Composite adaptive control barrier functions couple parameter estimation with safety in one energy function, recovering in three simulations the safe operating space robust methods give up

A composite energy function Vc unifying safety, stability, and estimation is constructed, and the parameter adaptation law is derived directly from it, creating a direct coupling between estimation accuracy and the safety margin. Prior adaptive CBFs largely adopt a modular structure in which the adaptation law (e.g., recursive least squares or gradient descent) remains agnostic to the barrier function, reducing parameter error globally; this work sums a logarithmic barrier potential, a control Lyapunov function, and a quadratic parameter-error term and derives an update law that strictly dissipates the total energy, so that near the boundary the effective learning rate prioritizes the parameters needed to maintain safety. The derivation is given in Section 3, from Vc in (13) through (17)–(25) to the dissipation inequality; the projection operator satisfies ˜θ^T Γ^{-1}(PΘ(τ)−τ) ≤ 0 (Lemma 1) and confines ˆθ(t) ∈ Θ (Lemma 2).

The safe set C is proved forward invariant for all bounded parameters, without requiring persistence of excitation, and for any κ > 0. Robust methods maintain safety via worst-case bounds at a performance cost, while modular learning can violate constraints during transients; this work provides a safety proof independent of the classical Theorem 1, using boundedness of Vc to force h(x(t)) > 0, thereby guaranteeing safety even without persistence of excitation. Theorem 3 uses a finite-time contradiction argument in which κ appears only in the finite bound cα + κδmax; Remark 5 states explicitly that forward invariance holds for any κ > 0.

The safety guarantee is proved unconditionally robust to bounded errors in the state-derivative measurement, and all closed-loop signals are uniformly ultimately bounded. Prior methods typically require noise to be sufficiently small or rely on parameter convergence; Theorem 5 preserves forward invariance for any bounded noise ¯w, and Theorem 4 gives UUB under the explicit computable condition κ > cα/ε (54), with Corollary 1 giving the sufficient bound κ > α0λ/(2A²). The proof of Theorem 5 introduces the noise term cw ≜ 2γθmax c̄F ¯w in (71) and states that ¯w need not be sufficiently small; Theorem 4 proves boundedness via forward invariance of the sublevel set Ωs.

The CaCBF admissible control set is proved to always contain the robust CBF admissible control set, quantifying the reduction in conservatism. Robust CBF shrinks the safe control set using a static worst-case margin σrob, which can render the feasible set empty; this work proves Urob(x, θe) ⊆ Uadp(x, ˆθ), so the adaptive QP is always at least as feasible as the robust one, with the gap growing as the estimate moves away from the worst-case direction. Theorem 6 gives an algebraic proof of the set inclusion; Remark 13 notes the inclusion holds for all ˆθ ∈ Θ as a structural property of the constraint geometry; in Example 3 the R-CBF fails entirely because the inflated margins overlap, while CaCBF traverses the gate.

The results target nonlinear control-affine systems with linear parametric uncertainty and relative degree 1, requiring a known parameter bound θmax, a non-empty safe set with the equilibrium in its interior (h(0) > 0), and constant parameters over the operational horizon. They apply to engineering systems well modeled this way, such as adaptive cruise control, omnidirectional robots, and planar drones, and provide a design path for optimization-based safety filters using arbitrary nominal controllers. For high-order relative-degree constraints, the paper constructs extended barrier functions via backstepping to recover relative degree 1, so the theorems apply directly to the extended system.

The adaptation law relies on the state derivative ˙x, approximated by finite differences with bounded error, but the effect of sampling period and delay on the convergence rate σ requires a sampled-data analysis and is left as future work; under persistent excitation the parameter error converges exponentially, yet under bounded noise it converges only to a residual ball, with exact convergence requiring ¯w = 0. The constant-parameter assumption needs a leakage modification for slowly time-varying cases. The current analysis assumes a fixed safe set; time-varying safe sets in multi-agent and moving-obstacle scenarios remain open. Hardware validation—where noise, latency, and actuator saturation occur simultaneously—remains an open question for deployment readiness.