⚡ electrical engineering

Safety Filtering with an Infinite Number of Constraints

This paper extends Control Barrier Function (CBF) theory to handle an infinite number of safety constraints by establishing necessary and sufficient conditions for forward invariance, defining regularity requirements for continuous controllers, and demonstrating how these theoretical advancements resolve limitations in existing backup CBF approaches.

Original authors: Max H. Cohen, Pio Ong, Pol Mestres, Aaron D. Ames

Published 2026-04-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Max H. Cohen, Pio Ong, Pol Mestres, Aaron D. Ames

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are driving a self-driving car. Your most important job is to keep the car safe: don't hit the curb, don't crash into other cars, and don't drive off a cliff.

In the world of robotics and AI, mathematicians use a tool called a Control Barrier Function (CBF). Think of a CBF as an invisible, magical forcefield around your car. If the car tries to cross the forcefield, the computer instantly slams the brakes or steers away to keep it inside the safe zone.

The Problem: The "Infinite" Wall

For a long time, this magic forcefield was easy to draw. If you just needed to stay inside a single box (like a parking spot), you only needed one rule: "Stay inside the box."

But real life is messy. Sometimes, the "safe zone" isn't a simple box. It's a complex shape defined by infinite rules.

Imagine trying to describe a safe path for a robot arm. You can't just say "stay in this box." You have to say:

"Don't hit the wall at 1 second."
"Don't hit the wall at 1.0001 seconds."
"Don't hit the wall at 1.0002 seconds."
...and so on, forever.

This is what the paper calls an infinite number of constraints. It's like trying to build a fence using an infinite number of tiny, invisible fence posts.

The big question the authors asked:
If you have an infinite number of rules, how do you:

Know for sure the car will never crash?
Write a computer program that can actually calculate the steering wheel angle without getting stuck or freezing?
Make sure the car's movements are smooth, not jerky?

The Solution: The "Smart Fence"

The authors of this paper (Cohen, Ong, Mestres, and Ames) figured out how to handle this infinite mess. Here is how they did it, using simple analogies:

1. The "Active" Fence Posts (Nagumo's Theorem)

Imagine you are walking along a cliff edge. You don't need to worry about the entire cliff; you only need to worry about the specific spot where your foot is touching the edge.

The paper proves that even if there are infinite rules, you only need to check the active rules—the ones that are currently "touching" the boundary of safety. If you can prove the car won't fall off right now at the edge, the math guarantees it won't fall off later. They found a way to turn these infinite rules into a manageable list of "active" rules.

2. The "Optimal Decay" Brake (The OD-CBF)

In the old way of doing things, the computer had to guess how hard to brake to stay safe. Sometimes it guessed too hard, sometimes too soft, and sometimes it couldn't guess at all, causing the car to freeze (a "crash" in the software).

The authors introduced a new concept called Optimal-Decay CBF.

Old Way: "You must slow down by exactly 5 mph." (Rigid, might break the rules).
New Way: "Slow down just enough to stay safe, but not more than necessary."

They added a special "dial" (a variable) to the computer's brain. This dial automatically adjusts the braking force. If the car is far from the danger, it brakes gently. If it's about to hit the wall, it brakes hard. This "dial" ensures the computer never gets stuck trying to find a solution; it always finds a smooth, safe path.

3. From Infinite to Finite (The Sampling Trick)

You can't ask a computer to check an infinite number of rules every millisecond. It would take forever.

The paper shows a clever trick: Sampling.
Imagine you want to check if a rope is strong enough. You don't need to test every single fiber of the rope. You just need to test a few specific, evenly spaced points. If those points hold, the whole rope holds.

The authors proved that if you pick a finite number of "checkpoints" (like checking the fence every 10 feet instead of every inch), and your car passes those checkpoints, it is mathematically guaranteed to pass the infinite number of rules in between. This allows the computer to run fast and smooth.

Real-World Example: The Backup Plan

The paper uses a "Backup CBF" as a test case. Imagine a robot that is supposed to stay in a room.

The Problem: The robot might get pushed. If it gets pushed too hard, it might fall out of the room.
The Backup: The robot has a "panic button" (a backup controller) that knows how to get back to a tiny, super-safe corner of the room if things go wrong.
The Infinite Rule: The "safe zone" is actually the path the robot would take if it hit the panic button. This path changes every fraction of a second, creating infinite rules.

Using their new method, the robot can now calculate a smooth, safe path that respects this infinite "panic path" without freezing up. The paper shows a simulation where the old method caused the robot to glitch and stop, but their new method kept the robot moving smoothly and safely.

The Bottom Line

This paper is like upgrading the safety system of a self-driving car from a rigid, broken checklist to a flexible, intelligent guardian.

Before: "Here are 1,000,000 rules. If you break one, you crash. Good luck calculating that!"
After: "Here is a smart system that only checks the rules that matter right now, adjusts its braking automatically, and uses a few smart samples to guarantee safety for the whole infinite future."

It turns a theoretical nightmare (infinite constraints) into a practical, working tool for keeping our future robots and cars safe.

1. Problem Statement

Control Barrier Functions (CBFs) are a standard framework for ensuring safety in autonomous systems by enforcing forward invariance of a safe set $S$ . While CBF theory is well-established for systems with a finite number of constraints (e.g., $h_i(x) \geq 0$ for $i=1 \dots N$ ), many critical applications require infinite constraints.

Key Examples:
- Backup CBFs: Expanding a conservative safe set using the system's flow under a backup controller over a time horizon.
- Collision Avoidance: Ensuring no collision occurs for any configuration of rigid bodies.
- Uncertainty: Enforcing safety where the safe set is defined by all possible state estimates.
The Gap: Existing CBF theory struggles with infinite constraints. Fundamental questions remain unanswered regarding:
1. Necessary and sufficient conditions for forward invariance when the safe set is defined by infinitely many inequalities.
2. The definition of CBFs and the regularity (continuity/Lipschitz properties) of controllers derived from infinite constraints.
3. Whether infinite constraints can be reduced to a finite set for practical implementation without losing safety guarantees.

2. Methodology

The authors extend CBF theory to handle safe sets defined as:
$S = \{x \in \mathbb{R}^n \mid h(\tau, x) \geq 0, \forall \tau \in T\}$
where $T \subset \mathbb{R}^p$ is a compact parameter set (e.g., time horizon) and $h$ is continuously differentiable in $x$ and continuous in $\tau$ .

Key Theoretical Steps:

Regularity and Well-Posedness: The paper introduces Assumption 1, a generalization of the "regular value" assumption. It requires that for any point on the boundary of $S$ , there exists a direction $v$ such that the Lie derivative of $h$ is strictly positive for all active constraints (where $h(\tau, x)=0$ ).
Smooth Ascent Direction (Lemma 1): Using the compactness of $T$ and Assumption 1, the authors prove the existence of a smooth ( $C^\infty$ ) vector field $\phi(x)$ that acts as a "strict ascent" direction for all active constraints simultaneously. This bridges the gap between pointwise existence and the existence of a smooth function required for stability proofs.
Invariance Conditions:
- Theorem 1: Extends Nagumo's Theorem to infinite constraints. It proves that $S$ is forward invariant if and only if the Lie derivative condition holds for all active constraints on the boundary.
- Optimal-Decay Barrier Functions (OD-BF): To close the gap between necessary/sufficient conditions and practical controller synthesis, the authors introduce OD-BFs. They prove that the existence of an OD-BF is necessary and sufficient for forward invariance under the stated regularity conditions.
Controller Synthesis:
- OD-CBF Definition: A Control Barrier Function tailored for infinite constraints using a single "Optimal Decay" (OD) variable $\omega$ . This avoids the need for infinite decision variables in optimization.
- Continuity (Proposition 4): The paper proves that optimization-based controllers (solving a Quadratic Program with infinite constraints) are continuous (though not necessarily Lipschitz) under strict feasibility conditions.
- Safety under Continuity (Theorem 2): Since continuous controllers do not guarantee unique solutions (Lipschitz continuity is usually required for uniqueness), the authors provide a weaker safety guarantee using Dini derivatives and Danskin's Theorem, ensuring invariance even for continuous (non-Lipschitz) closed-loop dynamics.
Finite Reduction (Proposition 5): The paper provides a constructive method to reduce the infinite constraint set to a finite set. By discretizing the parameter space $T$ with a sufficiently fine grid (based on Lipschitz constants), satisfying the finite set implies satisfaction of the infinite set, enabling the use of standard numerical solvers.

3. Key Contributions

Theoretical Extension: Formalized CBF theory for safe sets defined by an infinite number of continuously differentiable constraints.
Necessary and Sufficient Conditions: Established that forward invariance is equivalent to barrier-like inequalities on the boundary (generalizing Nagumo's Theorem) and linked this to the existence of Optimal-Decay Barrier Functions.
Regularity Guarantees: Proved that optimization-based controllers satisfying infinite constraints are continuous, and established safety guarantees for systems with continuous (non-Lipschitz) vector fields.
Practical Reduction: Developed a rigorous procedure to discretize infinite constraints into a finite set while preserving safety guarantees, making the theory computationally tractable.
Unified Framework for Backup CBFs: Demonstrated how the proposed framework resolves limitations in existing Backup CBF methods, specifically regarding the feasibility of safety filters.

4. Results

Theoretical Validation: The paper proves that under Assumption 1 (well-posedness), the infinite constraint set behaves similarly to finite sets regarding invariance, provided the controller satisfies specific barrier conditions.
Feasibility of OD-CBFs: The introduction of a single OD variable $\omega$ allows the formulation of a semi-infinite optimization problem (Eq. 29) that is strictly feasible near the boundary, ensuring the controller does not fail due to infeasibility.
Numerical Example (Double Integrator):
- The authors applied the theory to a double integrator with a backup CBF setup.
- Standard Backup CBF: Using fixed decay rates ( $\alpha$ ) led to infeasibility at certain states within the safe set, causing the controller to be undefined and the system to potentially leave the safe region.
- Proposed OD-CBF: Using the single OD variable approach, the controller remained feasible across the entire safe set. The resulting closed-loop vector field was well-defined everywhere, and the system successfully maintained safety.

5. Significance

This paper addresses a critical bottleneck in safety-critical control: the gap between theoretical safety guarantees and practical implementation for complex, infinite-horizon, or uncertainty-based safety constraints.

For Backup CBFs: It solves the "infeasibility" problem often encountered when trying to enforce safety over a time horizon with fixed decay rates, offering a robust alternative that guarantees the existence of a valid control input.
For General Safety: It provides the mathematical foundation for using CBFs in scenarios where the safe set is implicitly defined (e.g., via reachability or robust optimization), moving beyond simple geometric constraints.
Implementation: By proving that infinite constraints can be reduced to finite ones with guaranteed safety, the work enables the direct application of these advanced safety filters using standard optimization solvers (like QP) in real-time control systems.