Unified Overview of Safety Certificate Functions: Techniques from Hamilton--Jacobi Reachability and Control Barrier Functions

arXiv:2611.07537 · cs.RO, cs.SY, eess.SY · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

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

Rosa: We're looking at the paper: "Unified Overview of Safety Certificate Functions".

Dev: Here's the gist of the paper. Safety certificates are functions whose existence guarantees that a system meets its safety specification, with synthesis methods largely determined by whether a closed-form dynamics model is available and the state dimension.

Rosa: For low-dimensional systems with closed-form dynamics, classical methods apply; numerical Hamilton–Jacobi (HJ) reachability recovers approximations to the maximal safe set, while control barrier functions (CBFs) allow for analytical constructions under special dynamic structures at negligible computational cost.

Dev: When state dimension is high, learning-based methods are preferred due to reduced manual tuning requirements.

Rosa: Classical synthesis methods include: Analytical explicit methods like High-order CBFs (HOCBFs) and CBF backstepping; optimization-based methods such as Sum-of-Squares (SOS) optimization for polynomial dynamics; and implicit methods like Backup CBFs, which construct an implicit safe set via forward rollout under a known backup controller.

Paper summary: Taro: Thinking about the authors, Hao Wang and Max Cohen and Zeyuan Feng and Donggeon David Oh and Oswin So Songyuan Zhang Javier Borquez Chuchu Fan Jaime Fernandez Fisac Haimin Hu Jason J Choi Somil Bansal, it seems like they brought together a lot of different expertise here.

Rosa: Yeah, they really covered the whole spectrum from field robotics to control engineering to autonomy research, which is why this overview feels so comprehensive for the general audience interested in safety certificates.

Dev: So what does this paper actually mean for someone who isn't necessarily deep into the math right now? It means that instead of starting from scratch and trying to figure out if you should use HJ reachability or a CBF, there’s this unified view that shows you the options available.

Taro: I think it tells us that safety certificate synthesis is maturing, and it’s moving away from just having a few specific techniques to having a broader framework where you can choose based on the system's size and your computational budget.

Rosa: That’s right. It shows that we have these powerful tools like Hamilton--Jacobi reachability and control barrier functions, and this paper is the guide showing how they fit together for creating those safety guarantees for autonomous systems.

Dev: So if you take away the core message of this paper, it’s that there are now established techniques for creating safety certificates using both Hamilton--Jacobi reachability and control barrier functions, and this tutorial gives practitioners a holistic way to understand how they work.

Taro: It sets a clear direction for future work by showing where the current limitations are—like the need for adaptive sampling strategies or more accurate safety value approximations when you're dealing with really complex, high dimensional systems.

Rosa: Exactly. The paper points out that while we have these methods, we still have to keep working on making them more scalable and robust against real world uncertainty, which is where the next wave of research needs to go.

Conclusion: Rosa: So we've been looking at how different math tools—like Hamilton--Jacobi reachability and control barrier functions—can be used to build safety certificates for systems, and this paper is trying to put them all in one place under the title "Unified Overview of Safety Certificate Functions."

Dev: It basically takes these two separate families of techniques and shows how they work together, depending on whether your system is small or big.

Rosa: Right. The authors are a team from different areas, which you can see because they bring in expertise from field robotics to deep learning in the process.

Dev: They’re showing how you can pick the right tool for the job—like using something analytical when your system is simple, or switching to learning methods when things get really high dimensional.

Rosa: It means for a robot designer or a control engineer, there's a clearer map now of which approach makes sense based on the size of their problem and how much computation they can handle.

Dev: Exactly. It’s about making the choice between using traditional math methods and those new learning-based methods more organized.

Rosa: The implication for us is that we can start thinking about these certificates as a toolkit rather than just one specific technique you have to master on your own.

Dev: And it points out some of the real hurdles too, like how hard it is to verify the learned methods or dealing with those high dimensional problems where things get messy.

Rosa: It’s a lot of information packed into that overview, but it gives us a solid starting point for understanding this whole safety verification landscape.

University of Southern California · Stanford University · North Carolina State University · Princeton University · Massachusetts Institute of Technology · ÉUniversidad de Santiago de Chile

cs.RO, cs.SY, eess.SY

Submitted: 2026-11-05

Updated: 2026-11-05

Code: https://github.com/HJReachability/helperOC

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

Importance score: 82/100

The gist: Safety certificates are functions whose existence guarantees that a system meets its safety specification…

Key concepts

Hamilton–Jacobi Reachability
This method constructs safety certificates by analyzing the Hamilton–Jacobi–Bellman Variational Inequality (HJB-VI). It defines the safe set as where a specific value function, representing the best possible future cost, is non-negative. This approach is useful for low-dimensional systems.
Control Barrier Functions (CBFs)
CBFs are used to enforce safety by requiring that the system's dynamics stay within a safe region. They achieve this by ensuring a specific inequality holds for the derivative of the function defining the boundary of that safe set, allowing for analytical constructions under certain dynamic conditions.
PINN-Based HJ Reachability
This learning method uses physics-informed neural networks (PINNs) to approximate the value function. It minimizes errors between the network's predictions and known system dynamics, effectively using deep learning to solve complex reachability problems for high-dimensional systems.
Curse of Dimensionality
This limitation refers to the exponential increase in computational complexity that occurs when trying to apply methods like HJ reachability to systems with many states. It severely restricts grid-based classical methods, often limiting their practical use to low-dimensional problems.

Terminology

Summary

Safety certificates are functions whose existence guarantees that a system meets its safety specification, with synthesis methods largely determined by whether a closed-form dynamics model is available and the state dimension. p. 0

Where the paper says it

We note that the choice of synthesis method is largely determined by two properties of the problem: whether a closed-form dynamics model is available, and the dimension of the state. For low-dimensional systems with closed-form dynamics, classical methods apply.

For low-dimensional systems with closed-form dynamics, classical methods apply; numerical Hamilton–Jacobi (HJ) reachability recovers approximations to the maximal safe set, while control barrier functions (CBFs) allow for analytical constructions under special dynamic structures at negligible computational cost. p. 0
Where the paper says it

For low-dimensional systems with closed-form dynamics, classical methods apply. Numerical HJ reachability [8], [9] recovers the closest approximation to the maximal safe set, whereas CBFs [3], [10] admit analytical constructions under special structures of the dynamics [11], [12] at negligible computational cost, though control bounds complicate their design.

When state dimension is high, learning-based methods are preferred due to reduced manual tuning requirements. p. 1
Where the paper says it

II-F for the overview of the methods not covered in this tutorial. state dimension is high, learning-based methods require less manual tuning and are preferable.

Classical synthesis methods include: Analytical explicit methods like High-order CBFs (HOCBFs) and CBF backstepping; optimization-based methods such as Sum-of-Squares (SOS) optimization for polynomial dynamics; and implicit methods like Backup CBFs, which construct an implicit safe set via forward rollout under a known backup controller. p. 7

Where the paper says it

Formulations that bridge this gap, either by introducing CBF-like design flexibility into reachability formulations or by constructing CBFs that recover the maximal safe set, remain an active area of research [34], [51], [88]. Broadly, classical CBF synthesis methods can be grouped into three categories: 1) analytical explicit methods, 2) optimization-based methods, and 3) implicit methods.

HJ reachability analysis constructs certificates from the Hamilton–Jacobi–Bellman Variational Inequality (HJB-VI), where the safe set is defined by the zerosuperlevel set of the value function V s(x):= t to infinity V s(x, t) = u tau in[0, infinity) l(x(tau)), resulting in S = x in X V s(x) at least 0. p. 3
Where the paper says it

1 at a state x ∈ X and time t ∈ T is represented by the safety value function Vs(x, t) defined by Vs(x, t) := sup u min τ∈[0,t] l(x(τ )). (9) From the principle of optimality [29], we can derive that the value function Vs(x, t) is the viscosity solution [30] of the following Hamilton–Jacobi–Bellman Variational Inequality (HJB-VI) [31]: min ( − ∂Vs ∂t (x, t)+H(x, ∇Vs(x, t)), l(x)−Vs(x, t) ) = 0, Vs(x, 0) = l(x), (10) where ∇Vs(x, t) denotes the spatial gradient of Vs.

CBFs are particularly suited for enforcing forward invariance of safe sets by requiring that u in U ḣ(x, u) = u in U grad h(x), f(x, u) at least-alpha(h(x)) for x in S. p. 4
Where the paper says it

One may address this by introducing a discount factor in time into the definition of the value function [32]–[34]. Control Barrier Functions Control barrier functions (CBFs) are a class of safety certificates that are particularly amenable to the synthesis of controllers enforcing forward invariance of safe sets.

Learning-based methods encompass: PINN-Based HJ Reachability (DeepReach), which uses physics-informed neural networks to represent the value function V theta(x, t) by minimizing PDE residuals; Neural CBFs, which parameterize the CBF h theta and a controller pi phi jointly by penalizing violations of the CBF condition and deviations from a desired controller pi d; and RL-Based Reachability, which scales approximation via deep reinforcement learning algorithms solving time-discounted safety Bellman equations for black-box systems. p. 1

Where the paper says it

Neural CBFs [16]–[18] are computationally efficient but rely on a nominal controller. PINN-based reachability [19], [20] targets the exact safety value function, given continuous-time dynamics and an effective sampling scheme.

Central to these approaches is solving a dynamic programming problem, typically reachability. p. 6

Where the paper says it

Discrete-Time Hamilton–Jacobi Reachability Analysis The optimal control problem for HJ reachability in Prob. 1 can be similarly written in discrete time, whose dynamic programming principle yields the discrete-time counterpart of the HJB-VI (10) [21]: Vs(x, t) = min l(x), max u∈U Vs(fD(x, u), t + ∆t) , (27) where ∆t is the sampling period for discretization.

Overarching limitations include scalability due to the curse of dimensionality and real-world uncertainty stemming from system dynamics and environment changes. Learning methods are sensitive to sampling strategies and distribution shift, while verification schemes range from probabilistic (certifying distribution-level violation rates) to deterministic (checking CBF conditions pointwise or via forward reachable tubes), with scalable verification lagging behind synthesis development.

Specific challenges for methods are detailed: HJ reachability suffers from the curse of dimensionality, restricting grid-based methods to low dimensions (n x 6), and control-affine systems often yield bang-bang optimal controllers requiring smoothing. p. 8

Where the paper says it

The main limitation of the numerical method for HJ reachability is the curse of dimensionality: both storage and computation grow exponentially in the state dimension nx, since the number of grid nodes scales as (∆x −1 ) nx . In practice, systems with nx ≤ 6 are tractable on standard hardware, and higher dimensions are generally out of reach without additional structure [13], [95].

Classical CBFs (HOCBFs, backstepping) lack validity under hard input limits and rely on specific system structures. p. 9
Where the paper says it

For systems with strict control input bounds, the CBF condition (40) can be infeasible even when Lghr−1(x) ̸= 0. Even without hard input limits, the HOCBF condition (40) may become infeasible if there exist points within C where Lghr−1(x) = 0.

Backup CBFs require knowledge of a backup policy and safe set, introducing significant online computation. p. 5
Where the paper says it

CBFs can also be designed implicitly according to this principle, including the Backup CBF framework, which is discussed in Sec. More broadly, methods that leverage online computation to certify safety can be viewed as providing implicit safety certificates; a comprehensive overview of such approaches can be found in [74], [75].

PINN-Based HJ Reachability struggles with uniform sampling in high-dimensional, contact-rich systems due to non-physical samples. p. 1
Where the paper says it

Neural CBFs [16]–[18] are computationally efficient but rely on a nominal controller. PINN-based reachability [19], [20] targets the exact safety value function, given continuous-time dynamics and an effective sampling scheme.

Neural CBFs' resulting safe sets depend heavily on the quality of the fixed backup controller used for synthesis. RL-Based Reachability can be less data-efficient than model-based methods due to reliance on sampled transitions without explicit dynamics exploitation. p. 1
Where the paper says it

state dimension is high, learning-based methods require less manual tuning and are preferable. Neural CBFs [16]–[18] are computationally efficient but rely on a nominal controller.

Future research must focus on adaptive sampling strategies, more accurate safety value approximations with fewer interactions, and scalable formal guarantees for learned filters.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Adaptive Safety Filtering for Unknown Dynamics: The Q-CBF safety filter can be used to enforce a CBF constraint even when system dynamics fD is not available, allowing it to be applied to black-box systems with unknown dynamics by using the policy conditioned state action value function Qs(x, u).

  2. Scalable Safety Certificate Synthesis for High Dimensions: PINN-based HJ reachability methods can address the exponential scaling of numerical solvers by using a neural representation of the safety value function Vθ(x, t), which is trained via minimizing a weighted combination of PDE and boundary condition residuals.

  3. Robust Safety Policy Learning for Black-Box Systems: RL-based reachability enables scalable approximation of the safety value function by interacting with systems available only as black-box transition mechanisms, allowing it to learn safety certificates and best-effort safety policies directly from RL rollout data.

  4. Enhanced Geometric Observation for Safety Filtering: Using observation representations like 47D Frenet-based observations or 66D LiDAR measurements in conjunction with RL-based reachability allows the Q-CBF filter to achieve a 100% empirical safety rate across all seeds, even in high-fidelity racing scenarios.

  5. Verification of Learned Safety Certificates: Posthoc verification schemes can establish safety guarantees for learned certificates by treating them as black boxes and verifying safety from trajectory rollouts, using a probabilistic guarantee based on the empirical safety score J˜(x) to ensure that with probability at least 1 minus β, Px∈Sδ J˜(x) ≤ 0.

Abstract

Safety is a central concern in modern autonomous systems. Safety certificates---functions whose existence guarantees that a system meets its safety specification---have long been used as a principled tool for rigorously reasoning about safety. Methods for synthesizing such certificates are scattered across research areas, with little guidance on when to use a particular approach. This tutorial provides a holistic treatment of safety certificate synthesis centered on two widely used families: Hamilton--Jacobi (HJ) reachability analysis and control barrier functions (CBFs). For each family of methods, we first introduce its general formulation, along with classical and learning-based tools for certificate synthesis. Next, we introduce two benchmark problems consisting of vehicle systems of differing complexity to assess the certificate quality of different synthesis methods and their effectiveness in safety-critical control. We use these benchmarks to compare various methods and provide practical guidance for selecting an approach.

Sources

Related papers