Safety-Critical Control Framework

• Safety-Critical Control Framework is a methodology that enforces system safety by maintaining safe set invariance through control barrier functions and real-time optimization.
• It integrates mathematical guarantees and optimization-based architectures, addressing uncertainties, delays, and multi-agent challenges with techniques like QPs and MPC.
• Applications in automotive, aerospace, and robotics validate its robust performance in enforcing safety under hard constraints despite dynamic and uncertain environments.

A safety-critical control framework is a structured methodology that synthesizes controllers for dynamical systems subject to hard safety requirements, ensuring the forward invariance of user-specified safe sets under model dynamics and environmental uncertainties. Such frameworks systematically encode safety constraints—typically via control barrier functions (CBFs), their generalizations, or other certificates—into real-time optimization-based control algorithms, frequently realized as quadratic programs (QPs) or model predictive control (MPC) schemes, while optionally mediating other performance objectives and addressing actuation, estimation, or input/output delay limitations.

1. Mathematical Foundations: Safe Set Invariance and Control Barrier Functions

Core to safety-critical control is the formal notion of set invariance, whereby the system trajectory is constrained to remain within a predefined safe set, typically specified as the zero-superlevel set of a smooth function. For a general nonlinear control-affine system

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

with a safe set

$\mathcal{C} = \{ x \in X \mid h(x) \geq 0 \}$

for a continuously differentiable $h: X \to \mathbb{R}$, safety-critical control seeks to synthesize feedback controls $u(x)$ such that $\mathcal{C}$ is forward invariant. This is typically certified via control barrier functions (CBFs), such that

$L_f h(x) + L_g h(x) u \geq -\alpha(h(x))$

holds for all $x \in \mathcal{C}$, where $L_f h$ and $L_g h$ are Lie derivatives and $\alpha$ is an extended class-$\mathcal{K}$ function; see (Ames et al., 2016).

CBFs have been extended to handle systems with higher relative degree, as high-order or zeroing control barrier functions (HOCBFs, ZCBFs), to input/output time delays via environmental CBFs (ECBFs) or control barrier functionals (CBFals) (Molnar et al., 2021, Kiss et al., 2022), and to multi-constraint or distributionally uncertain settings (Oh et al., 2023, Zhou et al., 2024).

2. Safety-Critical Control Synthesis: Optimization-Based Architectures

Most safety-critical control frameworks implement safety constraints by formulating and solving real-time optimization problems, most notably quadratic programs (QPs): $$\min_{u \in U} \ \| u - u_{\mathrm{nom}}(x) \|^2 \quad \text{s.t.} \quad L_f h(x) + L_g h(x) u \geq -\alpha(h(x))$$ where $u_{\mathrm{nom}}$ encodes tracking or performance objectives. The unification of safety and performance is systematically treated using QPs incorporating both CBF (safety) and control Lyapunov function (CLF, performance) constraints with appropriate slack variables (Ames et al., 2016, Xie et al., 2024). For multi-constraint environments, QP feasibility is preserved via hierarchy, slack prioritization, or the introduction of auxiliary certificates such as volume CBFs (VCBFs) (Dong et al., 18 Mar 2025) or Ultimate Invariant Sets (Oh et al., 2023).

Model predictive control (MPC) can embed CBF constraints at each prediction step, yielding recursive feasibility and robust invariance even under dynamics uncertainty, actuator constraints, or prediction errors (Bisgaard et al., 2021, Oh et al., 2023). Volume-based CBFs and robustified barrier constraints (with disturbance bounds or observer-based tightening) ensure constraint satisfaction under input/state uncertainties (Dong et al., 18 Mar 2025, Wang et al., 2021).

3. Extensions for Uncertainty, Partial Observability, and Delays

Modern frameworks have generalized safety-critical control to address:

• Environmental uncertainty and stochasticity: Adaptive conformal prediction modules coupled to probabilistic CBFs enforce high-probability safety despite unknown noise distributions, by calibrating deterministic safety margins from streaming nonconformity scores; the resulting tightened constraints are integrated into MPC (Zhou et al., 2024).
• Partial observability: Observer-based CBF schemes employ estimation-error-quantified observers and adaptive function approximation to tighten the barrier constraint, ensuring forward invariance despite bounded state estimation errors (Wang et al., 2021).
• Input and state delays: Delay compensation is achieved by integrating system/environment prediction over the delay interval. Environmental CBFs (ECBFs) are evaluated at predicted future states, and safety margins account for error bounds in prediction, as shown for input delays in robotics and vehicle systems (Molnar et al., 2021). For state delays, control barrier functionals over functional state spaces yield generalized affine safety constraints (Kiss et al., 2022).
• Disturbance and model-uncertainty robustness: Nonlinear disturbance observers (such as RISE) are combined with CBFs, and disturbance bounds are propagated into tightened barrier constraints to guarantee robust safety while maintaining feasibility (Dong et al., 18 Mar 2025). Volume CBFs ensure the invariance of feasible input domains even as multiple CBFs and actuator constraints become active.

4. Hierarchical, Learning-Augmented, and Model-Free Safety-Critical Control

To address infeasibility and tradeoff priorities among conflicting objectives:

• Hierarchical optimization: Multi-stage QP hierarchies first minimize slack in safety (CBF) constraints, then in performance (CLF) constraints, and only then optimize nominal action deviation, systematically ensuring feasibility and maximal enforcement of hard safety constraints (Xie et al., 2024).
• Learning-based frameworks: Model-free and learning-augmented safety-critical controllers employ data-driven residual correction around the CBF derivative, actively learning uncertainty dynamics to guarantee safety via an iteratively updated QP filter (Taylor et al., 2019). Transfer learning or adaptive aggregation strategies separate RL objective learning from safety enforcement using explicit backup controllers (safeguards) (Zhang et al., 2023).
• Rule-based and hybrid architectures: Parallel deployment of a learned controller and a rule-based safety controller, with high-frequency switching logic or supervised superposition, ensures hard safety guarantees at runtime by falling back on certified safe trajectories when the learned policy proposes a potentially unsafe action (Aksjonov et al., 2022).

5. Multi-Agent, Distributed, and Application-Specific Extensions

Frameworks have been extended to multi-robot or distributed scenarios:

• Authority distribution: Alternative Authority Control (AAC) assigns trajectory-planning authority to a single robot at each time step, treating others as dynamic obstacles, and rotates this assignment for fairness and scalability (Shi et al., 2024).
• Flexible and time-varying CBFs: For dynamic, shape-varying obstacles and moving agents, flexible CBFs explicitly incorporate obstacle orientation, velocity, and shape, adapting the safety set and control barrier in real time (Shi et al., 2024).
• Safety-critical locomotion: Application-specific frameworks, such as those for quadrupedal robots in highly structured environments, integrate multiple CBFs with trajectory generation, environment-aware foothold planning, and high-level behavior switching to ensure safety through hybrid discrete/continuous control (Lee et al., 2023).

6. Theoretical Guarantees, Tools, and Limitations

Safety-critical control frameworks universally emphasize provable forward invariance of the safe set, local existence and uniqueness of closed-loop dynamics, and recursive feasibility of the underlying QP or MPC scheme. Proofs rely on set-invariance theory (Nagumo's lemma), robust Lyapunov or barrier certificates, and, in multi-objective optimization, limit-weight equivalence or monotonicity properties. Tools such as QP solvers (e.g., OSQP, CVXGEN), nonlinear programmings solvers (e.g., IPOPT), and observer/estimator implementations are standard.

Limitations include:

• Sensitivity to hyper-parameter tuning, especially for tradeoff weights in integrated QPs and the calibration of robustness margins.
• Modeling assumptions (e.g., full-state observability or accurate projective prediction of environmental state).
• Potential for undesirable closed-loop equilibria (deadlock) near the boundary of multiple constraints; strategies such as adaptive CLF shaping or dual-QP approaches mitigate this (Reis et al., 2024).
• Computational scalability for embedded applications with large constraint sets or complex environment models.

7. Representative Applications and Case Studies

The safety-critical control framework paradigm is validated through an array of representative applications:

• Automotive: Adaptive cruise control (ACC) and lane keeping, with guarantees under actuator saturation, lead-vehicle braking, and multi-constraint interactions. Both simulation and hardware results confirm runtime QP feasibility and strict safety (Ames et al., 2016, Molnar et al., 2021, Bohara et al., 2023).
• Aerospace: Fixed-wing UAV collision avoidance under multiple state and input constraints via ultimate invariant set carving and one-step (recursively feasible) MPC (Oh et al., 2023).
• Robotics: Model-free safe-velocity synthesis in configuration space for legged, wheeled, and aerial robots (Molnar et al., 2021); neural and rule-based navigation stacks in cluttered and dynamic environments (Zheng et al., 3 May 2025).
• Multi-robot coordination: AAC and F-CBF methods for cooperative lane changes, search-and-rescue, and formation-keeping with dynamic obstacle avoidance (Shi et al., 2024).
• Reinforcement learning environments: Adaptive aggregation for safe transfer RL, with explicit safeguard modules ensuring constraint satisfaction at every learning phase (Zhang et al., 2023).

These examples demonstrate the generality, tractability, and robustness of contemporary safety-critical control frameworks across diverse domains and uncertainty regimes.