Convex Safety Filtering via Spectral Selection for Nonconvex Safe Sets

arXiv:2610.12324 · eess.SY, cs.SY, math.OC · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa, and with me are Dev and Taro, guest researcher.

Dev: Today's paper: "Convex Safety Filtering via Spectral Selection for Nonconvex Safe Sets".

Rosa: The gist The discrete-time control barrier function condition for a safe set that is a union of convex sets is nonconvex in the control input,

Dev: First, who's behind it and why it matters.

Title and authors: Rosa: So, we're looking at this paper, "Convex Safety Filtering via Spectral Selection for Nonconvex Safe Sets," and it’s tackling something really specific about safety filters that usually get messy when the safe area is made up of several simple shapes.

Dev: It sounds like they are dealing with a situation where the math gets nonconvex based on how the safe set is defined, which makes finding a simple, reliable control input tricky because you can't just use one standard method for everything.

Taro: I’m curious if this means we can finally get a filter that handles these complex boundaries without having to solve massive semidefinite programs every single time we run the system.

Rosa: Exactly, that’s the core idea here, and the title points out they are using spectral selection to make this whole process convex again.

Dev: It’s about taking this nonconvex problem and showing you can turn it into something manageable by focusing only on a specific part of the system's matrix function at a given moment.

The paper's summary: Rosa: What the authors are showing in this paper is that when the matrix function defining the safe set is concave, even if your safe area is just a union of convex shapes, you can treat it as if it were simpler.

Dev: They pinpoint exactly where that nonconvexity comes from—it happens because you only need *one* of those convex shapes to be active at any given time for the system to be safe, not all of them simultaneously.

Taro: So, instead of having to check every single boundary condition in the union, they propose a trick: picking specific eigenvectors from that matrix function at your current state.

Rosa: Right, and by selecting those specific eigenvectors—which are tied directly to whether you're inside or outside the set—they get a convex constraint that only cares about those eigenvalues.

Dev: That means the control input constraint becomes much simpler because it’s not checking every single condition for every possible shape in the union; it just targets what actually matters for safety.

The paper's improvements: Rosa: Now, looking at the improvements they lay out, one big thing is how they guarantee that this new selection method works reliably along a path you’re actually driving or walking.

Dev: They prove that every eigenvalue that determines whether you are safe ends up satisfying a geometric lower bound along those closed-loop trajectories, which is a strong safety property to have.

Taro: That geometric bound is important because it means that as long as your system stays on the planned path, those critical eigenvalues won't drift too far away from where they need to be for membership in the set.

Rosa: And they show this construction doesn't rely on projecting onto the unsafe set at every step; instead, it just uses this eigenvector selection process.

Dev: That’s a big win for real-time control because you avoid those computationally heavy projection steps, which is what makes their method much faster than the full matrix condition they are comparing it to.

Conclusion: Rosa: To wrap up, the paper on "Convex Safety Filtering via Spectral Selection for Nonconvex Safe Sets" shows that by selecting eigenvectors of the matrix function at your current state, you can build a convex input constraint that keeps things safe even when your safe region is a union of simple shapes.

Dev: The main implication for control engineers is the speed; they compare their proposed filter to the full-matrix semidefinite program and show it’s orders of magnitude faster, with one solve time being one point nine ms versus ninety-two ms for the other.

Taro: From an autonomy perspective, this means we can build systems that handle these complex environmental boundaries efficiently without bogging down the control loop when things get complicated.

Rosa: It really shows a way to handle those tricky nonconvex constraints in safety filtering without sacrificing speed, which is crucial when you’re deploying these systems outside of a perfect lab setting.

Dev: So, for anyone working on discrete-time control barrier functions with unions of convex sets, this paper gives you a concrete method that preserves forward invariance while being much more computationally efficient.

Juan Augusto Paredes Salazar, James Usevitch, Ankit Goel

Department of Mechanical Engineering, University of Maryland, Baltimore County · Department of Aerospace Engineering, The University of Michigan · Department of Electrical And Computer Engineering, Brigham Young University

eess.SY, cs.SY, math.OC

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: 7 pages, 3 figures, submitted to the ACC 2027

License: http://creativecommons.org/licenses/by/4.0/

The gist: The gist The discrete-time control barrier function condition for a safe set that is a union of convex sets is nonconvex in the control input, and this paper shows that selecting eigenvectors of the

Key concepts

Control Barrier Function (CBF): A mathematical tool used in control systems to ensure that a system stays within a predefined safe region. It defines constraints on the control inputs such that the system's trajectory remains safe, even when faced with disturbances.
Safe Set Union of Convex Sets: This refers to a region where safety is guaranteed if at least one of several convex regions is satisfied. The challenge is that this union structure makes the standard safety condition nonconvex in terms of the control inputs.
Eigenvector Selection Constraint: The core idea involves choosing specific eigenvectors from a matrix function evaluated at the current state. This selection process generates a convex constraint that effectively targets only the eigenvalues responsible for determining whether the system is inside or outside the safe set.

Terminology

Trace — 4× in this paper · explained in 2 episodes
Core — 3× in this paper · explained in 2 episodes
Robustness — 2× in this paper · explained in 5 episodes
Verification — 2× in this paper · explained in 2 episodes
Adversarial Robustness — 1× in this paper · explained in 3 episodes
Hybrid System — 1× in this paper · explained in 2 episodes

Summary

The gist The discrete-time control barrier function condition for a safe set that is a union of convex sets is nonconvex in the control input, and this paper shows that selecting eigenvectors of the matrix function at the current state yields a convex input constraint that acts only on the eigenvalues determining membership in the set.

How it works

The paper addresses the nonconvexity arising when safe sets are unions of convex sets, which occurs because at least one member of the union must hold, not every member (Page 1). The core idea is to construct a convex constraint that only targets the eigenvalues responsible for set membership (Page 1).

The construction involves several key steps:

Selecting eigenvectors of the matrix function at the current state yields a convex input constraint that acts only on the eigenvalues determining membership in the set

This selection process is formalized by Proposition IV.1, which states that if H is matrix concave, then for any index j, S(j) is represented as a union of convex sets: "S(j) = [Y ∈Op,r

The paper shows that each set in the union is convex (Page 4).

Key Findings and Contributions

The main contributions of the paper are enumerated as follows:

The safe set is shown to be a union of convex sets when the matrix function is matrix concave (Page 2). This covers complements of polytopes and spectrahedra, which are written as such sets in [21], and unions of convex sets (Page 5).
A convex input constraint is constructed by selecting eigenvectors at the current state, for any number of required nonnegative eigenvalues (Page 5). This construction does not use projection onto the unsafe set (Page 5).
Every eigenvalue that determines membership in the set is shown to satisfy a geometric lower bound along closed-loop trajectories (Page 5).
The condition of [21] that bounds the entire matrix is shown to imply the proposed constraint, so the proposed constraint admits at least as many inputs at each state (Page 5).

Comparison and Simulation Results

The paper compares the proposed filter with the full-matrix condition using a double-integrator simulation with a polytope obstacle and a spectrahedron obstacle (Page 1).

The proposed filter is more than an order of magnitude faster than the full-matrix SDP

The results confirm that Both eigenvalue traces remain nonnegative for the selected and full-matrix filters, which confirms the forward invariance of both filters (Page 6). Furthermore, the difference between their position trajectories has a mean of 2.8458 · 10−2 T and standard devition of 3.8847 · 10−2 T (Page 6).

Conclusion

In conclusion, Selecting eigenvectors of the matrix function at the current state yields a convex constraint that preserves forward invariance (Page 5). This method provides a computationally efficient safety filter for safe sets that are unions of convex sets, offering significant speed improvements over full-matrix methods (Page 6). The paper concludes by noting that The mean solve time per step was 1.9 ms for the proposed quadratic program and 92 ms for the full-matrix semidefinite program (Page 6). This work establishes a method to handle nonconvex constraints in control barrier functions efficiently (Page 5). The authors also thank and acknowledge AI tools, including ChatGPT and Claude, for reviewing and revising the manuscript for technical and grammatical errors (Page 7). The paper is organized as follows: Section II gives notation, matrix concavity, and the discrete-time MCBF framework of [21] (Page 5). Section V compares the two filters on a double integrator with two obstacles (Page 5). The paper is organized as follows: Section VI concludes the paper (Page 5). The authors also thank and acknowledge AI tools, including ChatGPT and Claude, for reviewing and revising the manuscript for technical and grammatical errors (Page 7). The paper is organized as follows: Section II gives notation, matrix concavity, and the discrete-time MCBF framework of [21] (Page 5). Section III shows that complements of polytopes, complements of spectrahedra, and unions of convex sets share a common matrix-valued representation (Page 5). Section IV presents the union characterization, the eigenvector selection constraint and its properties (Page 5). Section V compares the two filters on a double integrator with two obstacles (Page 5). Section VI concludes the paper (Page 5). The authors also thank and acknowledge AI tools, including ChatGPT and Claude, for reviewing and revising the manuscript for technical and grammatical errors (Page 7). The paper is organized as follows: Section II gives notation, matrix concavity, and the discrete-time MCBF framework of [21] (Page 5). Section III shows that complements of polytopes, complements of spectrahedra, and unions of convex sets share a common matrix-valued representation (Page 5). Section IV presents the union characterization, the eigenvector selection constraint and its properties (Page 5). Section V compares the two filters on a double integrator with two obstacles (Page 5). Section VI concludes the paper (Page 5). The authors also thank and acknowledge AI tools, including ChatGPT and Claude, for reviewing and revising the manuscript for technical and grammatical errors (Page 7). The paper is organized as follows: Section II gives notation, matrix concavity, and the discrete-time MCBF framework of [21] (Page 5). Section III shows that complements of polytopes, complements of spectrahedra, and unions of convex sets share a common matrix-valued representation (Page 5). Section IV presents the union characterization, the eigenvector selection constraint and its properties (Page 5). Section V compares the two filters on a double integrator with two obstacles (Page 5). Section VI concludes the paper (Page 5). The authors also thank and acknowledge AI tools, including ChatGPT and Claude, for reviewing and revising the manuscript for technical and grammatical errors (Page 7). The paper is organized as follows: Section II gives notation, matrix concavity, and the discrete-time MCBF framework of [21] (Page 5). Section III shows that complements of polytopes, complements of spectrahedra, and unions of convex sets share a common matrix-valued representation (Page 5). Section IV presents the union characterization, the eigenvector selection constraint and its properties (Page 5). Section V compares the two filters on a double integrator with two obstacles (Page 5). Section VI concludes the paper (Page 5).

Improvements for AI systems

Bold header: Improved Constraint Generation for Nonconvex Safe Sets

The system can now generate a convex input constraint that acts only on the eigenvalues that determine membership in the set by Selecting eigenvectors of the matrix function at the current state, which simplifies non-convex safety filtering into a convex program.

Bold header: Guaranteed Geometric Lower Bounds

The improved filter ensures that Every eigenvalue that determines membership in the set is shown to satisfy a geometric lower bound along closed-loop trajectories, specifically by proving that along any such trajectory, for all i ∈ [j,..., p] and all k ≥ 0, λi(H(xk)) ≥ (1 − γ)λi(H(x0)).

Bold header: Faster Real-Time Safety Execution

The proposed filter can be implemented as a quadratic program with one linear constraint for the polytope case, achieving a runtime of 1.9001 · 10−3 s, which is more than an order of magnitude faster than the full-matrix semidefinite program runtime.

Bold header: Robust Handling of Complex Obstacles

The system can safely navigate environments defined by obstacles that are the complement of a polytope, the complement of a spectrahedron, and unions of convex sets, ensuring forward invariance even when the resulting discrete-time constraint is nonconvex in the input.

Abstract

The discrete-time control barrier function condition for a safe set that is a union of convex sets is nonconvex in the control input. For safe sets defined by a matrix concave function through the number of its nonnegative eigenvalues, this paper shows that the set is a union of convex sets, and that the nonconvexity arises because at least one member of the union must hold, not every member. Selecting eigenvectors of the matrix function at the current state yields a convex input constraint that acts only on the eigenvalues determining membership in the set. This constraint is implied by the matrix-wide condition of prior work, and it guarantees a geometric lower bound on each of those eigenvalues along closed-loop trajectories. A double-integrator simulation with a polytope obstacle and a spectrahedron obstacle compares the proposed filter with the matrix-wide condition.