Generalizations of Backup Control Barrier Functions: Expansion and Adaptation for Input-Bounded Safety-Critical Control

1. Problem Statement

Ensuring the safety of nonlinear systems with bounded inputs is a fundamental challenge in autonomous control. While Control Barrier Functions (CBFs) provide a rigorous framework for guaranteeing forward invariance of safe sets, verifying a valid CBF for complex systems with tight input constraints is notoriously difficult.

Existing methods, specifically Backup Control Barrier Functions (bCBFs), address this by constructing a larger "implicit safe set" through the propagation of trajectories under a pre-verified backup controller. However, standard bCBF methods suffer from a critical limitation:

• Coupling Constraint: The same controller is used both to expand the safe set (by simulating forward trajectories) and to certify the invariance of the backup set.
• Conservatism: This coupling restricts the choice of the expansion controller. To ensure the backup set remains invariant, the expansion controller must often be conservative (e.g., low-gain), resulting in a smaller safe set and reduced system performance.

The paper aims to decouple these roles to allow for more aggressive set expansion while maintaining formal safety guarantees.

2. Methodology

The authors propose a Generalized Backup Control Barrier Function (GB-CBF) framework that separates the set-expanding controller ($k_e$) from the backup controller ($k_b$).

A. Generalized Backup Set Construction

Instead of using a single controller for both expansion and invariance, the method introduces a switching controller ($k_s$):
$$k_s(x) = (1 - \eta(x))k_e(x) + \eta(x)k_b(x)$$

• $k_b$ (Backup Controller): A pre-verified controller that renders the backup set ($C_B$) forward invariant.
• $k_e$ (Expanding Controller): A controller chosen specifically to maximize the expansion of the safe set. It does not need to render $C_B$ invariant on its own.
• $\eta(x)$ (Switching Function): A smooth function that transitions between controllers. Crucially, $\eta(x) = 1$ when $x \in C_B$, ensuring that once the system enters the backup set, it is governed solely by the verified $k_b$.

The expanded implicit safe set ($C_I^\star$) is defined as the set of states from which the system, under the switched flow $\phi_s$, remains within the primary safe set ($C_S$) for a horizon $T$ and reaches $C_B$ at time $T$.

B. Theoretical Guarantees

The authors prove that $C_I^\star$ is controlled invariant. They derive forward invariance conditions using the chain rule and sensitivity matrices ($\Phi$) of the switched flow. These conditions are formulated as constraints in a Quadratic Program (QP):

• GB-QP: Minimizes the deviation from a primary controller ($k_p$) subject to constraints ensuring the safety of the trajectory over the horizon $[0, T]$ and the terminal condition at $T$.
• Feasibility: The existence of at least one feasible controller (the switching controller $k_s$) is guaranteed, ensuring the QP is always solvable.

C. Adaptive Set Expansion

The framework is further extended to parameterized controller families ($k_e(x, \theta)$).

• Augmented State: The parameters $\theta$ are treated as auxiliary states with their own dynamics.
• Adaptation Law: An auxiliary control input ($u_\theta$) is designed to adapt $\theta$ in real-time. The adaptation direction is derived from the gradient of the cost function in an optimal control problem, specifically aiming to push the terminal state deeper into the backup set.
• Result: This allows the system to dynamically adjust its expansion strategy (e.g., tuning controller gains) to maximize the reachable safe set during operation.

3. Key Contributions

1. Decoupling of Roles: The primary contribution is the theoretical separation of the set-expanding controller from the backup controller. This allows the use of aggressive, high-performance controllers for expansion without compromising the formal safety certification of the backup set.
2. Generalized Invariance Conditions: The paper establishes sufficient conditions for the forward invariance of the generalized implicit safe set ($C_I^\star$) and formulates the corresponding QP (GB-QP) that handles input bounds.
3. Online Adaptation: The framework is extended to support online adaptation of the expansion controller parameters, enabling the safe set to grow dynamically as the system state evolves.
4. Recovery of Existing Methods: The authors demonstrate that standard bCBF methods are a special case of this generalized framework (where $k_e = k_b$).

4. Results and Case Studies

The paper validates the approach through two case studies:

• Case Study 1: Double Integrator

  • Setup: A 2D system with input saturation.
  • Comparison: The standard bCBF (using a conservative gain for both expansion and backup) vs. the GB-CBF (using a high-gain, time-optimal controller for expansion).
  • Outcome: The generalized method produced a significantly larger controlled invariant set ($C_I^\star$) compared to the standard method, demonstrating that decoupling allows for much less conservative safety regions.

• Case Study 2: Quadrotor Landing

  • Setup: A planar quadrotor must land in a narrow zone with obstacles, subject to strict thrust and moment limits.
  • Comparison:
    1. Standard bCBF: Failed to reach the goal because the conservative expansion could not navigate the tight constraints.
    2. Generalized bCBF (GB-QP): Successfully expanded the safe set further than the standard method but still struggled to reach the goal due to fixed expansion parameters.
    3. Adaptive bCBF (aGB-QP): By adapting the controller gains ($\theta$) online, the system achieved the largest expansion, successfully navigating the obstacles and reaching the goal.
  • Outcome: The adaptive approach demonstrated superior task performance while strictly adhering to input bounds and safety constraints.

5. Significance

This work represents a significant advancement in safe autonomy for systems with input constraints.

• Performance vs. Safety Trade-off: It resolves the traditional trade-off where safety guarantees force conservative behavior. By allowing different controllers for expansion and certification, systems can operate closer to their physical limits without violating safety.
• Scalability: The method scales to high-dimensional systems (as shown with the quadrotor) and integrates seamlessly with existing QP-based control architectures.
• Adaptability: The introduction of online adaptation suggests a path toward "learning" or optimizing safety certificates in real-time, making safety-critical control more robust to changing environments and system dynamics.

In summary, the paper provides a rigorous mathematical foundation for generalizing backup sets, proving that separating the expansion strategy from the safety certification strategy yields larger, less conservative safe sets and enables more capable autonomous behaviors.