Adaptive Safe Behavior Generation for Heterogeneous Autonomous Vehicles Using Parametric-Control Barrier Functions (Student Abstract)

Yiwei Lyu, Wenhao Luo, John M. DolanAAAI 2022 · DOI

Abstract

Ensuring safety in autonomous driving often relies on Control Barrier Functions (CBFs), which typically assume uniform or cooperative behavior from surrounding vehicles. However, real-world traffic is heterogeneous, consisting of drivers with varying levels of aggressiveness. This paper introduces Parametric-Control Barrier Functions (Parametric-CBFs), a framework that uses weighted polynomial functions to model diverse driving styles. By estimating these parameters in real-time via ridge regression, an autonomous vehicle can adapt its behavior, improving merging efficiency by nearly 40% while maintaining formal safety guarantees.

Introduction

Safety is the primary constraint in autonomous vehicle (AV) navigation. Traditional methods, such as standard Control Barrier Functions (CBFs), define a "safe set" to prevent collisions. However, these models often lack the expressivity to account for the diverse driving styles encountered in real-world traffic. Standard CBFs typically assume linear safety constraints, which can lead to overly conservative or inefficient maneuvers when interacting with aggressive or timid human drivers. This paper addresses the challenge of heterogeneous traffic by proposing Parametric-CBFs, which allow AVs to learn and adapt to the specific behaviors of surrounding vehicles on the fly.

Methodology

The core innovation is the Parametric-CBF, which replaces simple linear safety constraints with a richer, weighted polynomial structure.

  1. Mathematical Framework: The safety constraint is defined using a vector $H(x)$ of independent odd-powered functions (e.g., $h(x), h^3(x), \dots$). A weight vector $\alpha$ determines the influence of each term. This allows the system to represent a wide spectrum of behaviors, from "hard" avoidance to "soft" gradual adjustments.
  2. Real-Time Parameter Estimation: The "ego" vehicle observes the states of neighboring cars and employs ridge linear regression to estimate their $\alpha$ parameters. This enables the robot to identify a neighbor's driving style within milliseconds.
  3. Adaptive Control: With the learned parameters, the ego vehicle solves a Quadratic Program (QP) to determine the optimal control input. This ensures the vehicle follows its desired path as closely as possible while strictly adhering to the safety boundaries defined by the predicted behavior of others.

Experimental Results

The framework was evaluated in a simulated highway ramp merging scenario.

Overview of the Parametric-CBF framework: (a) ramp merging scenario, (b) parameter convergence, (c) behavioral expressivity, and (d) performance comparison between control with and without prediction.

The results demonstrate significant improvements in both speed and efficiency:

  • Rapid Convergence: The system accurately predicted the driving styles of surrounding vehicles within 0.1 seconds (10 timesteps), as shown in Figure 1(b).
  • Enhanced Efficiency: Implementing Parametric-CBF prediction led to a 39.6% increase in task efficiency for the ego vehicle compared to a baseline without prediction.
  • Improved Traffic Flow: The overall coordination efficiency of the multi-vehicle system improved by 16.1%, reducing congestion at the merge point.
  • Behavioral Expressivity: As illustrated in Figure 1(c), the use of different $\alpha$ values allows the model to capture a variety of safety-critical trajectories that linear CBFs cannot represent.

Discussion and Conclusion

Parametric-CBFs provide a computationally efficient way to bridge the gap between formal safety guarantees and the unpredictability of human driving. By moving beyond linear constraints, AVs can "empathize" with surrounding drivers, leading to smoother and more efficient interactions. While this study focuses on ramp merging, future work will explore more complex urban scenarios, such as unprotected left turns. Additionally, further research is needed to ensure robustness when human behavior does not perfectly align with the Parametric-CBF mathematical model and to scale the optimization process for high-density traffic environments.

Summary updated

BibTeX
@inproceedings{lyu2022adaptive,
  author = {Yiwei Lyu and Wenhao Luo and John M. Dolan},
  title = {{Adaptive Safe Behavior Generation for Heterogeneous Autonomous Vehicles Using Parametric-Control Barrier Functions (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.1609/aaai.v36i11.21641},
}
Figure