Occlusion-Agnostic CBF-QP: Reacts too late after the hidden vehicle becomes visible.
OcclusionCBF: Acts before detection and safely completes the crossing.
Robots navigating under occlusion may detect dynamic obstacles only after it is too late to avoid them. OcclusionCBF addresses this by reasoning about where potentially hidden obstacles may move before they become visible. OcclusionCBF is a safety filter that certifies a backup rollout against reachable hidden-obstacle occupancy and minimally modifies the nominal control input only when needed.
Motivation
Occlusion-Agnostic CBF-QP: Reacts only after the hidden obstacle becomes visible, when it is already too late to avoid collision.
OcclusionCBF: Accounts for potentially hidden dynamic obstacles and intervenes before detection to safely navigate the crossing.
Conventional CBFs react to obstacles after they are detected. Under occlusion, this may be too late: by the time a hidden (dynamic) obstacle becomes visible, the robot may no longer have enough control authority to avoid it.
Can we design a CBF safety filter for dynamic occluded environments with millisecond-scale computation?OcclusionCBF reasons about the reachable motion of hidden obstacles before detection and preserves a safe backup maneuver.
Core Idea: Reachable Occupancy + Backup Rollout
OcclusionCBF over-approximates and expands each occluded region into a time-varying reachable occupancy that covers where a hidden obstacle may move. It then rolls out a backup policy and verifies that the robot remains outside this occupancy over the horizon and reaches a safe terminal set. These conditions are converted into affine constraints for a QP safety filter.
Q: What does OcclusionCBF guarantee?A: Under the occupancy-coverage and the valid terminal-set assumptions, the backup input remains feasible on the certified recoverable set. The OcclusionCBF-QP preserves this set recursively and avoids every hidden-obstacle motion represented by the predictor.
Preview Experiments
1. Blind-Crossing Scenario
Motivated by a real-world blind-crossing accident in which a bus occluded an approaching vehicle from a pedestrian’s view, we first consider the same crossing scenario. OcclusionCBF intervenes before the hidden vehicle becomes visible, whereas the occlusion-agnostic CBF-QP reacts too late.
Occlusion-Agnostic CBF-QP
OcclusionCBF
2. Randomized Crowd Navigation Benchmark
We compare OcclusionCBF with five reactive and occlusion-aware baselines under identical randomized scenarios with moving obstacles and line-of-sight occlusions.
•
Double Integrator
More videos
•
Unicycle
Across 100 paired trials, OcclusionCBF achieves the highest success rate among the hard-constrained methods at every tested obstacle density, while requiring only millisecond-scale computation per control update.
Success, infeasibility, and collision rates over 100 paired trials for increasing obstacle densities.
Average controller computation time per update. OcclusionCBF requires 4.1–4.5 ms for the unicycle and 1.6–3.9 ms for the double integrator.
3. MetaUrban Simulations
We further evaluate closed-loop execution in the high-fidelity MetaUrban simulator using both unicycle and double-integrator robot models under realistic urban geometry and dynamic occlusions.
•
Double Integrator
OcclusionCBF: OcclusionCBF intervenes when needed to maintain a safe backup maneuver during closed-loop navigation.
Representative baseline behavior in the same occluded urban scenario.
•
Unicycle
OcclusionCBF: OcclusionCBF safely navigates the scene while accounting for hidden dynamic obstacles.
Representative baseline behavior under dynamic occlusions.
Hardware Experiments
We also evaluate OcclusionCBF in a blind-crossing hardware scenario using wheeled mobile robots and compare it with an occlusion-agnostic CBF-QP.
Occlusion-Agnostic CBF-QP: Reacts after detection and collides with the hidden obstacle.
OcclusionCBF-QP: Intervenes before detection and safely reaches the goal.
BibTex
@inproceedings{kim2026occlusioncbf,
author = {Kim, Taekyung and Park, Hun Kuk and Wada, Renya and Atanasov, Nikolay and Koga, Shumon and Panagou, Dimitra},
title = {OcclusionCBF: Backup Control Barrier Functions for Safe Navigation Among Hidden Dynamic Obstacles},
booktitle = {arXiv preprint arXiv:2609.06342},
shorttitle = {OcclusionCBF},
year = {2026}
}
LaTeX
복사