Technical Summary: OcclusionCBF for Safe Navigation Among Hidden Dynamic Obstacles

Problem Statement

Autonomous robots often operate with incomplete environmental information due to occlusions caused by static structures (e.g., buildings, parked vehicles) or dynamic agents. A controller that reacts only to currently detected obstacles may steer a robot into a state where, once a hidden obstacle becomes visible, no admissible control input can prevent a collision. This failure arises not merely from sensing latency but from the interaction between uncertain hidden-obstacle occupancy, robot dynamics, and finite control authority.

Existing approaches, such as visibility-aware planning or set-based verification, often embed safety reasoning directly into trajectory optimization or fail to model the worst-case future motion of unseen dynamic obstacles. While effective, these methods can couple safety to specific planning architectures or require solving complex trajectory optimizations at every control update. The challenge is to construct a modular safety filter that explicitly accounts for robot dynamics and input constraints while reasoning about hidden obstacles, without being tied to a specific planner.

Methodology: OcclusionCBF

The paper proposes OcclusionCBF, a safety filter that extends the Backup Control Barrier Function (CBF) framework to handle time-varying reachable-occupancy predictions for potentially hidden dynamic obstacles. The method operates as a minimally invasive filter interposed between an arbitrary nominal controller and the system dynamics.

Core Components

1. Reachable-Occupancy Prediction:

   • The system models the set of workspace positions where an undetected obstacle may be located, denoted as $\mathcal{O}(P(x(t)), t)$.
   • This set is represented as a union of occluded regions. For each region, a collision-inflated reachable-occupancy set $\hat{H}^{(j)}(t, s)$ is constructed. This set accounts for the obstacle's maximum speed bound $\bar{v}_o$, the elapsed time since the last update, and a collision inflation radius $r_{col}$.
   • The occupancy prediction satisfies a temporal consistency condition: as the prediction horizon shifts, previously certified sets do not arbitrarily enlarge, ensuring the safety certificate remains valid.

2. Polyhedral Approximation and Margins:

   • To ensure differentiability for CBF formulation, the occluded regions are enclosed by convex polytopes.
   • A differentiable separation margin $h^C_j(y, t, s)$ is derived from the facet margins of these polytopes using a shifted log-sum-exp function. This margin ensures the robot's position remains outside the interior of the inflated occupancy set.
   • The margin satisfies a specific temporal derivative condition ($\frac{\partial h^C}{\partial t} - \frac{\partial h^C}{\partial s} \geq 0$) to maintain consistency as the time horizon advances.

3. Backup Rollout and Terminal Set:

   • The method defines a backup policy $\pi_b$ (e.g., a saturated lane-keeping or PD controller) that generates a backup trajectory $\phi_b(x, s)$ over a finite horizon $T$.
   • A time-varying terminal set $S_0(q)$ is defined such that if the backup trajectory reaches this set at time $T$, it guarantees a safe state with additional clearance $\rho_T$.
   • The terminal set is verified to be forward invariant under the backup policy.

4. Occlusion-Aware QP (OCBF-QP):

   • The safety filter solves a Quadratic Program (QP) to minimally modify the nominal input $u_{nom}$.
   • Constraints: The QP enforces affine constraints derived from differentiating the occupancy margins and terminal margins along the backup flow.
     • Occupancy Constraints: Ensure the backup trajectory remains outside the time-varying reachable occupancy sets for all $s \in [0, T]$.
     • Terminal Constraints: Ensure the backup trajectory terminates within the verified terminal set.
   • The constraints are affine in the control input $u$, allowing for efficient real-time computation.

Key Contributions

The paper outlines three primary contributions:

1. Planner-Agnostic Safety Filter: The extension of Backup CBFs to conservative reachable-occupancy predictions. This yields a filter that certifies a backup maneuver for hidden dynamic obstacles without embedding safety logic into a specific trajectory optimizer.
2. Theoretical Guarantees: The derivation of input-affine rollout and terminal constraints for a moving prediction horizon. The authors prove that the backup policy certifies recursive feasibility of the QP and guarantees collision avoidance for every obstacle motion covered by the occupancy prediction.
3. Smooth Polyhedral Realization: A method to handle multiple occluded regions using smooth polyhedral approximations. This enables the filter to operate with millisecond-scale computation while achieving high success rates in randomized benchmarks, MetaUrban simulations, and hardware experiments.

Results

The performance of OcclusionCBF was evaluated against five baselines: reactive CBF-QP, OA-MPC, Control-Tree MPC, Single-Hypothesis MPC, and OACP.

• Randomized Benchmarks:

  • In planar simulations with double-integrator and unicycle robots, OcclusionCBF achieved the highest hard-constrained success rate across all tested obstacle densities.
  • At high densities, it succeeded in 86% of unicycle trials and 70% of double-integrator trials, whereas other methods frequently failed due to collisions or controller infeasibility.
  • Computation Time: OcclusionCBP operates at millisecond scales (1.6–4.5 ms per update), significantly faster than predictive planning baselines (which range from ~7 ms to over 100 ms) and only slightly slower than reactive CBF-QP.

• High-Fidelity Simulation (MetaUrban):

  • Demonstrations with unicycle and quadruped models showed the ability to intervene proactively before obstacles become visible, successfully completing navigation tasks in complex urban environments.

• Hardware Experiments:

  • A "blind crossing" scenario was tested using TurtleBot robots. The baseline reactive CBF-QP, which ignores hidden occupancy, collided with a hidden obstacle after detection. In contrast, OcclusionCBF intervened before detection, slowing down and steering away from the occluded region to reach the goal safely.

Significance and Claims

The paper claims that OcclusionCBF addresses a critical gap in safe navigation: the inability of reactive controllers to handle the "state of no return" caused by hidden dynamic obstacles. By decoupling safety certification from the planning architecture, the method provides a modular, computationally efficient, and theoretically rigorous solution.

The authors emphasize that their approach does not merely react to sensing latency but proactively reasons about the interaction between uncertainty, dynamics, and control authority. The method guarantees that as long as the occupancy prediction covers the true obstacle motion, the robot will remain in a recoverable set and avoid collisions, even if the obstacle remains hidden until the last moment. The results demonstrate that this level of safety can be achieved without the computational burden of online branched trajectory optimization, making it suitable for real-time robotic applications.