RAPTOR: Robustness Assessment of Patch Transfer and Operational Risk in Edge Object Detectors for Mobile Robots

Abstract

Autonomous mobile robots rely on object detection for navigation and safety, yet edge-deployed detectors remain vulnerable to adversarial patches under limited compute, power, and latency constraints. This paper presents a deploymentoriented robustness benchmark using nine CARLA-GeAR billboard scenarios and five detector configurations spanning three detector design families. Scenario-specific patches are optimized with white-box access to YOLOv8n using Expectation over Transformation and are then evaluated through black-box transfer without patch re-optimization or target-model access. On YOLOv8n, the attack reduces mAP50 by 31.4 percentage points and increases high-confidence false positives by approximately 40×. This reveals a perception-overload failure mode not fully captured by standard accuracy metrics. The attack also transfers to other detector configurations, while a patch trained on billboard01 reduces mAP50 by 13.3 percentage points across eight unseen billboard scenarios. Full-pipeline measurements on NVIDIA Jetson AGX Orin show a 3.5% increase in end-to-end latency and only a 0.05 W change in mean power, suggesting that severe semantic degradation may produce only small changes in platform telemetry. Among five non-adaptive preprocessing defenses, Local Gradient Smoothing provides the strongest overload–accuracy trade-off, reducing low-threshold false-positive flooding by 63.6% at a 6.6 percentage-point clean-accuracy cost. These results highlight the need to evaluate edge perception using cross-model transfer, output-load, platform, and defense metrics alongside conventional detection accuracy.