In context
In 2025, underground robotics—spanning mining, tunnel construction, and rescue—faced persistent perception challenges: low-light conditions degrade visual detection, while onboard computing constraints limit heavy models. Existing solutions often relied on single sensors or computationally intensive fusion, leaving a gap for efficient, robust detection suitable for real-time autonomous operation in such harsh settings.
What was reported
Researchers proposed a lightweight dual-modal object detection network that fuses LiDAR point clouds with RGB images via a dual-branch architecture. The design integrates multi-scale feature fusion at shallow, intermediate, and deep levels, introducing a StarFusion module that uses element-wise multiplication to enhance cross-modal feature interaction. Depthwise separable convolutions and channel compression reduce model parameters to 2.3 million, roughly half of the original backbone's size.
To validate the approach, the team constructed a low-light multimodal dataset featuring four classes of typical underground targets, with image brightness (25±8.3) and sharpness (18.6±6.9) significantly lower than conventional datasets. On this dataset, the method achieved an mAP50 of 86.1%, a 2.6% improvement over the YOLOv8 baseline, while maintaining an inference speed of 20 frames per second.
The model was deployed on a Jetson Orin NX platform aboard an exploration robot. Field tests confirmed that the dual-modal complementary mechanism effectively overcomes single-sensor blind spots in low-light conditions, providing reliable real-time environmental perception for autonomous underground operations.
Why it mattered
This work demonstrated a practical path to balancing detection accuracy and computational efficiency for underground robots, highlighting the value of LiDAR-RGB fusion in extreme lighting. It underscored a shift toward multimodal perception as a key enabler for robust autonomy in industrial environments where vision alone is insufficient, offering a reference for future edge-deployable perception systems in similar constrained settings.
"The dual-modal complementary mechanism effectively overcomes the perception limitations of single-sensor systems in low-light conditions."
Source: 《机器人》期刊 (robot.sia.cn) · Published 2025-08-19 · “面向地下环境机器人的多模态目标检测方法”