Technical Summary: Policy Library Control Barrier Function (PL-CBF)

Problem Statement

The deployment of autonomous systems in unstructured, real-world environments faces significant challenges regarding online safety certification. Unlike classical settings with fixed constraints and fully modeled parameters, deployed robots operate under perception-limited information where safe areas are revealed online, dynamic obstacles appear and disappear, and operating conditions (e.g., friction) change over time.

While formal verification tools like Hamilton-Jacobi (HJ) reachability analysis offer strong guarantees, they suffer from the curse of dimensionality and are computationally intractable for high-dimensional robotic systems. Control Barrier Functions (CBFs) provide a scalable alternative but typically require an explicit barrier function for the full nonlinear system under fixed constraints, which is difficult to obtain when constraints evolve online.

Existing runtime safety filters based on rollouts, such as Policy CBFs (PCBFs), attempt to construct safety certificates from trajectory rollouts without offline value-function computation. However, these methods rely on a single designated fallback policy. In unstructured environments, no single maneuver is reliably safe across all perception updates. For instance, a stop maneuver may fail under sudden friction loss, or a fixed evasive maneuver may become infeasible due to newly detected obstacles. This dependence on a single fallback induces unnecessary conservatism and can lead to safety violations even when alternative safe maneuvers exist.

Methodology: Policy Library Control Barrier Function (PL-CBF)

The authors propose PL-CBF, a runtime safety filter that evaluates a finite library of diverse fallback policies via parallel finite-horizon rollouts to certify safety and select the least invasive safe mode.

Core Formulation

Instead of relying on a single backup policy, PL-CBF defines a library of $K$ candidate fallback policies $\Pi = \{\pi_1, \pi_2, \dots, \pi_K\}$. It constructs a finite-horizon policy library value function:
$$H^\Pi_T(x) := \max_{\pi_i \in \Pi} \left\{ \inf_{\tau \in [0,T]} h(\phi^{\pi_i}_\tau(x)) \right\}$$
where $h$ is the constraint function defining the safe set. If at least one policy in the library maintains non-negative constraint values over the horizon $T$, the state is considered safe.

Safety Filter Synthesis

The safety filter operates by solving a Quadratic Program (QP) that minimally modifies a nominal control input $u_{nom}$ while ensuring safety.

1. Parallel Rollouts: The system performs parallel rollouts for all policies in $\Pi$ to compute their respective safety values $H^{\pi_i}_T(x)$.
2. Policy Selection: Among the certified safe candidates (where $H^{\pi_i}_T(x) > 0$), the filter selects the "least invasive" policy $\pi_{i^*}$. Unlike methods that simply pick the highest safety margin, PL-CBF selects the policy that maximizes the volume of the admissible control set $K_{cbf}(x; \pi_i)$. This is defined as:
   $$i^* = \arg \max_{i \in I_{safe}(x)} \text{Vol}(K_{cbf}(x; \pi_i))$$
3. QP Execution: The final control input $u^*$ is obtained by solving:
   $$u^* = \arg \min_{u \in U} \|u - u_{nom}\|^2_2 \quad \text{s.t.} \quad \nabla H^{\pi_{i^*}}_T(x)^\top(f(x) + g(x)u) \geq -\alpha(H^{\pi_{i^*}}_T(x))$$
   Gradients are computed efficiently at runtime using automatic differentiation through the dynamics.

Theoretical Analysis

The paper establishes a completeness condition for finite-horizon safety certification using a finite-horizon language metric. This metric measures the worst-case state deviation between two policy rollouts over a horizon $T$. The authors prove that if the policy library $\Pi$ has a sufficiently low approximation precision (i.e., the library covers the space of admissible behaviors closely enough), and if there exists a strictly safe policy with a non-zero clearance, then the library is guaranteed to contain a policy that PL-CBF can certify as safe.

The framework also discusses the conceptual connection to classical CBFs and forward invariance in an idealized infinite-horizon limit, noting that if the library contains a valid backup policy, PL-CBF recovers the recursive feasibility guarantees of Backup CBF approaches.

Key Contributions

1. Policy-Library Finite-Horizon Safety Filtering: The introduction of PL-CBF, which certifies safety by evaluating a finite library of candidate fallback policies via parallel rollouts. It selects the least restrictive constraint by maximizing the volume of the admissible control set among certified candidates.
2. Theoretical Characterization: The establishment of a completeness condition for finite-horizon safety certification based on a finite-horizon language metric over closed-loop behaviors. This characterizes the policy-library coverage requirements necessary to certify safety.
3. Empirical Validation and Runtime Efficiency: Demonstration of improved safety coverage over single-policy safety filters across three simulation benchmarks (planar double-integrator, highway driving with abrupt friction changes, and 3D quadrotor navigation) while maintaining millisecond-level runtime.

Experimental Results

The authors evaluated PL-CBF against four representative safety filters (Model Predictive Shielding, Gatekeeper, Backup CBF, and Policy CBF) using single fallback policies.

• Planar Double-Integrator: PL-CBF demonstrated the ability to recover the largest safe set by evaluating a library of stopping and evasive policies, whereas single-policy methods recovered different, often smaller, safe sets depending on the chosen fallback.
• Highway Driving (Sudden Friction Change): In a scenario with a realistic nonlinear vehicle model (8 states) and abrupt friction changes (black ice), PL-CBF achieved 0/50 failures at a target speed of 10 m/s. In contrast, all single-fallback baselines suffered collision or infeasibility rates ranging from 22% to 92%. PL-CBF maintained a runtime of 7.5 ms per step.
• 3D Quadrotor Navigation: In a crowded dynamic environment (12 states), single-policy baselines failed in 68% to 79% of trials. PL-CBF reduced failure rates to 0% when using a library of 65 maneuvers (64 evasive + 1 nominal), with a runtime of 27.9 ms. The paper notes that a Mixed-Integer MPC baseline with similar maneuver selection was significantly slower (688 ms).

Significance and Claims

The paper claims that PL-CBF addresses the brittleness of relying on a single fixed fallback maneuver in unstructured environments. By retaining multimodal safe maneuvers through a library and selecting the least invasive option online, the method improves safety coverage without sacrificing real-time performance.

The authors position PL-CBF as a practical, deployable high-rate safety layer. They explicitly state that the method avoids the computational intractability of HJ reachability and the offline synthesis requirements of traditional CBFs. The significance lies in its ability to handle evolving constraints and perception updates where a single maneuver cannot guarantee safety, thereby reducing conservatism and preventing safety violations that could occur with single-policy filters.

The paper concludes modestly, noting that future work will focus on systematic policy library construction using learning-based approaches, handling model mismatch and perception uncertainty, and adaptive mechanisms to refine the library online.