Introduction
In the field of safety-critical control, ensuring that a system remains within a defined "safe set"—such as a drone maintaining a minimum distance from obstacles—is a fundamental requirement. Control Barrier Functions (CBFs) are the standard tool for achieving this. However, real-world applications introduce two significant challenges: external disturbances (e.g., wind or sensor noise) and gain uncertainty (e.g., fluctuations in actuator power). While Input-to-State Safety (ISSf) was developed to handle disturbances, the robustness of these controllers to gain variations—known as the gain margin—has remained poorly understood. This paper bridges the gap between safety-critical control and robust optimal control by quantifying these margins and proposing a method to ensure consistent safety across diverse nonlinear systems.
The Equivalence of Safety and Optimality
The core theoretical contribution of this work is the establishment of a three-way equivalence. The authors prove that the achievability of ISSf via feedback is equivalent to both the existence of an inverse optimal controller and the solvability of a specific Hamilton-Jacobi-Isaacs (HJI) equation.
This finding suggests that safety is not merely an external constraint imposed on a system; rather, it is a fundamental property of the system's optimal energy landscape. By demonstrating that a safe controller is effectively the solution to an optimization problem—one that minimizes control effort while maximizing safety—the authors show that these controllers can inherit the inherent robustness properties of optimal control theory.
Quantifying Asymmetric Safety Margins
A critical insight of the paper is that standard "safe" controllers do not provide uniform protection. The safety margin—the range of control gain scaling $\sigma$ that the system can tolerate—depends heavily on the system's natural dynamics at the boundary of the safe set:
- Unsafe Open-Loop Dynamics: If the system naturally drifts toward the boundary, the controller is highly robust to increasing the gain but sensitive to reductions. If both the open-loop system and the nominal controller are "unsafe," the margin is $[1/2, \infty)$.
- Conflicting Dynamics: If the open-loop system is naturally safe but the nominal (task-oriented) controller pushes it toward the boundary, the system becomes fragile to gain reductions, with a margin of only $[1/2, 1]$.
This asymmetry means that a system might remain safe if its motors become twice as powerful as expected, but could fail immediately if they lose even a small fraction of their rated power, depending on the underlying physics.
Proposed Solution: The Improved Inverse Optimal Controller
To rectify these inconsistencies, the authors propose a redesigned inverse optimal controller. By incorporating a specific "safety-boosting" term into the control law, they provide a "safety override" that is effectively twice as strong as what is nominally required.
This redesign guarantees a robust gain margin of $[1/2, \infty)$ regardless of whether the open-loop dynamics are safe or unsafe. Furthermore, this approach upgrades the safety guarantee from Local Asymptotic Stability (LAS) to Global Asymptotic Stability (GAS). This ensures that the system will return to the safe set from any initial condition, providing a crucial recovery mechanism after massive external disturbances.
Limitations and Future Directions
While the proposed controller offers superior robustness, the authors note that the improved gain margin comes at the cost of increased control effort. In resource-constrained applications, such as battery-powered robotics, this extra "safety push" may need to be balanced against energy efficiency.
Additionally, the current framework is designed for systems with a relative degree of one, where control inputs directly affect the safety function. Extending these proofs to high-relative-degree systems—where control affects acceleration to eventually change position safety—remains a challenge for future research. Finally, the computational complexity of solving HJI-related equations for high-dimensional systems suggests a need for more efficient approximation methods in practical implementations.
Summary updated