Control Barrier Function-Based Quadratic Programs Introduce Undesirable Asymptotically Stable Equilibria

Matheus F. Reis, A. Pedro Aguiar, Paulo Tabuada

This paper, published in 2020 by researchers from the University of Porto and UCLA, identifies a critical flaw in a popular framework for safe autonomous control and proposes an elegant mathematical "workaround" to keep robots from getting stuck.

TL;DR

Standard safety-critical control methods that combine stability (Control Lyapunov Functions) and safety (Control Barrier Functions) via Quadratic Programs can accidentally create "trap" points on the edges of obstacles. This paper proves these undesirable equilibria exist, shows they can be "sticky" (asymptotically stable), and introduces a "Lyapunov Shaping" technique to rotate the control objective and steer the system around these traps.

What Problem Does This Paper Address?

In modern robotics, we often use Control Lyapunov Functions (CLFs) to tell a robot where to go (stability) and Control Barrier Functions (CBFs) to tell it what to avoid (safety). A popular way to combine these is through a Quadratic Program (QP).

The QP acts like a real-time negotiator: it tries to follow the CLF's "gradient" toward the goal while treating the CBF as a hard constraint. If the goal is behind an obstacle, the QP "softens" the stability requirement to ensure the robot doesn't crash.

However, the authors point out a major catch: this negotiation can lead to a stalemate. If the direction to the goal is perfectly opposed to the direction away from an obstacle (collinearity), the robot might reach a point on the obstacle's boundary where the "best" safe move is to simply stop. The robot becomes trapped in an "undesirable equilibrium" that isn't the actual goal.

Methods

The paper approaches the problem in two phases: mathematical analysis and controller redesign.

  1. Stability Analysis: The authors use the Karush-Kuhn-Tucker (KKT) conditions—the mathematical requirements for a solution to an optimization problem—to prove exactly where these "trap" points occur. They derive Theorem 2, which shows that if the "curvature" of your goal-seeking function (CLF) is sharper than the curvature of your obstacle (CBF), the trap becomes asymptotically stable—meaning if the robot gets near it, it will be sucked in and stay there.
  2. Lyapunov Shaping: To fix this, they propose a novel extension. Instead of a fixed, "bowl-shaped" CLF, they use a non-radial CLF that can rotate. They introduce a virtual state (a rotation matrix $Q$) and a new barrier function ($h_D$) that measures how close the system is to a "stuck" configuration (collinearity).
  3. Modified QP: They add a third constraint to the QP. When the robot approaches a potential trap, the QP is forced to rotate the CLF. This effectively "re-shapes" the robot's perceived landscape, turning a dead-end into a path that flows around the obstacle.

Datasets

As a theoretical control theory paper, no external datasets were used. The authors evaluated their approach using numerical simulations of:

  • A 2D Integrator system: A basic point-mass model navigating around a circular obstacle.
  • Nonlinear systems: Two different mathematical models designed to test the robustness of the controller under more complex dynamics.

Key Insights and Ideas

  • The Geometry of Failure: The "aha moment" is realizing that traps occur precisely when the gradients of the CLF and CBF are collinear. If the robot is trying to go "through" the center of an obstacle to reach a goal, it will likely get stuck on the "front" of that obstacle.
  • Curvature is Key: It’s not just about the direction; the shape of the functions matters. If your goal-seeking function is too "aggressive" (high curvature) compared to the obstacle's boundary, the trap becomes much harder to escape.
  • Dynamic Objectives: Traditionally, we think of the goal as fixed. This paper suggests that by dynamically rotating our perception of the goal (the CLF shape) without changing the goal itself, we can navigate complex environments more reliably.

Key Results

  • Formal Proof: The paper provides a rigorous mathematical foundation for why these traps exist in the standard Ames et al. (2014) QP framework.
  • Trap Avoidance: In simulations where a standard QP-based controller gets stuck (see Fig. 1 in the paper), the proposed "Shaped" controller successfully steers the system around the obstacle to the origin (see Fig. 2).
  • Convergence: The authors demonstrate that their method ensures the "trap" points are no longer equilibria, meaning the only remaining stable point in the system is the actual target.

Limitations and Future Work

  • Convexity: The current mathematical proofs and the specific "shaping" logic are primarily designed for convex obstacles (like circles or boxes). Handling non-convex obstacles (like U-shapes or "dead-end" rooms) is significantly more complex and is noted as a direction for future work.
  • Computational Overhead: Adding a virtual rotation state and a third constraint to the QP increases the computational load. While manageable for simple systems, it might require optimization for high-speed or resource-constrained hardware.
  • Parameter Tuning: The method introduces new hyperparameters (like the rotation speed and the "proximity to collinearity" threshold) that may need careful tuning for different environments.

Summary updated

Abstract

Control Lyapunov functions (CLFs) and control barrier functions (CBFs) have been used to develop provably safe controllers by means of quadratic programs (QPs), guaranteeing safety in the form of trajectory invariance with respect to a given set. In this letter, we show that this framework can introduce equilibrium points (particularly at the boundary of the safe set) other than the minimum of the Lyapunov function into the closed-loop system. We derive explicit conditions under which these undesired equilibria (which can even appear in the simple case of linear systems with just one convex unsafe set) are asymptotically stable. To address this issue, we propose an extension to the QP-based controller unifying CLFs and CBFs such that the resulting system trajectories avoid the undesirable equilibria problem on the boundary of the safe set. The solution is illustrated in the design of a collision-free controller.
BibTeX
@inproceedings{reis2021control,
  author = {Matheus F. Reis and A. Pedro Aguiar and Paulo Tabuada},
  title = {{Control Barrier Function-Based Quadratic Programs Introduce Undesirable Asymptotically Stable Equilibria}},
  booktitle = {IEEE Control. Syst. Lett},
  year = {2021},
  url = {https://openreview.net/forum?id=AHFN4P4P7fz},
}

Loading related content...

Figure