Composite Adaptive Control Barrier Function

• Composite Adaptive Control Barrier Function (CaCBF) is an adaptive framework that consolidates multiple control barrier functions to enforce safety in nonlinear control-affine systems.
• It employs online convex optimization and Lyapunov-based adaptation laws to update combination weights or parameters, ensuring the forward invariance of safe sets despite uncertainties.
• CaCBF has been successfully applied in adaptive cruise control, robotic navigation, and multi-robot coordination, offering reduced conservatism compared to traditional methods.

A Composite Adaptive @@@@1@@@@ (CaCBF) is an adaptation-based approach to enforcing safety constraints in nonlinear control-affine systems, particularly when multiple, possibly time-varying state and input constraints, uncertainty in system parameters, or higher/mixed relative-degree constraints are present. The CaCBF paradigm synthesizes multiple candidate control barrier functions (CBFs) into a unified barrier, adapts internal combination weights or parameter estimates online via convex optimization or Lyapunov-based adaptation laws, and certifies the safety—guaranteeing forward invariance of the safe set—while addressing critical challenges in input-constrained, multi-constraint, or uncertain settings (Black et al., 2023, Kamaldar, 25 Jan 2026, Black et al., 2022).

1. Fundamental Principles and Problem Formulation

CaCBF frameworks address the problem of safe control for nonlinear, control-affine systems

$\dot{x} = f(x) + G(x)u$

or, in the parametric uncertainty setting,

$\dot{x} = f(x) + F(x)\theta_* + G(x)u$

where $x \in \mathbb{R}^n$, $u \in \mathbb{R}^m$, $f$, $G$, $F$ are known (typically $C^1$) vector fields, and $\theta_*$ is an unknown but bounded parameter vector (Kamaldar, 25 Jan 2026).

Safety requirements are specified by one or more smooth constraint (barrier) functions $h_i(t,x)$, each encoding a safe set $S_i(t) = \{x \mid h_i(t,x) \ge 0\}$, for time-varying or static state constraints. In classical CBF theory, each $h_i$ must satisfy a differential inequality of extended class-$\mathcal{K}$ form under admissible inputs, certifying forward invariance via Nagumo’s theorem or its extensions (Black et al., 2023, Black et al., 2022).

2. Construction of the Consolidated/Composite Barrier Function

Given $c$ constraint functions $h_i(t,x)$, CaCBF frameworks consolidate these into a single, smooth candidate barrier $H$ using a weighted, decreasing function of the form

$H(t, w, x) = 1 - \sum_{i=1}^c \phi(h_i(t,x), w_i)$

where $w = [w_1,\dots,w_c]^\top$ are non-negative, time-varying adaptation weights, and canonical choices for $\phi$ include $\phi(h, w) = \exp(-hw)$ or $\phi(h, w) = v/(hw + v)$ with $v > 0$. The properties of $\phi$ are such that $\phi(0, \cdot) = \phi(\cdot, 0) = 1$ and $\phi$ is strictly decreasing in both arguments (Black et al., 2023, Black et al., 2022).

The zero-superlevel set $D(t) = \{ x \mid H(t, w(t), x) \ge 0 \}$ strictly under-approximates the intersection of the individual safe sets. Thus, invariance of $D(t)$ implies satisfaction of all constraints.

This construction supports arbitrarily many constraints and enables seamless handling of mixed relative-degree constraints, since the consolidated barrier incorporates all partial derivatives as needed (Black et al., 2023, Black et al., 2022).

3. Adaptive Weight or Parameter Estimation Laws

3.1 Predictor-Corrector and Interior-Point ODE (Constituent-Weight Adaptation)

The challenge is to ensure, online, that the consolidated barrier $H$ remains a valid CBF, i.e., for all $x \in D(t)$, there exists $u$ within constraints such that

$\frac{d}{dt} H(t,w,x) \ge -\alpha(H)$

even as $h_i$ or system parameters evolve. CaCBF adapts $w(t)$ using a predictor-corrector scheme derived from (i) an interior-point (log-barrier) approach for strictly feasible adaptation, and (ii) convex preference cost $J(t,w,x)$. The adaptation law is derived as the solution to an unconstrained log-barrier objective, resulting in the ODE

$\dot{w} = \mu^f(t,w,x) + \nu^f(t,w,x)u$

where $\mu^f$, $\nu^f$ are functions of gradients and Hessians of the barrier objective and are designed to enforce feasibility and linearity in $u$ (Black et al., 2023).

3.2 Parameter Adaptation under Model Uncertainty

For affine systems with linear-in-parameter uncertainty,

$\dot{x} = f(x) + F(x)\theta_* + G(x)u$

a composite Lyapunov-like function is used: $$V_c(x, \hat{\theta}) = -\ln \left( \frac{h(x)}{1+h(x)} \right) + \kappa V(x) + \frac{1}{2} (\hat{\theta} - \theta_*)^\top \Gamma^{-1} (\hat{\theta} - \theta_*)$$ where $V(x)$ is a CLF for stability, and $\Gamma \succ 0$, $\kappa > 0$. The online parameter adaptation law

$$\dot{\hat{\theta}} = \mathcal{P}_\Theta\Big(\Gamma [\kappa \phi(x)^\top - \psi(x)^\top/(h(h+1)) + \gamma F(x)^\top e ], \hat{\theta} \Big)$$

utilizes projections to maintain feasibility under a known bound $\|\theta_*\| \leq \theta_{max}$, with regressor terms $\psi(x)$ and $\phi(x)$ defined via Lie-derivatives of CBF and CLF, and instantaneous prediction error $e$ computed from system measurements (Kamaldar, 25 Jan 2026).

4. Validity Certification and Theoretical Guarantees

The central guarantee in CaCBF is that the adaptive law ensures persistent validity of the barrier condition, under bounded controls and parametric uncertainty:

• The convexity of the log-barrier or corresponding QP adaptation law keeps the adapted weights or parameters strictly feasible;
• The composite barrier $H$, maintained via the adaptive law, satisfies for all times the CBF inequality under input constraints, which via Nagumo’s theorem or its generalizations, implies forward invariance of the consolidated safe set;
• In the uncertainty setting, boundedness of the composite energy $V_c$ implies $h(x) > 0$ for all $t$, i.e., forward invariance is preserved even if parameter convergence is not achieved (Kamaldar, 25 Jan 2026).

In cases of mixed relative-degree constraints, the structure of consolidated $H$ and control-affine weight dynamics allow satisfying the barrier condition without separate high-order CBF constructions (Black et al., 2023).

5. Synthesis of CaCBF-Based Quadratic Programs (QPs)

Safety and liveness controllers are synthesized using a QP with a single consolidated CBF constraint: $$u^* = \arg\min_{u \in U} \|u - u_{des}(t, x) \|^2 \quad \text{s.t.} \quad \frac{d}{dt} H(t, w(t), x) \ge -\alpha(H)$$ where the derivative incorporates both the system dynamics and the adaptation law for $w$ or $\hat{\theta}$ (Black et al., 2023, Black et al., 2022, Kamaldar, 25 Jan 2026).

In adaptive schemes addressing uncertainty, the QP also enforces CLF constraints for stabilization and an auxiliary variable $\delta$ for relaxation: $$\begin{array}{rl} \min_{u, \delta \ge 0} & \frac{1}{2} u^\top R(x) u + \rho \delta^2 \ \text{s.t.} & L_f h(x) + \psi(x) \hat{\theta} + L_G h(x) u \ge -\alpha(h(x)) \ & L_f V(x) + \phi(x) \hat{\theta} + L_G V(x) u \le -\lambda V(x) + \delta \end{array}$$ ensuring both safety and stability constraints are compatible and that the set of admissible controls remains convex (Kamaldar, 25 Jan 2026).

6. Applications and Empirical Studies

Significant implementation and empirical evaluation of CaCBF methods have been reported:

• Reach-avoid tasks with mixed constraints: In simulations of bicycle robot models subject to multiple static obstacle, speed, slip, and time-based reachability constraints (many of which have relative degree two), the CaCBF approach maintains all safety constraints and goal achievement, where standard high-order CBF methods fail due to feasibility loss (Black et al., 2023).
• Multi-robot systems: Laboratory and warehouse simulations with multiple ground robots, consolidating speed, collision, and corridor constraints. The consolidated weight adaptation ensures smooth, coordinated maneuvers without loss of feasibility or excessive chattering (Black et al., 2022).
• Parametric uncertainty: For adaptive cruise control, omnidirectional robots with unknown drift, and planar drones subject to unknown wind, CaCBF demonstrably reduces conservatism compared to worst-case robust CBFs, recovers nearly the full safe set, and enables safe operation near constraint boundaries without parameter convergence (Kamaldar, 25 Jan 2026).

The table below summarizes comparison regimes from (Kamaldar, 25 Jan 2026):

Scenario	CaCBF Performance	R-CBF Baseline
Adaptive cruise control	95% safe set, h_min>0	55% safe set, larger margins
Omnidirectional robot	Path/clearance ~33% shorter	Longer path, conservative
Planar drone through gate	Gap successful, safe passage	Gap blocked by margin

7. Limitations, Challenges, and Future Directions

• Input bounds: Early CaCBF methods required unbounded control; input bounds increase technical complexity, which has been partially addressed by embedding such constraints in the adaptive law and QP formulations (Black et al., 2023).
• Handling of high-order or nonlinear uncertainties: CaCBF currently certifies only linear-in-parameter uncertainty. Extensions to unstructured and nonlinear model uncertainties would require nonparametric function approximators and Lyapunov analysis (Kamaldar, 25 Jan 2026).
• Output feedback: All current guarantees assume full-state measurements; extension to output-feedback (with state estimation) is an open area.
• Recursive feasibility under large disturbances or initialization errors, and experimental validation in the presence of noise/disturbance, remain important topics for future investigation (Kamaldar, 25 Jan 2026).

A notable property is that the CaCBF approach allows compositionality and smooth adaptation of safety margins in real time, resulting in a system that is both less conservative and more robust to constraint geometry and uncertainty than robust or switched CBF approaches.

• (Black et al., 2023) "Consolidated Control Barrier Functions: Synthesis and Online Verification via Adaptation under Input Constraints"
• (Black et al., 2022) "Adaptation for Validation of a Consolidated Control Barrier Function based Control Synthesis"
• (Kamaldar, 25 Jan 2026) "Composite Adaptive Control Barrier Functions for Safety-Critical Systems with Parametric Uncertainty"