On the Degree of Safety: Beyond Safe or Unsafe with Control Barrier Functions

summary

The gist

A valid control barrier function (CBF) certifies if its represented safe set can be rendered forward invariant, and the sign of its value indicates whether a state is safe or not…

In short

Control Barrier Functions (CBFs) certify if a safe set is forward invariant and indicate safety via their value's sign, but they don't quantify a true degree of safety. The authors show that CBF values and gradients can be changed arbitrarily while maintaining the same set, meaning they are representation-dependent. They introduce the Invariance Authority Demand (IAD) as a new, representation-independent measure of safety based on required control authority.

Key concepts

Control Barrier Function (CBF)
A mathematical function used to certify if a system's state remains within a safe set by ensuring that the system's dynamics always push it back toward the safe region. Its value tells you if the current state is safe or unsafe, but not how 'close' it is to being unsafe.
Intrinsic Infeasibility
A type of problem where a candidate CBF fails because the underlying safe set itself cannot be made controlled invariant, regardless of how the mathematical function is represented. This means the system dynamics fundamentally prevent safety from being guaranteed.
Representational Infeasibility
A problem where a candidate CBF fails because its specific mathematical representation (the chosen function $h(x)$ and its gradient $ abla h(x)$) is not compatible with the system's constraints, even if the safe set is controllable. This can be fixed by changing the function representation.
Invariance Authority Demand (IAD)
A new measure of safety that quantifies exactly how much control authority (actuator effort) is needed to keep the system within the safe set. A lower IAD value means less control effort is needed, signifying a higher degree of safety.

Terminology used across episodes

This episode discusses

The paper

On the Degree of Safety: Beyond Safe or Unsafe with Control Barrier Functions · Read on arXiv

University of California, Irvine · University of North Carolina at Chapel Hill

Transcript

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

Rosa: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "On the Degree of Safety".

Dev: Here's the gist of the paper. A valid control barrier function (CBF) certifies if its represented safe set can be rendered forward invariant, and the sign of its value indicates whether a state is safe or not, but it does not quantify a degree of safety beyond the binary indication The main contributions of this paper are as follows.

Rosa: Limitations of CBF Metrics The authors demonstrate that CBF values and gradients cannot by themselves quantify a representation-independent degree of safety because they can be arbitrarily modified at any finite number of interior states while still representing the same set Theorem III.1 shows that if we have a valid CBF representing the safe set C, we can construct another valid CBF representing the same set C by arbitrarily manipulating the CBF values and gradients at any finite collection of interior states This implies that quantities such as h˙ and h˙ + α(h) remain representation-dependent in the interior despite incorporating the system dynamics Corollary III.1 shows that if a measure of the degree of safety is required to be independent of the representation of the safe set, then h(x) and ∇h(x), ∀x ∈ Int(C), cannot provide any nontrivial information to such measure.

Dev: Diagnosis of Infeasibility The paper introduces a distinction between intrinsic and representational infeasibility based on the feasibility of a candidate CBF-OP Definition III.1 defines intrinsic infeasibility as occurring when JC < 0, and representational infeasibility occurs when JC ≥ 0 and Jh,α < 0.

Rosa: Proposition III.1 provides a two-stage diagnosis: the safe set should first be tested for intrinsic feasibility, and representational feasibility should be assessed only when the set is controlled invariant.

Paper summary: Rosa: So we're looking at this paper, "On the Degree of Safety: Beyond Safe or Unsafe with Control Barrier Functions". Essentially, they take a Control Barrier Function, or CBF. They use it to check if a set is safe—meaning if the system stays within that set. But the paper says that just having a valid CBF doesn't tell you much about *how* safe it actually is beyond just being on or off.

Dev: Right, so a valid CBF certifies forward invariance for the represented safe set, and its sign tells you if you're in it or not—it’s binary at best. But the paper argues that this binary indication isn't enough to quantify a real degree of safety beyond that simple yes or no.

Taro: So what they claim is that you can change the interior values and the gradients of these CBFs however you want, and they still represent the same safe set.

Rosa: That's what they show in Theorem III.one where it says among valid CBFs representing the same set C, interior values and gradients can be changed arbitrarily <ref:2609.03319#pg1,interior values and gradients can be changed arbitrarily>. This means neither those quantities—like h-dot or h-dot plus alpha of h—can determine a degree of safety that is independent of how you chose to represent that set.

Dev: So if you want a measure of safety that doesn't depend on the specific math you pick for the representation, then both the function itself and its gradient don't give you any extra information about how safe it is in those interior points.

Taro: And they also point out that just checking if a candidate CBF-based inequality constraint is feasible doesn't quantify safety on its own. Feasibility can fail for two reasons: either the actual safe set isn't controlled invariant, or the specific way you wrote the CBF fails to represent it correctly.

Rosa: So they distinguish between intrinsic and representational infeasibility there. Intrinsic infeasibility happens when JC is less than zero, while representational infeasibility happens when JC is greater than or equal to zero and Jh-alpha is less than zero.

Dev: That distinction helps diagnose the problem in two stages, according to Proposition III.one <ref:2609.03319#pg1>. First, you test for intrinsic feasibility of the safe set C itself. Then, you only check for representational feasibility if the set is already known to be controlled invariant—that's a necessary condition before worrying about how you wrote your CBF.

Taro: So it sounds like the system first needs to be checked for fundamental safety, and only then do we look at whether our specific mathematical formulation of that safety is working correctly.

Rosa: That brings us to their main contribution: the invariance authority demand, or IAD. They introduce this as a representation-independent degree of safety. It quantifies the control authority you actually need to maintain that controlled invariance.

Dev: The IAD, denoted as gamma C of x, is defined pointwise on the boundary of the safe set C and then summarized by setwise version Gamma C over the whole boundary. They show that if Gamma C is less than or equal to one, meaning less required control authority, then the safe set is controlled invariant.

Taro: So a smaller Gamma C means you need less control effort to keep things safe, which they frame as a higher degree of safety. This sounds useful for real-world systems where you might be constrained by actuator limits.

Rosa: Exactly. They use this IAD to suggest repair strategies when problems happen, depending on what kind of problem it is. If the set C needs to stay the same, you need more control authority, and Gamma C tells you exactly how much scaling of your control input U is needed to fix it.

Dev: Or if the actuator itself can't change, then you have to repair the safe set C—the representation itself—instead of changing the control input. They give an example where changing alpha or h can fix representational infeasibility without touching C or U.

Taro: It sounds like this IAD moves us past just saying "safe" or "unsafe" and gives us a concrete number about how much control authority is at stake right now, which can guide actuator redesigns.

Rosa: So the main point of this paper is that we need a new way to measure safety that isn't dependent on how we choose to write our mathematical model of the safe set. It suggests that CBF values and gradients alone aren't enough to give you a representation-independent safety degree.

Dev: And they provide this IAD as an example of that, showing it quantifies the control authority required for controlled invariance in a way that doesn't care about the specific math used to define the boundary C.

Taro: It puts real constraints on how we think about these functions in autonomous systems, because it shows that if you want a reliable safety measure, you have to look at the underlying control authority demand.

Rosa: So for anyone designing a system with safety constraints, this paper suggests looking beyond just the sign of the function; you need to consider how much effort it takes to keep things safe when things get tricky.

Conclusion: Rosa: So we're wrapping up this look at "On the Degree of Safety: Beyond Safe or Unsafe with Control Barrier Functions". The main thing here is that just checking if a system stays safe isn't enough to know how safe it really is.

Dev: Right, so you can have a valid Control Barrier Function, or CBF, and it tells you if a set is forward invariant and whether the state is in or out of it. But this paper shows that those numbers—the values and the gradients—can be changed in ways that still represent the exact same safe set.

Taro: So, if you change those mathematical representations inside the boundary, they can't tell you a real degree of safety that doesn't depend on how you wrote it down.

Rosa: Exactly. They show that manipulating those interior values and gradients arbitrarily means neither the function itself nor its gradient can give you any information about a safety level independent of your representation. It stops being just binary.

Dev: And they also point out that feasibility checks on those CBF constraints don't quantify safety on their own because you can fail feasibility for two reasons, not just one.

Taro: Intrinsic infeasibility and representational infeasibility are different things, Rosa. One is about whether the actual safe set is controlled invariant at all, while the other is about whether your specific CBF formulation is working correctly even if the set *is* invariant.

Rosa: That distinction helps you diagnose problems in two stages, first checking intrinsic safety then checking representational feasibility only when you know it's already controlled invariant.

Dev: The paper introduces this idea of an invariance authority demand, or IAD, which they call gamma C of x. This is a way to measure the control authority actually needed to maintain that controlled invariance.

Taro: So if you have a system boundary C and you calculate this IAD on the boundary, it tells you how much effort your actuators need to put in right now.

Rosa: If gamma C is less than one, it means less control authority is required, which they define as a higher degree of safety. It’s a representation-independent way to quantify safety based on the physical demands.

Dev: This moves beyond just saying safe or unsafe and gives you a concrete number about the control effort at stake, which can guide how you redesign actuators or fix the set itself.

Taro: It gives us a tool to decide whether we need more control power or if we need to change what we consider safe in the first place.

Rosa: So while CBF values and gradients alone aren't enough, this IAD is one representation-independent example of a degree of safety measure that quantifies the required control authority. Next up, they show how this IAD helps guide repair strategies when problems happen.

More episodes

← Home