DMMG-RPF: A Dual-Modal Mutual Guidance Framework for Robust 3-D Object Detection Under Sensor Degradation

While multimodal perception is important for autonomous operations in hazardous industrial environments, conventional fusion architectures remain vulnerable to asymmetric sensor degradation. By relying on unconstrained feature aggregation, these paradigms may propagate localized soft failures–such as volumetric dust scattering or optical attenuation–across modalities. To address this issue, this paper introduces the Dual-Modal Mutual Guidance Robust Perception Framework (DMMG-RPF), a physics-informed cross-sensor validation framework for robust 3D object detection. Rooted in sensor phenomenology, the architecture exploits physical complementarity to attenuate Mie-scattered LiDAR noise and anchor visual representations under Beer-Lambert energy decay. A quality-aware gating mechanism based on latent entropy is further used to down-weight degraded modalities and reduce cross-modal contamination. Evaluated on the real-world AutoMine dataset, DMMG-RPF achieves 57.99 3D AP for safety-critical Trucks among publicly reported AutoMine results, with KITTI degradation experiments further indicating adverse-weather robustness.