Adaptive Control Barrier Functions

• Adaptive Control Barrier Functions are a safety framework that uses online adaptation and barrier constraints to maintain forward invariance in systems with uncertainty.
• They integrate parameter estimation, adaptive gain modulation, and auxiliary variables to dynamically adjust the safe set and ensure QP feasibility.
• Applications include impulsive robotic manipulators, multi-agent formations, and autonomous vehicles, demonstrating low conservatism and robust safety in real time.

Adaptive Control Barrier Functions (aCBFs) are a principled framework for enforcing safety in dynamic systems subject to parametric uncertainty, disturbances, output estimation errors, and nonideal events such as impulsive state jumps. By coupling classical Control Barrier Functions (CBFs) with online adaptation mechanisms—including parameter estimation, auxiliary variable dynamics, and gain modulation—these methods dynamically shape the barrier constraint to maintain forward invariance of the safe set, often under challenging model and environmental conditions. Modern aCBF frameworks leverage quadratic program (QP) synthesis, observer models, data-driven parameter identification, and auxiliary optimization policies to guarantee safety, feasibility, and low conservatism in real time.

1. Mathematical Foundations

aCBFs generalize the classical CBF condition

$$\sup_{u\in \mathcal{U}}\left\{L_f h(x) + L_g h(x) u\right\} \geq -\alpha(h(x))$$

by allowing the barrier function $h$ and/or the safety constraint to depend on time-varying or estimated parameters, auxiliary variables, or system states affected by uncertainty (Panja, 2024). A typical control-affine model under uncertainty is

$\dot{x} = f(x) + F(x)\theta^* + g(x)u + d(x, t)$

where $\theta^*$ is unknown but constant, $d(x, t)$ is bounded disturbance.

The adaptive safety constraint is often formulated as

$$\sup_{u\in\mathcal{U}} \left\{ L_f h_a(x, \hat{\theta}) + L_F h_a(x, \hat{\theta}) \lambda_{cbf}(x, \hat{\theta}) + L_g h_a(x, \hat{\theta}) u \right\} \geq -\alpha(h_a(x, \hat{\theta}))$$

where $\lambda_{cbf}(x, \hat{\theta})$ incorporates adaptive correction (e.g., $$\lambda_{cbf} = \hat{\theta} + \Gamma \partial_{\theta} h_a(x, \hat{\theta})^T$$ as in (Taylor et al., 2019)).

For impulsive systems,

$$\dot{x}(t) = f(x(t)) + g(x(t))u(t), \quad x(t_k^+) = x(t_k^-) + p(x(t_k))$$

the safe set $\mathcal{C}$ must be invariant both under continuous flow and at state jumps, enforced via CBF conditions at both regular and impulse times (Liu et al., 13 Mar 2025).

2. Adaptive Mechanisms

Modern aCBF approaches employ several adaptation strategies, summarized in the table below.

Mechanism	Description	Reference
Parameter estimation	Online update of $\hat{\theta}$ via RLS, batch LS, or observer	(Gutierrez et al., 2024, Shen et al., 2024)
Adaptive gain modulation	State-dependent or event-dependent gains $\kappa(x)$	(Liu et al., 13 Mar 2025, Chriat et al., 2023)
Auxiliary variables	Time-varying $\alpha_i(t)$/virtual states to modify constraints	(Liu et al., 20 Feb 2025, Xiao et al., 2020)
Observer-state estimation	Safety margins compensate for estimation error in $\hat{x}$	(Wang et al., 2021)

Parameter estimation is typically achieved via least-squares or recursive algorithms, ensuring boundedness or convergence under persistency-of-excitation conditions. Adaptive gains dynamically increase safety correction near the safe set boundary or in response to state impulses. Auxiliary variables (e.g., as unbounded optimization variables in QP) allow real-time expansion of the feasible control set, particularly when classic CBF constraints become conflicting under actuator saturation or state constraints.

3. Quadratic Program Synthesis

Adaptive CBF constraints are operationally enforced by embedding them in convex QPs. A general form is

$$\min_{u \in \mathcal{U}} \: \frac{1}{2} \|u - u_{\text{nom}}(x, \hat{\theta})\|^2$$

subject to one or more adaptive CBF inequality constraints,

$$L_f h_a(x, \hat{\theta}) + L_F h_a(x, \hat{\theta}) \lambda_{cbf} + L_g h_a(x, \hat{\theta}) u \geq -\alpha(h_a(x, \hat{\theta}))$$

and possibly control bounds or auxiliary slack variables. At impulse times, the QP center is shifted by an adaptive gain proportional to jump magnitude (Liu et al., 13 Mar 2025). For multi-step/discrete-time safety, the QP constraints may include multi-step recursions or risk-aware expectations (via CVaR measures) (Liu et al., 24 Mar 2025).

Auxiliary-variable adaptation (AVCBF) techniques guarantee feasibility by making all control components enter the highest-order constraint via time-varying auxiliary states, with hyperparameters tuned via sensitivity analysis or real-time optimization (Liu et al., 20 Feb 2025).

4. Theoretical Guarantees and Safety Analysis

Nearly all modern aCBF frameworks prove forward invariance of the adaptive safe set—i.e., $x(t) \in \mathcal{C}$ for all $t$—under mild initializations and regularity (e.g., Lipschitz continuity, control authority near the safe set boundary). Typical conditions include:

• Boundedness and uniform ultimate boundedness of parameter estimates or adaptive variables;
• Feasibility and continuity of the QP solution for all admissible states and control bounds;
• Nonincreasing conservativeness of the adaptive safety margin, which approaches nominal as uncertainty vanishes (under PE) (Gutierrez et al., 2024);
• Piecewise Lipschitz regularity and bounded jumps for closed-loop control under impulsive dynamics (Liu et al., 13 Mar 2025);
• Exclusion of Zeno behavior in event-triggered adaptation schemes (Shen et al., 2024).

Theoretical results often rely on constructing composite Lyapunov or barrier functions which combine adaptive parameter and state terms, showing exponential or monotonic decay, and applying forward invariance/comparison lemmas.

5. Feasibility, Conservatism, and Robustness

Adaptive mechanisms in aCBF frameworks directly address feasibility constraints in the safety QP: tuning auxiliary variables, penalty functions, or gain schedules ensures that the QP remains solvable even as time-varying bounds or actuator saturations cause conflicts for classical formulations (Xiao et al., 2020, Liu et al., 20 Feb 2025). Risk-aware aCBFs incorporate coherent risk measures (e.g., CVaR) that mediate between worst-case conservatism and practical average-case safety (Liu et al., 24 Mar 2025).

Set-membership identification (SMID) and batch LS estimation reduce conservatism in the adaptive margin by refining parameter bounds as new data arrives (Gutierrez et al., 2024, Lopez et al., 2020, Kim et al., 17 Jun 2025). Adaptive neural networks can learn model drift online, and their estimation error bounds propagate directly into the safety margin computation (Sweatland et al., 2024).

6. Applications and Case Studies

Adaptive CBFs have demonstrated efficacy across a breadth of domains:

• Impulsive robotic manipulators: QP-adaptive CBFs show zero safety violations and fast recovery in the face of state jumps (Liu et al., 13 Mar 2025).
• Multi-agent formation: Distributed aCBF controllers integrating reference tracking and safety filters guarantee obstacle avoidance and convergence in leader-follower graphs (Solano-Castellanos et al., 2024).
• Stochastic control: Risk-aware multi-step aCBFs with auxiliary variables preserve both feasibility and robustness under tight bounds and disturbance propagation (Liu et al., 24 Mar 2025).
• Time-varying robotic tasks: RaCBFs with ISSf and SMID enforce dynamic force constraints for quality assurance in surface treatment, with experimentally verified reductions in conservatism (Kim et al., 17 Jun 2025).
• Cruise control and obstacle avoidance: AVCBFs, penalty-based AdaCBFs, and learning-based aCBFs maintain feasibility and safety under vehicle dynamics noise and rapidly-varying actuation limits (Liu et al., 20 Feb 2025, Xiao et al., 2020, Sweatland et al., 2024).
• Real-time adaptive parameter estimation: Triggered batch LS safety adaptation achieves finite-time convergence and safety with provable exclusion of Zeno phenomena (Shen et al., 2024).

These applications highlight the flexibility of aCBF frameworks across continuous, discrete, impulsive, stochastic, and multi-agent environments.

7. Limitations and Directions for Research

Current aCBF approaches generally assume linearly parameterized uncertainties, known regressor structures, and enforce feasibility via convex optimization or adaptive relaxation terms. Known limitations include:

• Potential infeasibility under conflicting or overly conservative constraints (especially in high-dimensional or nonlinear-parameter environments) (Panja, 2024);
• Manual tuning of gains and adaptation policies, often requiring expert knowledge of system dynamics (Taylor et al., 2019);
• Lack of nonparametric uncertainty handling and limited integration with model-free learning;
• Assumption of persistency of excitation for parameter convergence;
• The need for improved event-triggered or batch adaptation schemes to handle switching, sampling, and intermittent feedback loss (Sweatland et al., 2024, Shen et al., 2024).

Recent work explores data-driven and reinforcement learning approaches to shape class-$\mathcal{K}$ functions adaptively through end-to-end training (Chriat et al., 2023), offers nonconservative safety under PE (Gutierrez et al., 2024), and presents solutions for mixed relative degree or nonuniform input mappings (Liu et al., 20 Feb 2025).

This suggests a trajectory for ongoing research: integrating aCBF frameworks with nonlinear regression or Gaussian process models, developing scalable tuning algorithms, and formalizing guarantees under switched or networked control topologies.