Rule-Based Safety-Critical Control Design using Control Barrier Functions with Application to Autonomous Lane Change

Suiyi He, Jun Zeng, Bike Zhang, Koushil Sreenath

Introduction

Lane changing remains one of the most challenging maneuvers for autonomous vehicles (AVs), contributing to approximately 9% of all vehicle crashes according to the National Highway Traffic Safety Administration (NHTSA). The primary difficulty lies in the dynamic nature of traffic; a safe gap can vanish in milliseconds if a surrounding vehicle accelerates or cuts in unexpectedly.

Traditional autonomous driving architectures often struggle with these fast-changing environments due to inherent latency. Typically, these systems separate decision-making into a hierarchical structure: a mid-level "path planner" generates a geometric trajectory over a relatively slow timeframe, while a high-frequency low-level "controller" attempts to track that trajectory. If a threat emerges after a path is planned, the controller may continue to follow a now-dangerous trajectory until the next planning cycle, creating a critical safety gap.

Comparison between traditional lane change strategies and the proposed rule-based safety-critical control architecture.

Methodology: Unifying Planning and Control

To address the latency problem, this research proposes a novel control strategy that pushes safety-critical decision-making directly into the high-frequency control loop. The system replaces the traditional path-following approach with a CLF-CBF-QP (Control Lyapunov Function - Control Barrier Function - Quadratic Program) formulation guided by a Finite State Machine (FSM).

The Control Framework

  1. Control Lyapunov Functions (CLFs): These act as the objective-seeking component, providing the "pull" that guides the vehicle toward its target lane and desired cruising speed.
  2. Control Barrier Functions (CBFs): These serve as a mathematical "shield." They enforce hard safety constraints by defining a safe set of states (e.g., minimum distance from surrounding vehicles) that the controller is forbidden to exit.
  3. Quadratic Programming (QP): By formulating the control problem as a convex QP, the system can solve for optimal steering and acceleration commands at 100Hz, allowing for near-instantaneous reactions to environmental changes.

The Finite State Machine (FSM)

The FSM serves as the high-level logic that dictates which safety constraints are active. Unlike many algorithms that assume a maneuver must be completed once started, this FSM explicitly includes "Back to Current Lane" (BL/BR) states.

The Finite State Machine (FSM) logic showing transitions between Adaptive Cruise Control, Lane Changing, and Abort states.

If the CBFs detect a threat mid-maneuver—such as a vehicle in the target lane suddenly accelerating—the FSM triggers an immediate transition to an abort state. A key theoretical contribution of this work is proving constraint continuity: ensuring that when the FSM switches states, the vehicle's current physical state remains valid within the new set of safety constraints, preventing controller instability.

Experimental Results

The proposed controller was evaluated using a kinematic bicycle model in a custom numerical simulator. The testing suite included 5,000 randomly generated traffic scenarios based on US Federal Highway Administration (FHWA) guidelines for lane width and speed.

  • Performance: In urban environments, the vehicle successfully completed lane changes in 62.46% of cases and safely aborted or waited in 37.06% of cases. In highway environments, the success rate was 55.58%, with 44.22% safe waits/aborts.
  • Safety: The system demonstrated a near-zero failure rate. Optimization failures occurred in less than 0.5% of cases, primarily due to simulator artifacts where NPC vehicles occupied the same physical space, creating mathematically impossible constraints.

Discussion and Future Work

While the results demonstrate a robust framework for reactive lane changes, several limitations remain. The current model utilizes a kinematic bicycle model which assumes zero tire slip, limiting its application in high-lateral-acceleration scenarios. Additionally, the simulation assumes "perfect perception"—the ego vehicle has zero-latency knowledge of all surrounding vehicle states.

Future research will focus on incorporating dynamic vehicle models that account for tire forces and integrating state estimators to handle sensor noise and latency. Furthermore, the authors aim to explore interactive behaviors, where surrounding vehicles respond to the ego vehicle's maneuvers, providing a more realistic representation of human-driven traffic.

Summary updated

BibTeX
@inproceedings{he2021rule,
  author = {Suiyi He and Jun Zeng and Bike Zhang and Koushil Sreenath},
  title = {{Rule-Based Safety-Critical Control Design using Control Barrier Functions with Application to Autonomous Lane Change}},
  booktitle = {ACC},
  year = {2021},
  doi = {10.23919/acc50511.2021.9482848},
}
Figure