Parameterized Barrier Functions to Guarantee Safety Under Uncertainty

Anil Alan, Tamás G. Molnár, Aaron D. Ames, Gábor OroszIEEE Control. Syst. Lett 2023 · DOI

TL;DR

This paper introduces Parameterized Barrier Functions (PBFs) as a unified framework to handle various types of uncertainties in safety-critical control systems. PBFs generalize existing robust control barrier function (RCBF) methods, offering a more flexible way to quantify safety guarantees, including degradation and conservativeness, and showing that Input-to-State Safety (ISSf) is a special case.

What Problem Does This Paper Address?

Ensuring the safety of autonomous systems, like self-driving cars or robots, is paramount. Control Barrier Functions (CBFs) are a popular mathematical tool for guaranteeing safety by defining a safe set of states. However, real-world systems operate with uncertainties – in their models, sensors, or environments. Existing methods for handling these uncertainties, often called robust CBFs, have led to a fragmented landscape of solutions, each tailored to specific uncertainty types. This makes it difficult to compare approaches and understand the trade-offs between safety and performance. The paper aims to bridge this gap by proposing a general framework that unifies these diverse robustification techniques and provides a more nuanced understanding of safety under uncertainty.

Methods

The core contribution is the introduction of Parameterized Barrier Functions (PBFs). A PBF, denoted by $H(x, h^*) = h(x) - h^*$, is built upon an existing Control Barrier Function $h(x)$ and introduces a new parameter $h^*$. This parameter $h^*$ allows us to define a new safe set $S^* = \{x \in \mathbb{R}^n | H(x, h^*) \ge 0\}$, which can be larger than, equal to, or smaller than the original safe set $S = \{x \in \mathbb{R}^n | h(x) \ge 0\}$.

The paper then extends the concept of Robust Control Barrier Functions (RCBFs) to this PBF framework. A key theorem (Theorem 3) establishes conditions under which a controller designed using a PBF can guarantee safety with respect to the set $S^*$, even in the presence of additive and multiplicative uncertainties in the system dynamics. This theorem generalizes existing RCBF results by allowing the safety guarantee to be established for a potentially modified safe set defined by $h^*$.

Furthermore, the paper demonstrates that Input-to-State Safety (ISSf), a concept used to quantify safety degradation under bounded disturbances, is a special case of the PBF framework where $h^* < 0$. By leveraging the PBF framework, the authors show how to derive more accurate safety guarantees for ISSf-based controllers, allowing for cases where $h^* = 0$ (standard safety) and $h^* > 0$ (conservative safety).

Datasets

Not discussed in the paper. The paper focuses on theoretical development and uses a simulated example (inverted pendulum) for illustration.

Key Insights and Ideas

  • Unification of Robustness: PBFs provide a single, elegant framework to encompass various existing methods for robustifying CBFs, simplifying the understanding and comparison of different approaches.
  • Quantifying Safety Trade-offs: The parameter $h^*$ in PBFs allows for a direct quantification of safety degradation (when $h^* < 0$) and conservativeness (when $h^* > 0$), offering a more nuanced view than traditional binary safety guarantees.
  • ISSf as a Special Case: Recognizing ISSf as a specific instance of PBFs ($h^* < 0$) provides a theoretical link and opens avenues for improving ISSf-based controllers.
  • Flexibility in Safety Specification: The ability to define safety with respect to $S^*$ (a superlevel set of $h$) rather than strictly $S$ (where $h \ge 0$) is crucial for handling uncertainties without overly conservative control actions.

Key Results

The paper presents theoretical results (Theorems 1-4 and Corollaries 1-2) that establish conditions for safety guarantees under uncertainty using PBFs. The primary "result" is the framework itself and its ability to generalize and improve upon existing safety guarantees.

Through the inverted pendulum example, the paper illustrates:

  • How under-approximating uncertainty can lead to safety violations (system leaving the safe set $S$).
  • How over-approximating uncertainty leads to conservative behavior (system staying within a smaller subset of $S$).
  • How PBFs, by defining $h^*$ based on the uncertainty estimation, can provide more accurate bounds on the system's behavior, reflecting both degradation and conservativeness. For instance, Figure 3 shows how the PBF-derived lower bound $h^*$ accurately tracks the actual $h(x(t))$ for different uncertainty estimations, unlike the fixed ISSf bound.

Limitations and Future Work

  • Computational Complexity: While the framework is general, the practical implementation of PBFs and the computation of $h^*$ might still involve complex optimization problems, especially for high-dimensional systems.
  • Data-Driven Integration: The authors suggest future work on merging the PBF framework with data-driven schemes to obtain online safety guarantees, which would be crucial for systems with unknown or time-varying uncertainties.
  • Specific Uncertainty Classes: While PBFs are general, the specific forms of the compensation term $\sigma$ and the functions $\alpha$ and $\epsilon$ used in the theorems and corollaries are still tied to certain assumptions about the uncertainty structure (e.g., bounded additive uncertainty).

The paper opens up avenues for developing more adaptive and less conservative safety controllers by providing a unified theoretical foundation for handling uncertainties.

Summary updated

BibTeX
@article{alan2023parameterized,
  author = {Anil Alan and Tamás G. Molnár and Aaron D. Ames and Gábor Orosz},
  title = {{Parameterized Barrier Functions to Guarantee Safety Under Uncertainty}},
  journal = {IEEE Control. Syst. Lett},
  year = {2023},
  doi = {10.1109/lcsys.2023.3285188},
}
Figure