A Hamilton-Jacobi Reachability Framework with Soft Constraints for Safety-Critical Systems

Chams Eddine Mballo, Donggun Lee, Claire J. TomlinarXiv 2025 · DOI · arXiv

Abstract

Traditional Hamilton-Jacobi (HJ) reachability analysis often fails in emergency scenarios because it treats all safety requirements as unbreakable "hard" constraints. This paper proposes a novel framework that incorporates "soft constraints"—requirements that can be temporarily violated within a specified time budget. By augmenting the system state to track cumulative violations and employing a regularization technique for discontinuous dynamics, the framework identifies safe control strategies where traditional methods find none.

Introduction

In safety-critical domains such as autonomous aviation and electric vehicles, Reachability Analysis is a cornerstone for ensuring system integrity. It identifies the "safe set"—the collection of initial states from which a system is guaranteed to remain safe despite worst-case disturbances. However, standard frameworks are often overly conservative, treating all constraints as binary and unbreakable. In real-world emergency scenarios, such as an engine failure, strictly adhering to secondary constraints (e.g., battery temperature or recommended speed limits) may lead to an empty safe set, rendering a catastrophic outcome inevitable. This research introduces a method to treat safety margins as a manageable resource, allowing systems to "spend" a violation budget to maintain overall mission safety.

Methodology

The proposed framework extends the classical Hamilton-Jacobi-Isaacs (HJI) approach through two primary innovations:

  1. State Augmentation: The system's state space is expanded with an auxiliary variable, $z$, which functions as a cumulative timer. This variable tracks the total duration the system spends outside the soft-constraint boundaries.
  2. Regularization of Discontinuous Dynamics: Tracking violations typically involves indicator functions, which introduce discontinuities that standard numerical solvers cannot handle. The authors introduce a Lipschitz-continuous approximation, $h_\epsilon$, to smooth these transitions. This allows the use of established toolboxes, such as Level Set or HelperOC, to solve the resulting partial differential equations (PDEs).

The resulting Soft-Constrained Reach-Avoid Set identifies all initial conditions from which a system can reach a target state while strictly adhering to hard constraints and limiting soft-constraint violations to a user-defined budget, $Q$.

Key Results

The framework was validated through high-fidelity numerical simulations, including a point-mass landing model and a DC9-30 fixed-wing aircraft model.

Expansion of the Reach-Avoid set with increasing violation budgets.

As shown in the figure above, the "Verified Safe Region" (the Reach-Avoid set) expands significantly as the allowed violation budget increases. In the aircraft engine-failure scenario, a safe landing from altitudes above 22 meters was mathematically impossible under strict hard constraints. By allowing a small violation of the recommended speed envelope, the safe region expanded to cover the entire operational altitude range.

Furthermore, the authors provided a formal convergence proof, demonstrating that as the smoothing parameter ($\epsilon$) approaches zero, the approximate solution converges to the true value function. Numerical tests on a $300 \times 300$ grid showed mean boundary errors as low as 0.0006, confirming that the regularization process maintains high precision.

Discussion and Future Work

The core insight of this work is the conceptualization of safety as a flexible resource rather than a binary state. By providing a mathematical bridge between discontinuous logic and continuous PDE solvers, the framework offers a practical tool for emergency response in autonomous systems.

However, challenges remain. Like most HJ methods, the framework suffers from the "curse of dimensionality," with computational costs growing exponentially with the number of states. Current simulations are limited to 2D and 3D models. Future research will explore neural-network-based approximations, such as DeepReach, to scale the method to higher-dimensional systems (e.g., 12-state drones). Additionally, future iterations may incorporate the magnitude of violations rather than just their duration.

Conclusion

This soft-constrained HJ reachability framework provides a robust mathematical foundation for balancing safety and flexibility. By allowing controlled violations of non-critical constraints, it enables autonomous systems to navigate emergency scenarios that would otherwise be deemed unrecoverable.

Summary updated

BibTeX
@misc{mballo2025hamilton,
  author = {Chams Eddine Mballo and Donggun Lee and Claire J. Tomlin},
  title = {{A Hamilton-Jacobi Reachability Framework with Soft Constraints for Safety-Critical Systems}},
  year = {2025},
  eprint = {2510.24933},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2510.24933},
}
Figure