Abstract
Autonomous navigation of agricultural robots faces persistent challenges in unstructured environments, including canopy occlusion, specular reflections from pooled water, and ridge-induced platform vibrations that cause temporally varying sensor degradation. Existing multimodal systems fail to maintain both reliability and real-time responsiveness, while accurate delineation of safety–critical classes (especially waterlogged zones) and the direct translation of perceptual quality into navigational safety remain key bottlenecks. We present a lightweight real-time adaptive multimodal perception system that establishes a complete perception–planning closed loop. Built on MobileNetV3-U-Net-Lite, the architecture integrates (1) six-channel early RGB–LiDAR fusion for deep cross-modal complementarity from the shallowest layers, (2) stepwise collaborative filtering with learnable anisotropic boundary weights supervised by a composite CE + Dice + boundary loss to handle directional terrain margins and low-frequency safety–critical classes, and (3) a 10-dimensional physical state vector (IMU vibration signatures + RGB illumination descriptors) that drives an MLP-based dynamic modality weighting mechanism reinforced by physical-consistency regularization. The fused output generates a quantitative risk map encoding slope, surface undulation, water-accumulation probability, and vegetation obscuration, which directly feeds the Agri Path planner for risk-guided path refinement and global map accumulation. Deployed on the NVIDIA Jetson Orin NX, the system achieves 47 FPS with 3.4 M parameters and 7.8 W power consumption. On the RELLIS-3D benchmark it attains 62.56 % mIoU, 40.42 % puddle IoU, and boundary F1 scores exceeding all baselines. Field trials at Xihua University demonstrate a 91.7 % improvement in minimum safe distance over the IMU-ablated baseline and reduction of peak yaw-rate to 37.5 % of the control group. These results demonstrate a viable pathway toward physically grounded, temporally adaptive safety navigation: shifting from static precision-centric fusion to physically grounded, temporally adaptive, risk-quantified safety navigation under tight edge-computing constraints. Field validation under dusk post-irrigation conditions confirms system robustness under compound sensor degradation; the proposed mechanisms demonstrate strong potential for transfer to orchards, vegetable fields, and wheat croplands, subject to further validation under full nocturnal and adverse-weather conditions.