acceptodds
Under review as a conference paper at ICLR 2027

Neural Feasibility Certificates

Abstract

Ensuring recursive feasibility is critical for safety-constrained online control under hard constraints. Since analytically constructing feasible regions and recovery policies is difficult for nonlinear systems, learning-based methods offer a flexible alternative. However, existing methods often rely on approximate boundaries and continuously constrain the nominal controller, leading to limited guarantees and unnecessary conservatism. To address these limitations, we propose a framework to synthesize and formally verify neural feasibility certificates for safety-constrained online control. Our contributions are threefold. First, we introduce a neural feasibility certificate defining an inner approximation of the instantaneous feasible set. The associated margin-triggered controller retains the original online controller when sufficient feasibility margin is available and activates a learned witness policy only near the boundary of the certified feasible region, where it drives the system back to the feasible interior in finite time. Second, we develop a synthesis and formal verification framework that jointly learns the feasibility certificate and witness policy and certifies recursive feasibility over the continuous state space. Third, we validate the proposed framework on adaptive cruise control, autonomous overtaking, and mobile robot navigation against multiple baselines, demonstrating improved feasibility with competitive safety, control performance, and computational efficiency.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Control Barrier Functions, Recursive Feasibility, Neural Certificates

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_77ezOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a framework that jointly learns a neural feasible-set certificate and a witness policy for discrete-time safety-constrained control. The certificate defines an inner approximation of the instantaneous feasible set, while a margin-triggered switch retains the original optimization-based controller in the certified interior and activates the witness policy near the boundary to restore the feasibility margin in finite steps. Set-inclusion checks and Lipschitz-based verification provide sufficient conditions for recursive feasibility and constraint satisfaction.

Critical strengths
  1. The framework clearly distinguishes instantaneous from closed-loop recursive feasibility and provides sufficient conditions for constraint satisfaction and finite-step recovery, addressing a practical failure mode of safety filters.

  2. Jointly learning the certificate and recovery policy enables intervention only when needed while preserving the original controller in the certified interior. Results across three tasks provide encouraging evidence of feasibility with modest online computational cost.

Critical weaknesses
  1. The formal-certification claims cannot be reproduced from the supplied details: verification domains, certified Lipschitz constants, covering radii, grid sizes, verifier encodings, and witness-admissibility checks are missing. Local certificates are also combined without establishing the composition assumptions, making it difficult to assess whether Theorem 6 covers the implemented controllers.

  2. All feasibility comparisons restrict initial states to the proposed certified set. Although useful for paired evaluation, this does not establish broader coverage or reduced conservatism. Certified-set size, baseline recoverable-set coverage, recovery frequency, and relevant ablations are unreported, while the strongest baselines already achieve 100% feasibility in two tasks.

  3. The paper adopts various concepts and techniques from the literature, including HOCBF, recursive feasibility, CEGIS, and neural representation of certificate. The key novelty in the methodology is not clearly elaborated.

Questions for authors
  1. What is the key innovation in the method? Considering that the various concepts and techniques are adopted from the literature, including HOCBF, recursive feasibility, CEGIS, and neural representation of certificate.

  2. What exact verification domains, certified Lipschitz bounds, grid covering radii and sizes, switching margins, and verifier encodings support each reported certification?

  3. How are optimizer regularity and successor containment in the set \mathcal{X} established?

  4. How are the nonlinear discrete-time robot/overtaking HOCBF constraints verified, and which composition theorem applies to the implemented shared/minimum certificates? Please provide evidence that its assumptions hold.

  5. What fraction of the evaluation domain is certified, and how do coverage and intervention frequency compare with Backup CBF and FGO? Does the advantage persist under a common initial-state distribution independent of the proposed certificate?

Reviewer_NoHGOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a Neural Feasibility Certificate Framework to maintain recursive feasibility in discrete-time control-affine systems under hard constraints. The framework synthesizes a neural feasibility certificate h_{feas} defining an inner approximation of the true instantaneous feasible set, paired with a learned neural witness policy \pi _{w}. A hard switching rule activates the witness policy near the boundary to drive the system state back to a designated interior zone in finite steps. Offline verification is conducted by mapping set-inclusion to an UNSAT query via Farkas' Lemma, and evaluating continuous-state recovery via a grid-sampling layout bounded by estimated Lipschitz constants.

Critical strengths
  1. The paper provides a mathematically rigorous formulation of recursive feasibility by treating it as a localized, finite-time recovery contract near the constraint boundaries.

  2. The framework demonstrates excellent empirical validation across classic safety-critical testbeds. It achieves a 100% closed-loop tracking feasibility rate on the evaluated low-dimensional benchmarks, outperforming several optimization-based and learned baselines.

Critical weaknesses
  1. Lack of Core Novelty (Incremental Engineering Integration)


The technical machinery presented is an integration of pre-existing components rather than a novel conceptual advancement:

The margin-triggered switching protocol between a nominal task controller and an emergency backup policy is structurally identical to classical Supervisory Control, Invariance Control, and traditional Backup CBF designs

Approximating feasibility sets via machine learning to construct barrier constraints directly follows the paradigms already established in the cited Xiao et al. (2020, 2023a, 2026) literature.

The formal verification step is delegated entirely to third-party, off-the-shelf neural network verifiers (α,β-CROWN / Marabou).

The paper functions as an engineering assembly of known techniques rather than a contribution of fundamental machine learning or control theory.

  1. Severe Computational Bottlenecks & Failure to Scale


The continuous-space safety verification guarantees rest on Lemma 5 and Theorem 6, which introduce a severe Curse of Dimensionality:


Exponential Grid Density: To maintain valid sound bounds over the continuous domain, the sampling radius \eta_{x} must be infinitesimally small. For systems beyond basic 3D/4D toy problems, the number of grid points \mathcal{X}_{grid} required to cover the state space scales exponentially.


Over-conservatism from Lipschitz Constants: Computing global or regional Lipschitz constants L_h, L_{F_w}, L_{F^*} for deep neural networks combined with nonlinear dynamics is notoriously conservative. This forces the switching margin δ to be excessively large (per Lemma 2), drastically shrinking the nominal controller's operating envelope and crippling the utility of the original online optimizer.


The compositional assumptions in Appendix C (independent input blocks or completely disjoint recovery zones) are trivial special cases that are immediately violated by highly coupled real-world dynamics.


 3. Hard Boundary Switching and Destructive Chattering


The crisp, discontinuous control law presented in Equation (5) is physically unviable:

uk={u∗(xk),xk∈Sδ,πw(xk),xk∈Bδ.u_{k}=\begin{cases}u^{*}(x_{k}),&x_{k}\in \mathcal{S}_{\delta },\\ \pi_{w}(x_{k}),&x_{k}\in \mathcal{B}_{\delta }.\end{cases}

When the state enters \mathcal{B}{\delta}, the witness policy engages and pushes the system toward the interior. The exact instant it crosses back over the h{feas}(x) = \delta hyperplane, control reverts to the nominal law u^{*}. If the nominal task goal is aggressive or close to the constraint boundary, it will immediately drive the system back down into Bδ\mathcal{B}_{\delta }.


This creates high-frequency chattering/oscillations directly on the switching surface. In physical hardware, this behavior causes severe actuator wear, saturation, and destabilizing unmodeled high-frequency dynamics. The authors do not propose any hysteresis margin, smoothing functions, or sliding-mode analysis to address this obvious control artifact.

Questions for authors
  1. The verification times in Table 3 hover around 3–4 minutes for extremely low-dimensional systems (ACC and 2D mobile robots). Provide explicit scaling formulations (or empirical numbers) showing the grid size and verification time expansion when moving to a standard 12-state quadrotor or a 6-DoF manipulator arm.
  2. Provide the high-frequency temporal plots of the raw control inputs uku_{k} when the system is operating near the boundary threshold δ. How does your framework guarantee that physical actuators will not suffer from destructive chattering?
  3. quantifiably state how large the safety margin δ becomes under realistic, validated Lipschitz bounds. What percentage of the actual valid feasible space Xfeas\mathcal{X}_{feas} remains accessible to the nominal controller u∗u^{*} after subtracting this conservative margin?
Reviewer_iFMFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper uses a discrete time model of a system under control, specifies constraints to be satisfied, uses a control barrier function methodology to inspire a neural approximation to a controller which maintains the constraints, and demonstrates results on some simulated control problems. While potentially interesting and useful, the paper ignores a large and very active area of safety based control design which should be included in the baselines.

Critical strengths

The paper presentation is fairly clear, the result presents a new result on recursive feasibility, and the simulation examples chosen on which to evaluate the baselines seem reasonable.

Critical weaknesses

This paper is that it ignores a large body of literature associated with recursive feasibility through model predictive control and through Hamilton-Jacobi reachability. In both areas, learning-based methods have been proposed to deal with the scalability issues of dynamic programming, and certification methods designed on top of these.

For example, see the work of Zeilinger’s group and that of Fisac and others, who build neural certificates based on reachability formulations, and provide methods to calibrate and verify these.

The certified set inclusion property has been likewise already considered in related work on the discriminating hyperplane formulation of reachable sets.

Also, please comment on the control chattering induced by your method.

Questions for authors

Please explain how the feasibility certificates of Section 4 help with these two problems:

(1) For the HOCBF condition, how do you know that the control authority that is required to keep the system safe with respect to one of the constraints has non-empty intersection with the control authority that is required to keep the system safe with respect to a second, or a third constraint?

(2) Similarly, some constraints may have very different relative degrees than others, so not only the variation in the constraints, but in their relative degrees, could make the discovery of such inputs very challenging if not impossible.

I don’t understand how the Ames 2016 paper performed so poorly on the benchmarks. Can you discuss this more?

Finally, can you comment on the chattering that your method exhibits – this would be prohibitive in any real control scenario.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.