Predictive Control Barrier Functions (PCBF)

• Predictive Control Barrier Functions (PCBFs) are optimization-based safety mechanisms that integrate MPC principles with slack-minimization to ensure robust set invariance and asymptotic stability.
• They leverage finite-horizon predictions to compute a value function that quantifies minimal constraint violations, enabling recovery from disturbances and initial infeasibility.
• PCBFs facilitate multiobjective control by embedding safety constraints into performance-driven MPC, with applications in space rendezvous and autonomous driving.

A Predictive Control Barrier Function (PCBF) is an optimization-based barrier function that encodes safety and invariance properties for constrained control systems by leveraging finite-horizon predictions and slack-minimization, providing strong guarantees of set-invariance and robust asymptotic stability for safety-critical applications. Unlike pointwise control barrier functions, PCBFs define a value function (the optimal cost of a slack-augmented predictive optimal control problem) whose properties intrinsically encode the ability of the system to maintain safety and return to a safe set despite disturbances, initial infeasibility, or approximation errors. This construction generalizes standard barrier-function techniques while supporting integration with modern Model Predictive Control (MPC) and learning-based filters.

1. Mathematical Formulation and Definition

A PCBF is defined through an auxiliary finite-horizon optimal control problem that minimizes constraint violations (slacks) over the entire horizon, favoring trajectories that strictly satisfy constraints but providing an always-feasible solution via relaxation. In its canonical discrete-time form for a nonlinear system $x_{k+1} = f(x_k, u_k)$ with hard input constraints $u_k\in U$ and state constraints $x_k\in X$, the PCBF value function $B(x)$ is defined as

$$B(x) = \min_{\{u_i\}_{i=0}^{N-1},\{\xi_i\}_{i=0}^{N}} \alpha_f\,\xi_N + \sum_{i=0}^{N-1} \|\xi_i\|_1$$

subject to

$$\begin{aligned} x_0 &= x \ x_{i+1} &= f(x_i, u_i), \quad i = 0, ..., N-1 \ u_i &\in U, \quad x_i\in X(\xi_i) := \{c_x(x_i) \le -\Delta_i + \xi_i\}, \quad \xi_i \ge 0 \ h_f(x_N) &\le \xi_N, \quad \xi_N \ge 0 \end{aligned}$$

where $\Delta_i$ are constraint-tightening sequences, $h_f$ is a terminal control barrier function, and $\alpha_f$ is a large weight on the terminal slack. The zero sub-level set

$S = \{x \mid B(x) = 0\}$

coincides with initial states from which strictly constraint-satisfying admissible trajectories exist within the imposed tightenings and terminal safe set. For $B(x) > 0$, $B(x)$ quantifies the minimal cumulative violation required along the horizon, hence acting as a barrier certificate.

This framework is connected to the value function of MPC with safety constraints, as in "Predictive Control Barrier Functions: Bridging model predictive control and control barrier functions" (Huang et al., 12 Feb 2025), where invariance and constraint satisfaction are established via forward propagation of feasible plans under receding horizon.

2. Invariance and Decrease Conditions

The invariance property is established via a set-wise decrease condition on $B(x)$:

• For all $x \in D\backslash S$ (where $D$ is an enlarged level set, e.g. $\{x \mid B(x) \leq \alpha_f\,\gamma_f\}$), there exists $u\in U$ such that

$$B(f(x, u)) - B(x) \le -\alpha_3( \mathrm{dist}(x, S) )$$

where $\alpha_3$ is a class-$\mathcal{K}$ comparison function.

• For $x \in S$, one can find $u \in U$ such that $B(f(x, u)) \le 0$.

This guarantees robust forward invariance and asymptotic stability of $S$, i.e., under repeated application, the trajectory remains in $D$, converges to $S$, and satisfies all constraints as long as $B(x(k)) = 0$ (Didier et al., 2024, Wabersich et al., 2021).

An important consequence is that, unlike classical CBFs which typically enforce decrease at each step of a predicted trajectory, the PCBF is concerned with the value function decrease relative to the last feasible trajectory, which can be enforced in an MPC setting using a warm-start based on the previous solution (Didier et al., 25 Mar 2025).

3. Robustness and Recovery Mechanisms

PCBFs provide robust recovery and feasibility guarantees even under disturbances, initialization outside $S$, or minor deviations due to approximations:

• The slack variables ensure (soft) feasibility of the underlying optimal control problem at every step, allowing the system to "recover"—drive $B(x(k)) \to 0$—even from constraint violations (Wabersich et al., 2021).
• With suitable tuning (e.g., large $\alpha_f$ for terminal slack), one can ensure that the value function is continuous, enabling the construction of recovery mechanisms that iteratively converge to the strict-feasibility set $S$.
• Main theoretical results ensure input-to-state stability (ISS) of $B(x)$ under bounded disturbances or approximation errors (e.g., due to learning-based surrogates), such that invariance is preserved up to a tube proportional to the error (Didier et al., 2024, Didier et al., 2022).

4. Integration with Performance Objectives and Multiobjective MPC

A major advancement of the PCBF framework is the ability to embed safety invariance into multiobjective optimal control, balancing task objectives and safety guarantees:

• The value function $B(x)$ can be included as a constraint (imposing a minimal decrease condition) within a broader MPC cost, allowing optimization of general control objectives while provably enforcing robust asymptotic stability of the desired safe set (Didier et al., 25 Mar 2025).
• The decrease constraint is computed with respect to a "warmstart" value from the prior feasible plan, ensuring that the controller always optimizes performance conditional on progress toward/within the robust safe set.
• Typical applications include linear space rendezvous and nonlinear lane-changing problems, where performance and safety invariance must be balanced over time.

5. Algorithmic Structure and Implementation

The PCBF methodology admits several algorithmic instantiations, with a common template:

• At each step, solve the PCBF slack-minimization OCP to obtain $B(x(k))$ and associated slack-optimal trajectory.
• Enforce a decrease constraint (via a small-scale QP or within a full-horizon MPC) where the chosen input guarantees $B(x(k+1)) \le B(x(k)) - \alpha_3( \cdot )$ (or is less than a warmstart value), thereby maintaining invariance.
• Warmstart the solver with the previous solution to ensure recursive feasibility and computational efficiency.
• For approximate PCBFs (e.g., via neural network surrogates), the decrease constraint and resulting safety filter can be implemented as a single-step QP with explicit error-margin compensation (Didier et al., 2024, Didier et al., 2022).

6. Theoretical and Practical Significance

Key properties and motivations of PCBFs include:

• Guaranteeing robust forward invariance and (under suitable conditions) robust asymptotic stability of the safe set defined by the zero-level of $B(x)$.
• Enabling recovery from infeasible states by ensuring the safety filter remains always feasible (through slack augmentation) and returning trajectories to the feasible set (Wabersich et al., 2021).
• Preserving computational tractability in high-dimensional scenarios, particularly via value-function approximation (Didier et al., 2024, Didier et al., 2022).
• Providing a systematic approach to integrate with model-based, learning-based, or data-driven controllers, and facilitating multi-objective trade-offs.

7. Application Examples and Empirical Validation

Empirical evaluations and application case studies highlight the efficacy of PCBFs:

• For space rendezvous and nonlinear autonomous driving problems, embedding the PCBF decrease constraint within the MPC allows tight constraint satisfaction and robust set stabilization, outperforming constraint softening or penalty-based alternatives (Didier et al., 25 Mar 2025).
• In high-speed predictive control of nonlinear systems, maintaining the decrease condition enables provable forward invariance and recovery without excessive conservatism or computational burden.
• Neural network surrogates for the PCBF value function yield significant speed-ups and scalability, with quantifiable approximation-induced relaxation of invariance (Didier et al., 2024, Didier et al., 2022).

• "A multiobjective approach to robust predictive control barrier functions for discrete-time systems" (Didier et al., 25 Mar 2025)
• "Approximate predictive control barrier function for discrete-time systems" (Didier et al., 2024)
• "Predictive control barrier functions: Enhanced safety mechanisms for learning-based control" (Wabersich et al., 2021)
• "Approximate Predictive Control Barrier Functions using Neural Networks: A Computationally Cheap and Permissive Safety Filter" (Didier et al., 2022)
• "Predictive Control Barrier Functions: Bridging model predictive control and control barrier functions" (Huang et al., 12 Feb 2025)