Safe Control for Nonlinear Systems With Stochastic Uncertainty via Risk Control Barrier Functions

Andrew Singletary, Mohamadreza Ahmadi, Aaron D. Ames

TL;DR

This paper introduces "Risk Control Barrier Functions" (RCBFs) to guarantee safety for nonlinear systems operating under uncertainty. By leveraging coherent risk measures, specifically Conditional Value-at-Risk (CVaR), RCBFs provide a more robust safety guarantee than traditional methods, ensuring that systems remain within safe bounds even in the face of unpredictable disturbances.

What Problem Does This Paper Address?

Ensuring the safety of autonomous systems, like robots, operating in real-world environments is a critical challenge. These systems often face "stochastic uncertainties" – unpredictable variations in their dynamics or environment. Traditional safety methods, which often guarantee safety with high probability or in expectation, can be insufficient when the consequences of failure are severe. This is because they might not adequately account for the worst-case scenarios that can occur with a small but non-negligible probability. The paper aims to bridge this gap by providing a framework for "risk-sensitive safety," where safety is guaranteed not just probabilistically, but in a way that considers the potential severity of undesirable outcomes as defined by coherent risk measures.

Methods

The core contribution of this paper is the development of Risk Control Barrier Functions (RCBFs). These are an extension of the well-established Control Barrier Functions (CBFs). While standard CBFs ensure that a system's state remains within a safe set, RCBFs do so under a risk-sensitive criterion.

Here's a breakdown of the key methodological elements:

  • Coherent Risk Measures: The paper grounds its approach in the theory of coherent risk measures, which are desirable properties for quantifying risk (e.g., monotonicity, convexity). A prominent example discussed is the Conditional Value-at-Risk (CVaR), which focuses on the expected loss in the worst-case scenarios (e.g., the worst 1% of outcomes).
  • Risk Control Barrier Functions (RCBFs): An RCBF is a function $h(x)$ associated with a safe set $S = \{x | h(x) \ge 0\}$. The existence of an RCBF guarantees that the system's state $x_{t+1}$ will satisfy a risk measure condition relative to $h(x_t)$, specifically $p(h(x_{t+1})) \ge \alpha(h(x_t))$, where $p$ is a coherent risk measure and $\alpha$ is a class K function. This condition ensures that the risk measure of the next state's safety function is bounded from below by a function of the current state's safety function.
  • Finite-Time RCBFs: To address the need for reaching a safe set within a specific timeframe, the authors introduce finite-time RCBFs. These functions provide a guarantee that the system will reach the safe set within a finite number of steps, $t^*$, with a bound on $t^*$ provided.
  • Boolean Compositions: The framework is extended to handle safety and reachability for sets defined by Boolean combinations (AND/OR) of multiple safety functions, leveraging the properties of min and max operations on these functions.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and simulation-based validation.

Key Insights and Ideas

  • Beyond Probabilistic Safety: The paper highlights that traditional probabilistic safety guarantees might not be sufficient for safety-critical applications. By using coherent risk measures like CVaR, we can achieve a more robust form of safety that explicitly considers the severity of potential failures.
  • RCBFs as a Unified Framework: RCBFs provide a powerful generalization of CBFs, allowing for risk-sensitive safety guarantees within a single, elegant mathematical framework.
  • Finite-Time Guarantees: The introduction of finite-time RCBFs is crucial for practical applications, as it allows for guaranteed convergence to a safe state within a predictable timeframe, not just asymptotic safety.
  • Tuning for Risk Aversion: The use of CVaR with different confidence levels ($\beta$) allows for tuning the level of risk aversion. A lower $\beta$ (e.g., 0.01) leads to a more risk-averse controller that prioritizes avoiding extreme negative outcomes.

Key Results

  • Theoretical Guarantees: The paper formally proves that the existence of an RCBF guarantees risk-sensitive safety (p-safety) for discrete-time nonlinear systems with stochastic uncertainty.
  • Finite-Time Reachability: The existence of a finite-time RCBF guarantees risk-sensitive reachability to a safe set within a finite time, with a derived upper bound on this time.
  • Simulation Efficacy: The proposed method was demonstrated on a nonlinear cart-pole system. Simulations showed that controllers designed with RCBFs (both standard and finite-time) successfully maintained safety, even in scenarios where a nominal controller would violate safety constraints. Specifically, the CVaR-based filter (with $\beta=0.01$) provided superior safety performance compared to a standard expectation-based filter.

Limitations and Future Work

  • Computational Complexity: While the paper mentions that solving the optimization problem for RCBFs is feasible in real-time (e.g., 0.7 ms per solution for the cart-pole example), the computational cost can increase with system dimension and the complexity of the risk measure.
  • Linearization of RCBFs: The authors suggest that future work could involve linearizing nonlinear RCBFs to create affine RCBF constraints, which might simplify the optimization problem.
  • Data-Driven Approaches: The paper mentions that future work could explore how nonlinear CBFs can be linearized to result in affine RCBF constraints, with the error included in the stochastic uncertainty to result in formal safety guarantees.
  • Generalization to Continuous-Time Systems: While the paper focuses on discrete-time systems, extending these concepts to continuous-time systems would be a valuable next step.

Summary updated

Abstract

Guaranteeing safety for robotic and autonomous systems in real-world environments is a challenging task that requires the mitigation of stochastic uncertainties. Control barrier functions have, in recent years, been widely used for enforcing safety related set-theoretic properties, such as forward invariance and reachability, of nonlinear dynamical systems. In this paper, we extend this rich framework to nonlinear discrete-time systems subject to stochastic uncertainty and propose a framework for assuring risk-sensitive safety in terms of coherent risk measures. To this end, we introduce risk control barrier functions (RCBFs), which are compositions of barrier functions and dynamic, coherent risk measures. We show that the existence of such barrier functions implies invariance in a coherent risk sense. Furthermore, we formulate conditions based on finite-time RCBFs to guarantee finite-time reachability to a desired set in the coherent risk. Conditions for risk-sensitive safety and finite-time reachability of sets composed of Boolean compositions of multiple RCBF are also formulated. We show the efficacy of the proposed method through its application to a cart-pole system in a safety-critical scenario.
BibTeX
@article{singletary2023safe,
  author = {Andrew Singletary and Mohamadreza Ahmadi and Aaron D. Ames},
  title = {{Safe Control for Nonlinear Systems With Stochastic Uncertainty via Risk Control Barrier Functions}},
  journal = {IEEE Control. Syst. Lett},
  year = {2023},
  doi = {10.1109/lcsys.2022.3187458},
}
Figure