From Global to Local: Hierarchical Probabilistic Verification for Reachability Learning

summary

The gist

The gist The authors propose a hierarchical probabilistic verification framework for reachability learning that bridges offline global certification and online local refinement to provide scalable

In short

The authors developed a hierarchical probabilistic verification framework to improve reachability learning safety in high-dimensional systems. They combine an efficient global certification step using scenario optimization with an online local refinement module. This structure allows the system to provide scalable safety guarantees while focusing computation on critical state regions, leading to better performance in drone racing simulations.

Key concepts

Hierarchical Verification Framework: A multi-tier control logic structure that manages safety. It moves from a fast target maintenance mode (Tier 0) to global verification using a learned policy (Tier 1), local refinement in critical areas (Tier 2), and finally, safe recovery if certification fails (Tier 3).
Scenario Optimization: A method used for efficient safe set expansion. Instead of complex chance constraints, it samples deterministic scenarios to find the minimum required constraints needed to guarantee a probabilistic safety bound on trajectory deviations.
Local Refinement Mode: A mechanism that recovers regions deemed unsafe by the global verification but potentially safe locally. It uses scenario optimization to determine a high-resolution local radius, focusing computation only on the boundary where refinement is necessary.

Terminology used across episodes

Reinforcement Learning (RL) — explained in 25 episodes
Sample Complexity — explained in 18 episodes
Federated Learning — explained in 17 episodes
Retrieval-Augmented Generation (RAG) — explained in 16 episodes
Interpretability — explained in 15 episodes
Quantum Fisher Information (QFI) — explained in 15 episodes

This episode discusses

The paper

From Global to Local: Hierarchical Probabilistic Verification for Reachability Learning · Read on arXiv

Ebonye Smith, Sampada Deglurkar, Jingqi Li, Gechen Qu, Claire J. Tomlin

University of California Berkeley · Oden Institute of Computational Engineering and Sciences

Hamilton-Jacobi (HJ) reachability provides formal safety guarantees for nonlinear systems. However, it becomes computationally intractable in high-dimensional settings, motivating learning-based approximations that may introduce unsafe errors or overly optimistic safe sets. In this work, we propose a hierarchical probabilistic verification framework for reachability learning that bridges offline global certification and online local refinement. We first construct a coarse, conservative under-approximation of the reach-avoid safe set using scenario optimization, providing an efficient global probabilistic certificate. We then introduce an online local refinement module that is structurally anchored to the global certificate and recovers regions excluded by the global verification. This refinement reduces conservatism while focusing computation on critical regions of the state space. Integrated with a condition-based switching mechanism between a learned reachability policy and a model-based controller, the proposed framework improves success rates in goal-reaching tasks with safety constraints, as demonstrated in drone-racing simulation experiments.

DOI: 10.1109/LCSYS.2026.3717040

Transcript

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

Rosa: Today's paper: "From Global to Local".

Dev: The gist The authors propose a hierarchical probabilistic verification framework for reachability learning that bridges offline global certification and online local refinement to provide scalable safety guarantees in high-dimensional systems.

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

Paper summary: Rosa: We started by looking at this paper "From Global to Local: Hierarchical Probabilistic Verification for Reachability Learning" and the core idea is that standard Hamilton-Jacobi reachability is too slow for complex systems, so they propose a hierarchical probabilistic verification framework.

Dev: It claims they bridge offline global certification with online local refinement, starting with a coarse safe set using scenario optimization to get an efficient global probabilistic certificate.

Rosa: Then they introduce an online local refinement module that’s structurally anchored to that global certificate and expands the certified safe set near its boundary by solving a sequence of convex programs.

Dev: This refinement is what recovers regions excluded by the global verification, which reduces conservatism while focusing computation on the critical parts of the state space.

Taro: So, they start big with a conservative check and then use local methods to get more detail where it matters most for real-world performance.

Rosa: They provide probabilistic safety guarantees for both these sets; Proposition one shows that if a state is in V global, you get a probabilistic guarantee on staying safe under certain conditions <ref:2603.24990#pg1,provide probabilistic safety guarantees for both>.

Dev: And if the state is outside V global but inside V local, Proposition two provides the local certificate guarantee, meaning safety is maintained even when the global check doesn't cover that spot <ref:2603.24990#pg1>.

Taro: It’s like having a system that gives you a general idea of what's safe across the whole map and then uses a high-resolution tool to check every single corner of that general area.

Rosa: They also discuss probabilistic dynamics bounding for the global verification where they estimate a trajectory deviation quantity x* t using samples to alleviate conservatism.

Dev: They collect N i.i.d. scenario pairs and solve x* t = min x t in R s.t. P pair (x t - t squared > x t) at most epsilon <ref:2603.24990#pg1>.

Taro: That's them using samples to quantify how much the actual trajectory might drift away from the nominal one, which is a good way to handle uncertainty in the dynamics.

Rosa: They also have this local growth iterative refinement which is a quickly computed, high-resolution local refinement on the boundary of V global by solving a scenario program to find the maximum verifiable radius r where the learned policy keeps the trajectory safe.

Dev: And they use Theorem one again to guarantee that this local growth refinement provides a probabilistic safety guarantee for V local, meaning those recovered regions are safe under local refinement <ref:2603.24990#pg1>.

Taro: They’re not just stopping at the boundary; they’re actively trying to reclaim unsafe areas by looking locally and solving a program to see how big that safe region can actually be.

Rosa: This whole paper, "From Global to Local: Hierarchical Probabilistic Verification for Reachability Learning," is about using this hierarchical structure to provide formal safety guarantees in high-dimensional systems.

Dev: The authors integrate a switching mechanism between a learned reachability policy and a model-based controller, which activates the model-based controller when the learned policy might be unreliable outside the certified set.

Conclusion: Rosa: So, looking at "From Global to Local: Hierarchical Probabilistic Verification for Reachability Learning" and its authors Ebonye Smith, Sampada Deglurkar, Jingqi Li, Gechen Qu, and Claire J. Tomlin. It’s about formalizing reachability learning with a multi-tier approach to safety.

Dev: The main implication is that you can handle high-dimensional systems by breaking the verification into global certification and local refinement rather than trying to solve one massive problem all at once.

Rosa: They successfully show that combining scenario optimization for a coarse global certificate with local convex refinement reduces conservatism while focusing computation on the critical regions of the state space, leading to probabilistic safety guarantees.

Dev: The framework is particularly interesting because it provides theoretical guarantees for both globally and locally verified sets, ensuring safety under local refinement.

Taro: It means that for autonomous systems, we can have a system that knows when it’s in a risky zone and can switch control strategies gracefully to a recovery mode if things go wrong.

Rosa: Ultimately, this work demonstrates that hierarchical verification is a way to make reachability learning practical by balancing the need for formal safety with the computational limits of complex systems.

More episodes

← Home