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 manuscript, we show that this framework can introduce equilibrium points (particularly at the boundary of the unsafe 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 that explicitly avoids undesirable equilibria on the boundary of the safe set. The solution is illustrated in the design of a collision-free controller.
Introduction
In modern autonomous systems, ensuring both stability (reaching a goal) and safety (avoiding obstacles) is paramount. A widely adopted framework for achieving this involves Control Lyapunov Functions (CLFs) for stability and Control Barrier Functions (CBFs) for safety, integrated via a Quadratic Program (QP). The QP acts as a real-time optimizer, attempting to follow the CLF's gradient toward the goal while treating the CBF as a hard constraint.
However, this paper identifies a critical flaw in this standard approach: the negotiation between stability and safety can lead to a stalemate. When the direction to the goal is perfectly opposed to the direction away from an obstacle—a condition known as collinearity—the system can reach an "undesirable equilibrium" on the obstacle's boundary. At these points, the robot stops moving toward its goal, effectively becoming trapped.
The Geometry of Failure
The authors utilize Karush-Kuhn-Tucker (KKT) conditions to prove exactly where these traps occur. They demonstrate that if the "curvature" of the goal-seeking CLF is sharper than the curvature of the obstacle's boundary (CBF), the trap becomes asymptotically stable. This means that if a system enters the vicinity of such a point, it will be drawn into it and remain stuck, failing to reach its objective.
As shown in Figure 1, trajectories under a nominal QP controller often terminate prematurely on the boundary of the unsafe set (the green region), rather than proceeding to the origin.
Proposed Solution: Lyapunov Shaping
To address this, the researchers propose a novel "Lyapunov Shaping" technique. Instead of using a fixed, radial CLF, they introduce a non-radial CLF that can dynamically rotate. This is achieved by:
- Introducing a Virtual State: A rotation matrix $Q$ is added to the system dynamics.
- Defining a New Barrier: A secondary barrier function ($h_D$) measures the proximity to a "stuck" configuration (collinearity).
- Modifying the QP: A third constraint is added to the optimization. When the system approaches a potential trap, the QP is forced to rotate the CLF. This effectively re-shapes the perceived landscape, turning a dead-end into a path that flows around the obstacle.
Comparative Results
The proposed method was evaluated using numerical simulations of 2D integrator systems and more complex nonlinear models. The results demonstrate that the "Shaped" controller successfully eliminates the undesirable equilibria, leaving the target origin as the only stable point in the system.
In Figure 4, the contrast is clear: while the nominal QP trajectories (blue) are diverted or trapped, the proposed QP trajectories (black) successfully navigate the obstacle to reach the goal. This confirms that by dynamically rotating the perception of the goal without changing the goal itself, robots can navigate complex environments more reliably.
Conclusion and Future Work
This work provides a rigorous mathematical foundation for a known but previously unproven issue in safe control. While the current "shaping" logic is optimized for convex obstacles, it offers a robust path forward for autonomous navigation. Future research will focus on extending these proofs to non-convex environments and optimizing the computational overhead introduced by the additional virtual states and constraints.
Summary updated