Control Barrier Function (CBF) Theory

• Control Barrier Function (CBF) Theory is a framework that formalizes safety constraints by converting set invariance into optimization-based control synthesis.
• It guarantees forward invariance of safe sets by employing quadratic program designs that satisfy affine input constraints and robust compatibility margins.
• The approach uses grid-based certification algorithms to verify multiple barrier functions, addressing input limitations and computational challenges.

Control Barrier Function (CBF) theory formalizes safety constraints in control systems as set-invariance conditions enforced through affine constraints on the control input. Developed to address state and input constraints, CBFs rigorously transform the problem of set invariance into real-time controller synthesis, particularly well-suited to modern optimization-based safety-critical control.

1. Formal Definition and Invariance Guarantees

A control-affine system is modeled as

$$\dot x = f(x) + g(x) u,\quad x \in \mathbb{R}^n,\, u \in \mathcal{U} \subset \mathbb{R}^m$$

where $f$ and $g$ are locally Lipschitz and $\mathcal{U}$ encodes control limits. The desired safe set is typically specified as a zero-superlevel set of $N$ continuously differentiable barrier functions: $$C = \{ x \in \mathbb{R}^n : h_i(x) \ge 0,\; i = 1, \dots, N \}$$ A $C^1$ function $h(x)$ is a control barrier function (CBF) on a domain $D \supset C$ if there exists an extended class–$\mathcal{K}$ function $\alpha$ such that

$$\forall x \in D,\, \exists u \in \mathcal{U} : \quad L_f h(x) + L_g h(x)\, u + \alpha(h(x)) \ge 0$$

with $$L_f h(x) = \nabla h(x)^\top f(x),\, L_g h(x) = \nabla h(x)^\top g(x)$$. For multiple constraints, the input set is determined by the intersection: $$K(x) = \Big\{ u \in \mathcal{U} : A(x) u + b(x) \ge 0 \Big\}$$ where $A(x)$ stacks the $L_g h_i(x)$ row vectors and $b(x)$ stacks $L_f h_i(x) + \alpha_i(h_i(x))$. Forward invariance of $C$ is guaranteed for all $u(x) \in K(x)$ via Nagumo’s set invariance theorem (Tan et al., 2022).

2. Multiple Barriers and Compatibility

When multiple CBF constraints are imposed, their compatibility may not be automatic, particularly under input bounds. Compatibility holds if $K(x) \neq \emptyset$ for all $x \in D$. If not, the associated quadratic program (QP) for real-time controller synthesis is infeasible, and safety cannot be enforced.

Robust compatibility strengthens each barrier inequality to

$$L_f h_i(x) + L_g h_i(x) u + \alpha_i(h_i(x)) \ge \eta$$

where $\eta > 0$ is a margin ensuring invariance against additive state disturbances with worst-case Lie derivative bounded by $\eta$. The maximal attainable uniform slack $c(x)$, defined as

$$c(x) = \max_{u \in \mathcal{U},\, t \in \mathbb{R}} t \;\; \text{s.t.}\;\; A(x) u + b(x) \ge t \mathbf{1}$$

quantifies the compatibility at each $x$ (Tan et al., 2022).

3. Quadratic Program (QP) Controller Synthesis

CBF-based controllers are commonly synthesized via an online QP: $$u^*(x) = \arg\min_{u \in \mathbb{R}^m} \frac{1}{2} \|u - u_{\rm nom}(x)\|^2$$ subject to

$$A(x) u + b(x) \ge 0,\quad u_{\min} \leq u \leq u_{\max}$$

where $u_{\rm nom}(x)$ is a performance-oriented nominal controller, and actuator limits are explicit. The QP’s feasibility directly depends on CBF compatibility (Tan et al., 2022).

Certified feasibility is essential; otherwise, some states may violate safety constraints. If the QP admits a solution at each $x \in C$, the closed-loop system is provably forward-invariant in $C$.

4. Algorithmic Verification of Compatibility

A rigorous grid-based offline certification algorithm establishes robust compatibility or finds counterexamples. The procedure exploits global Lipschitz bounds of $A(x)$ and $b(x)$: $$\|A(x) - A(y)\|_\infty \le L_A \|x - y\|_\infty,\quad \|b(x) - b(y)\|_\infty \le L_b \|x - y\|_\infty$$ At each lattice center $x_k$, calculate $c_k = c(x_k)$ and optimizer $u_k^*$. The cube $B(x_k,\rho_k)$, with radius

$\rho_k = \frac{2 c_k}{L_A \|u_k^*\|_\infty + L_b}$

is certified compatible. Shells where $\rho_k < r$ are recursively refined. The algorithm terminates finitely under a uniform robustness margin $\eta > 0$, providing either a compatibility certificate or an upper bound $\eta'$ (Tan et al., 2022).

5. Numerical Illustration and Complexity

Representative examples use 2-dimensional control-affine dynamics with box-bounded inputs and annular safety regions. Key parameters (grid size $r_0$, refinement factor $\lambda$, Lipschitz constants $L_A$, $L_b$) determine computational cost. Compatible and incompatible input bounds yield respectively (i) finite certification time—three iterations with $U_{\max}=3$, (ii) immediate detection of incompatibility with $U_{\max}=2$. Robust margin bounds are empirically validated.

The algorithm is restricted by the curse of dimensionality; practical application is limited to $n \leq 4$. Local Lipschitz estimations and boundary-only certification can mitigate computational burden (Tan et al., 2022).

6. Extensions and Limitations

Further improvements include:

• Restricting verification to neighborhoods of $\partial C$ (boundary-only certification)
• Employing non-cube coverings or adaptive tessellations
• Generalizing to time-varying systems and sets by state-space augmentation with time
• Utilizing local Lipschitz constants for reduced conservatism

Limitations center on dimensionality and conservatism. Grid refinement scales exponentially with state dimension. Over-conservatism may occur due to global Lipschitz bounds and can be reduced by employing local analysis.

The presented grid-based certification framework guarantees offline verification of feasibility for multiple CBF constraints under input bounds, ensuring QP-based controller synthesis can be conducted with provable online feasibility (Tan et al., 2022). This is fundamental for robust safety-critical control design in input-constrained, multi-barrier control systems.