Formal Safety Boundaries via Hamilton-Jacobi Value Functions

27 concepts26 papers1 proposal

When deploying autonomous systems in the real world, ensuring safety is paramount. Traditional safety-critical control often relies on finding a Control Barrier Function (CBF)—a mathematical certificate that proves a system will remain within a designated safe set. However, synthesizing a valid CBF for complex, nonlinear systems is notoriously difficult. If a barrier function is chosen poorly or fails to account for the system's underlying physics, the controller may inadvertently steer the system into a state where a collision becomes physically unavoidable.

To solve this, researchers use Hamilton-Jacobi (HJ) reachability analysis. Instead of guessing a barrier function, HJ reachability computes the exact boundaries of safety by analyzing the system's dynamics under worst-case conditions. The result is a mathematically rigorous framework that identifies precisely when and how to intervene to guarantee safety.

The Mechanics of Inevitable Failure

The core intuition behind HJ reachability is the concept of the Backward Reachable Tube (BRT). As Xinrui Wang et al. explain in Infusing Reachability-Based Safety into Planning and Control for Multi-agent Interactions, the BRT is the set of all states from which a system will inevitably propagate into an unsafe region (such as an obstacle or a collision), even if the controller executes the best possible evasive maneuvers.

This boundary is fundamentally conditioned on the system's physical capabilities. For example, a heavy truck requires a much larger distance to stop than a nimble sedan; thus, its BRT is significantly larger. In fact, Gokul Swamy et al. in Your Learned Constraint is Secretly a Backward Reachable Tube demonstrate that when we attempt to learn safety constraints from human demonstrations via Inverse Constraint Learning, we are not actually recovering the static locations of physical obstacles. Instead, the authors show that we are implicitly reconstructing this dynamics-conditioned BRT.

   [ Safe Region ]  -->  [ Boundary / Level Set ]  -->  [ Backward Reachable Tube ]  -->  [ Obstacle ]
   (Control wins)        (Value Function = 0)           (Inevitable Failure)

Why It Works: The Differential Game

To compute this boundary, HJ reachability frames safety as a zero-sum differential game between two players:

  1. The Controller ($u$): Tries to keep the system safe.
  2. The Disturbance ($d$): Acts adversarially to force the system into failure.

By formulating this interaction, Fang‐Chieh Chou et al. in Reachability Analysis for FollowerStopper: Safety Analysis and Experimental Results show how solving the Hamilton-Jacobi-Isaacs (HJI) partial differential equation yields a value function over the state space. The value at any given state represents the margin of safety under worst-case adversarial actions. This worst-case assumption allows the framework to handle severe real-world uncertainties, such as time-varying communication delays, which Shankar A. Deka et al. successfully model as an adversarial player in Towards cyber-physical systems robust to communication delays: A differential game approach.

The boundary of the safe set is defined by a specific level set (or threshold) of this value function. Pierre-François Massiani et al. provide the theoretical foundation for this in Safe Value Functions, proving that a strict mathematical gap exists at the boundary of the viability kernel. This gap allows practitioners to decouple safety from task performance: as long as the value function remains above the threshold, the system is guaranteed to have at least one valid control action available to remain safe indefinitely.

Bridging Reachability and Control Barrier Functions

Historically, HJ reachability and Control Barrier Functions (CBFs) were viewed as competing paradigms. HJ reachability provided global, worst-case guarantees but was computationally heavy to solve online. CBFs offered lightweight, optimization-based filtering in real-time but were incredibly difficult to construct globally for nonlinear systems.

The modern synthesis bridges this gap by using the precomputed HJ value function as a Control Barrier Function. Jason J. Choi et al. formalize this in Robust Control Barrier-Value Functions for Safety-Critical Control by introducing Robust Control Barrier-Value Functions (CBVFs). By treating the reachability value function as a robust barrier, the online controller can freely maximize performance inside the safe operating envelope, knowing exactly where the boundary lies.

This insight enables the systematic derivation of safety filters without manual tuning. As Javier Borquez et al. demonstrate in On Safety and Liveness Filtering Using Hamilton-Jacobi Reachability Analysis, these value functions naturally respect predefined control bounds, avoiding the aggressive, "jittery" behavior typical of traditional bang-bang controllers. Similarly, Azra Begzadić et al. leverage this in Back to Base: Towards Hands-Off Learning via Safe Resets with Reach-Avoid Safety Filters by designing a Viscosity-Based Control Barrier Function (VB-CBF). This filter ensures that an autonomous agent undergoing reinforcement learning can explore aggressively but is guaranteed to safely return to a "home base" for resets without human intervention.

Real-Time Deployment and Scaling

Because the HJI equation can be solved offline, the resulting value function acts as a precomputed lookup table or a differentiable surface that online controllers can query instantaneously.

This enables seamless integration with modern planning pipelines:

Overcoming the Curse of Dimensionality

The primary bottleneck of HJ reachability is that solving partial differential equations grids the state space, scaling exponentially with the number of state dimensions. To make these computations tractable for high-dimensional systems, Dylan Hirsch et al. introduce a singular perturbation approach in Approximate Hamilton-Jacobi Reachability Analysis for a Class of Two-Timescale Systems, with Application to Biological Models. By decomposing complex systems into lower-dimensional "fast" and "slow" subsystems, the authors demonstrate that you can compute rigorous inner and outer approximations of the safe set, preserving formal safety guarantees without the intractable computational burden.

Key Foundational Work

Reach-avoid problems with time-varying dynamics, targets and constraints (2015) — contributed 1 concept elements to this direction.
Hamilton-Jacobi Formulation for Reach-Avoid Differential Games (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_hamilton_jacobi_reachability_analysis_eec4638b,
  title = {Hamilton-Jacobi reachability analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hamilton-jacobi-reachability-analysis/rdr_eec4638bc07147ada80b89cf14d58bfe},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_hamilton_jacobi_reachability_analysis_eec4638b_v1,
  title = {Hamilton-Jacobi reachability analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/hamilton-jacobi-reachability-analysis/rdr_eec4638bc07147ada80b89cf14d58bfe/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake