In context
By 2025, neural radiance fields (NeRF) were being integrated into SLAM systems to enable dense mapping, but reflective surfaces—common in labs and industrial settings—remained a challenge due to view-dependent lighting and false features. This work from the University of Science and Technology of China addressed that gap for robotic navigation.
What was reported
Researchers proposed RefN-SLAM, a neural SLAM method that models high-light (specular) and low-light (diffuse) scene components with two separate NeRFs. These are combined using a learned tone-mapping coefficient to produce the final color, improving handling of reflections and complex illumination.
The method enhances depth perception through surface-aware and perspective-aware sampling, and uses a coarse-to-fine optimization to boost accuracy and efficiency. Camera pose and scene representation are jointly optimized on a global keyframe pixel database.
Experiments showed satisfactory reconstruction in a chemistry laboratory and excellent tracking performance on synthetic datasets (Replica) and real-world robotic tests. The reconstructed model met the accuracy requirements for trajectory planning of a chemical robot arm.
Why it mattered
RefN-SLAM advanced dense SLAM for reflective environments, a common pain point in industrial automation where shiny metals and glass disrupt visual odometry. By enabling reliable mapping and tracking in such settings, it supports autonomous robots in labs and factories, improving navigation and manipulation tasks.
“The reconstructed model accuracy can meet the requirements of trajectory planning tasks for a chemical robot arm.”
Source: 《机器人》期刊 (robot.sia.cn) · Published 2025-09-16 · “RefN-SLAM：反射场景下的神经SLAM方法”