenath, and Paulo Tabuada

    Aaron D. Ames, Samuel Coogan, Magnus Egerstedt, Gennaro Notomista, Koushil Sreenath, and Paulo Tabuada. Control barrier functions: Theory and applications. In 2019 18th European Control Conference (ECC), pages 3420–3431, 2019

    work page 2019

  5. [5]

    Bertsekas

    D. Bertsekas. Infinite time reachability of state-space regions by using feedback control. IEEE Transactions on Automatic Control, 17(5):604–613, 1972

    work page 1972

  6. [6]

    Set-Theoretic Methods in Control

    Franco Blanchini and Stefano Miani. Set-Theoretic Methods in Control. Springer International Publishing, 2015

    work page 2015

  7. [7]

    Learning stability certificates from data

    Nicholas Boffi, Stephen Tu, Nikolai Matni, Jean-Jacques Slo- tine, and Vikas Sindhwani. Learning stability certificates from data. In Jens Kober, Fabio Ramos, and Claire Tomlin, editors, Proceedings of the 2020 Conference on Robot Learning, volume 155 of Proceedings of Machine Learning Research, pages 1341–

    work page 2020

  8. [8]

    Bravo, D

    J.M. Bravo, D. Limon, T. Alamo, and E.F. Camacho. On the computation of invariant sets for constrained nonlinear systems: An interval arithmetic approach. Automatica, 41(9):1583–1589, 2005

    work page 2005

Show all 37 references
  1. [9]

    Brown, M

    S. Brown, M. Khajenejad, S. Z. Yong, and S. Mart ´ınez. Comput- ing controlled invariant sets of nonlinear control-affine systems. arXiv preprint arXiv:2304.11757, 2023

    2023 arXiv

  2. [10]

    Byrd, Peihuang Lu, Jorge Nocedal, and Ciyou Zhu

    Richard H. Byrd, Peihuang Lu, Jorge Nocedal, and Ciyou Zhu. A limited memory algorithm for bound constrained optimization. SIAM Journal on Scientific Computing, 16(5):1190–1208, 1995

    1995

  3. [11]

    Verification- aided learning of neural network barrier functions with termina- tion guarantees

    Shaoru Chen, Lekan Molu, and Mahyar Fazlyab. Verification- aided learning of neural network barrier functions with termina- tion guarantees. arXiv preprint arXiv:2403.07308, 2024

    2024 arXiv

  4. [12]

    G. Chesi. Domain of Attraction: Analysis and Control via SOS Programming. Lecture Notes in Control and Information Sciences. Springer London, 2011

    2011

  5. [13]

    Verification and synthesis of control barrier functions

    Andrew Clark. Verification and synthesis of control barrier functions. In 2021 60th IEEE Conference on Decision and Control (CDC), pages 6105–6112, 2021

    2021

  6. [14]

    Safe reinforcement learning using robust control barrier functions

    Yousef Emam, Gennaro Notomista, Paul Glotfelter, Zsolt Kira, and Magnus Egerstedt. Safe reinforcement learning using robust control barrier functions. IEEE Robotics and Automation Letters, pages 1–8, 2022

    2022

  7. [15]

    Efficient and accurate estimation of lipschitz constants for deep neural networks

    Mahyar Fazlyab, Alexander Robey, Hamed Hassani, Manfred Morari, and George Pappas. Efficient and accurate estimation of lipschitz constants for deep neural networks. Advances in neural information processing systems, 32, 2019

    2019

  8. [16]

    Fiacchini, T

    M. Fiacchini, T. Alamo, and E.F. Camacho. On the computation of convex robust control invariant sets for nonlinear systems. Automatica, 46(8):1334–1338, 2010

    2010

  9. [17]

    Victor Freire and Marco M. Nicotra. Systematic design of discrete-time control barrier functions using maximal output admissible sets. IEEE Control Systems Letters, 7:1891–1896, 2023

    2023

  10. [18]

    Ad- vances in the theory of control barrier functions: Addressing practical challenges in safe control synthesis for autonomous and robotic systems

    Kunal Garg, James Usevitch, Joseph Breeden, Mitchell Black, Devansh Agrawal, Hardik Parwana, and Dimitra Panagou. Ad- vances in the theory of control barrier functions: Addressing practical challenges in safe control synthesis for autonomous and robotic systems. Annual Reviews...

    2024

  11. [19]

    On the effectiveness of interval bound propagation for training verifiably robust models

    Sven Gowal, Krishnamurthy Dvijotham, Robert Stanforth, Rudy Bunel, Chongli Qin, Jonathan Uesato, Relja Arandjelovic, Timo- thy Mann, and Pushmeet Kohli. On the effectiveness of interval bound propagation for training verifiably robust models. arXiv preprint arXiv:1810.12715, 2018

    2018 arXiv

  12. [20]

    Kerrigan

    Eric C. Kerrigan. Robust Constraint Satisfaction: Invariant Sets and Predictive Control. PhD thesis, University of Cambridge, 2000

    2000

  13. [21]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2015

    2015 arXiv

  14. [22]

    Milan Korda, Didier Henrion, and Colin N. Jones. Convex com- putation of the maximum controlled invariant set for discrete- time polynomial control systems. In 52nd IEEE Conference on Decision and Control, pages 7107–7112, 2013

    2013

  15. [23]

    Empirical Bernstein Bounds and Sample-Variance Penalization

    Andreas Maurer and Massimiliano Pontil. Empirical Bernstein Bounds and Sample-Variance Penalization. In Proc. 22nd Annual Conference on Learning Theory (COLT), July 2009

    2009

  16. [24]

    Mitchell, A.M

    I.M. Mitchell, A.M. Bayen, and C.J. Tomlin. A time-dependent hamilton-jacobi formulation of reachable sets for continuous dynamic games. IEEE Transactions on Automatic Control, 50(7):947–957, 2005

    2005

  17. [25]

    Jorge Nocedal and Stephen J. Wright. Numerical Optimization. Springer, New York, NY , USA, 2e edition, 2006

    2006

  18. [26]

    Fast verification of control barrier functions via linear programming

    Ellie Pond and Matthew Hale. Fast verification of control barrier functions via linear programming. IFAC-PapersOnLine, 56(2):10595–10600, 2023. 22nd IFAC World Congress

    2023

  19. [27]

    Vboc: Learning the viability boundary of a robot manipulator us- ing optimal control

    Asia La Rocca, Matteo Saveriano, and Andrea Del Prete. Vboc: Learning the viability boundary of a robot manipulator us- ing optimal control. IEEE Robotics and Automation Letters, 8(11):6971–6978, 2023

    2023

  20. [28]

    Scalable computation of robust control invariant sets of nonlinear systems

    Lukas Sch ¨afer, Felix Gruber, and Matthias Althoff. Scalable computation of robust control invariant sets of nonlinear systems. IEEE Transactions on Automatic Control, 69(2):755–770, 2024

    2024

  21. [29]

    Shakhesi

    E. Shakhesi. Verification and synthesis of discrete-time control barrier functions. Master’s thesis, TU Eindhoven, Nov 2023. Accessed: date-of-access

    2023

  22. [30]

    Strong and Leila J

    Amy K. Strong and Leila J. Bridgeman. Data driven verification of positive invariant sets for discrete, nonlinear systems. In Alessandro Abate, Mark Cannon, Kostas Margellos, and Anto- nis Papachristodoulou, editors, Proceedings of the 6th Annual Learning for Dynamics &; Contr...

    2024

  23. [31]

    Intriguing properties of neural networks

    Christian Szegedy, Wojciech Zaremba, Ilya Sutskever, Joan Bruna, Dumitru Erhan, Ian Goodfellow, and Rob Fergus. Intriguing properties of neural networks. arXiv preprint arXiv:1312.6199, 2013

    2013 arXiv

  24. [32]

    Dimarogonas

    Xiao Tan and Dimos V . Dimarogonas. Compatibility checking of multiple control barrier functions for input constrained systems. In 2022 IEEE 61st Conference on Decision and Control (CDC), pages 939–944, 2022

    2022

  25. [33]

    Predictive control barrier functions: Enhanced safety mechanisms for learning- based control

    Kim P Wabersich and Melanie N Zeilinger. Predictive control barrier functions: Enhanced safety mechanisms for learning- based control. IEEE Transactions on Automatic Control, 68(5):2638–2651, 2022

    2022

  26. [34]

    Safety verification and controller synthesis for systems with input constraints

    Han Wang, Kostas Margellos, and Antonis Papachristodoulou. Safety verification and controller synthesis for systems with input constraints. IFAC-PapersOnLine, 56(2):1698–1703, 2023. 22nd IFAC World Congress

    2023

  27. [35]

    Safety-critical model predictive control with discrete-time control barrier func- tion

    Jun Zeng, Bike Zhang, and Koushil Sreenath. Safety-critical model predictive control with discrete-time control barrier func- tion. In 2021 American Control Conference (ACC), pages 3882– 3889, 2021

    2021

  28. [36]

    Exact verification of relu neural control barrier functions

    Hongchao Zhang, Junlin Wu, Yevgeniy V orobeychik, and An- drew Clark. Exact verification of relu neural control barrier functions. Advances in neural information processing systems, 36:5685–5705, 2023

    2023

  29. [1350]

    PMLR, 16–18 Nov 2021

    2021

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.