Safety Under Uncertainty: Tight Bounds with Risk-Aware Control Barrier Functions

Mitchell K. Black Jr., Mitchell Black, Georgios Fainekos, Bardh Hoxha, Danil V. Prokhorov, Dimitra Panagou

TL;DR

This paper introduces a new class of "risk-aware control barrier functions" (RA-CBFs) for ensuring the safety of stochastic systems. Unlike previous methods that can be overly conservative, RA-CBFs provide tighter, more realistic bounds on the probability of system failure over time, demonstrated through simulations on a mobile robot and an autonomous vehicle highway merging scenario.

What Problem Does This Paper Address?

Ensuring the safety of autonomous systems, especially in the presence of uncertainty and randomness, is a critical challenge. Control Barrier Functions (CBFs) are a popular tool for guaranteeing that a system remains within a safe set of states. However, existing methods for stochastic systems (S-CBFs) often rely on martingale theory, which can lead to overly conservative safety guarantees. This conservatism means the system might avoid taking actions that are actually safe, limiting its performance. The paper aims to bridge the gap between theoretical safety guarantees and practical system behavior by developing a new framework that provides more accurate risk bounds.

Methods

The core contribution is the introduction of Risk-Aware Control Barrier Functions (RA-CBFs). Instead of relying on martingale theory, RA-CBFs leverage a result from the stochastic level-crossing literature. The key innovation lies in a new condition for RA-CBFs that uses an "infinitesimal generator" derived from stochastic differential equations. This generator characterizes how a function (representing safety) changes over time due to both deterministic control inputs and random noise.

The RA-CBF condition is formulated as: $$ \inf_{u \in U} \Gamma_B(x, u) \leq \alpha(h(I_\Gamma(t))) $$ where $\Gamma_B$ is the generator, $u$ is the control input, $\alpha$ is a Lipschitz continuous function, and $h(I_\Gamma(t))$ is a term that incorporates the system's current state, the desired risk level ($p_d$), and an integral of the generator over time ($I_\Gamma(t)$). This formulation allows for a more direct bounding of the probability of exiting the safe set.

The paper also proposes a method for handling complex safety specifications by partitioning the safe set into multiple sub-level sets and applying RA-CBFs to each. This allows for a more granular control over risk and can lead to less conservative behavior.

Datasets

Not discussed in the paper. The paper focuses on theoretical development and numerical simulations using synthetic examples.

Key Insights and Ideas

  • Beyond Martingales: The paper demonstrates that moving beyond martingale theory for S-CBFs can lead to less conservative and more accurate safety bounds in stochastic systems.
  • Tighter Risk Bounds: RA-CBFs provide a more precise upper bound on the probability of a system becoming unsafe within a finite time interval, better reflecting real-world performance.
  • Performance vs. Safety Trade-off: The RA-CBF framework allows for a more nuanced trade-off between safety guarantees and system performance. By adjusting design parameters, controllers can be made to take more aggressive (and thus potentially higher-performing) actions while still maintaining a specified level of safety.
  • Hierarchical Safety: The proposed method of partitioning the safe set into sub-level sets and applying RA-CBFs hierarchically is a powerful technique for managing complex safety requirements and achieving desired risk levels.

Key Results

  • Mobile Robot Example: In a simulation with a mobile robot, the RA-CBF controller achieved a measured safety probability that closely matched a very low theoretical risk bound (0.01), while the S-CBF controller, with a much higher theoretical risk bound (0.505), still showed a significant fraction of unsafe outcomes. This highlights the conservatism of S-CBFs.
  • Autonomous Vehicle Highway Merging: The RA-CBF controller successfully navigated a complex highway merging scenario with 11 vehicles, achieving a safety guarantee of 99.1% (risk < 0.01). This demonstrates the practical applicability of the framework in safety-critical scenarios.
  • Theoretical Guarantees: The paper provides theoretical guarantees that RA-CBFs can achieve a smaller system risk than existing S-CBF methods under certain conditions.

Limitations and Future Work

  • Computational Complexity: While not explicitly detailed, the numerical simulations and the use of QP-based controllers suggest that real-time implementation might require efficient solvers.
  • Parameter Tuning: The RA-CBF framework involves several design parameters (e.g., $p_d$, $\alpha$, $\gamma$, and the partitioning of safe sets), and determining optimal values for these parameters might require further investigation.
  • Future Work: The authors suggest extending the framework to consider measurement noise and investigating applications in recovery problems within the context of safe, predictive control.

Summary updated

Abstract

We propose a novel class of risk-aware control barrier functions (RA-CBFs) for the control of stochastic safety-critical systems. Leveraging a result from the stochastic level-crossing literature, we deviate from the martingale theory that is currently used in stochastic CBF techniques and prove that a RA-CBF based control synthesis confers a tighter upper bound on the probability of the system becoming unsafe within a finite time interval than existing approaches. We highlight the advantages of our proposed approach over the state-of-the-art via a comparative study on an mobile-robot example, and further demonstrate its viability on an autonomous vehicle highway merging problem in dense traffic.
BibTeX
@inproceedings{jr2023safety,
  author = {Mitchell K. Black Jr. and Mitchell Black and Georgios Fainekos and Bardh Hoxha and Danil V. Prokhorov and Dimitra Panagou},
  title = {{Safety Under Uncertainty: Tight Bounds with Risk-Aware Control Barrier Functions}},
  booktitle = {IEEE International Conference on Robotics and Automation},
  year = {2023},
  doi = {10.1109/icra48891.2023.10161379},
}
Figure