Neural Certificates for Scalable Safety Verification of Dynamical Systems

This direction is merged into another direction: curse of dimensionality (state-space discretization bottleneck). Its last version stays readable and citable here.

Version 1 of 1176 papers196 conceptsfrozen 2026-10-04

When we deploy autonomous systems—like drones or self-driving cars—in the real world, "mostly safe" isn't good enough. We need formal guarantees that a robot will never crash or will always reach its goal. Traditionally, providing these guarantees required formal verification, a process of mathematically proving that a system's dynamics will never violate safety constraints.

However, traditional verification hits a wall known as the curse of dimensionality. As researchers like Federico Ramponi and John Lygeros have noted in On the connections between PCTL and dynamic programming, most classical methods rely on discretizing the state space into a grid. If you double the number of variables in your system (e.g., moving from a 2D point mass to a 4D car model), the number of grid cells doesn't just double—it grows exponentially. This makes verifying a 12-dimensional humanoid robot computationally impossible using grids.

A new research direction has emerged to solve this: replacing discrete grids with Neural Certificates. Instead of checking every point in a box, we learn a continuous neural network that acts as a mathematical "proof" of safety or stability.

The Geometry of Safety: Lyapunov and Barrier Functions

To understand why this works, we have to look at the geometry of control. In classical control theory, we prove a system is stable using a Lyapunov function—a mathematical "bowl" where the bottom is the goal state. If we can show that the system always moves "downhill" on this bowl, we have a proof of stability. Similarly, Barrier Functions act like a "fence" around unsafe regions; if the function's value increases as you approach an obstacle, and the controller ensures the value stays below a certain threshold, the system is guaranteed to stay safe.

The breakthrough, as detailed by Felix Berkenkamp and Andreas Krause in The Lyapunov Neural Network, is to use a neural network to represent these functions. Because neural networks are universal function approximators, they can "mold" themselves to the complex, high-dimensional shapes of safe regions that a simple quadratic function or a rigid grid could never capture.

From Deterministic Fences to Stochastic Supermartingales

Real-world systems are rarely deterministic; they are buffeted by wind, sensor noise, and unpredictable terrain. In these stochastic systems, a simple "fence" isn't enough because noise could theoretically push the system over any finite boundary.

To handle this, researchers like Đorđe Žikelić and Mathias Lechner have introduced Neural Ranking Supermartingales (RSMs) and Reach-Avoid Supermartingales (RASMs). As described in Learning Control Policies for Stochastic Systems with Reach-avoid Guarantees, these certificates provide a probabilistic guarantee. Instead of saying "the robot will never crash," they allow us to calculate a formal lower bound—for example, "the robot will reach the goal while avoiding obstacles with a probability of at least 99.9%."

The intuition here is that the Supermartingale acts as a "stochastic compass." Even if noise pushes the robot off course, the certificate ensures that, on average, the system is always moving toward safety or the goal.

Making the Proof Rigorous

A neural network is just a collection of weights; how do we know it actually constitutes a "proof"? The research direction focuses on three main ways to make these certificates rigorous:

  1. Architectural Constraints: By hard-coding mathematical properties into the network itself. For instance, Prashanth Krishnamurthy and Farshad Khorrami use Lipschitz Neural Networks to guarantee the network isn't too "wiggly," which allows for formal bounds on error.
  2. Convex Relaxations: Rayan Mazouz and Luca Laurenti show in Safety Guarantees for Neural Network Dynamic Systems via Stochastic Barrier Functions how to use optimization techniques to verify that the neural barrier function satisfies the required safety conditions across the entire state space.
  3. SMT Solvers and SOS Optimization: Methods like those proposed by Benoit Landry and Marco Pavone in Reach-Avoid Problems via Sum-of-Squares Optimization and Dynamic Programming use powerful mathematical solvers to "search" for a certificate that satisfies the safety equations, bypassing the need for grids entirely.

Why This Matters

By moving from grids to neural certificates, we shift the burden of verification from memory (storing a massive grid) to optimization (training a network). This allows us to scale safety proofs to the high-dimensional systems required for modern robotics. As Claire Tomlin demonstrates in Classification-based Approximate Reachability, replacing a grid with a neural classifier can enable real-time safety guarantees for systems that were previously considered "unverifiable."

The end goal of this direction is a "correct-by-construction" pipeline: we specify what we want the robot to do, and the system automatically learns both a control policy and a neural certificate that proves the policy is safe.

Papers in this version

Showing 60 of 176 papers.

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_neural_certificates_for_scalable_safety__49a0c86b,
  title = {Neural Certificates for Scalable Safety Verification of Dynamical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-certificates-for-scalable-safety-verification-of-dynamical-systems/rdr_49a0c86b2a284bbdb69e99c81ad88916},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_neural_certificates_for_scalable_safety__49a0c86b_v1,
  title = {Neural Certificates for Scalable Safety Verification of Dynamical Systems},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neural-certificates-for-scalable-safety-verification-of-dynamical-systems/rdr_49a0c86b2a284bbdb69e99c81ad88916/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}