· Research & technology· 日本語で読む
Team at Chinese Academy of Sciences’ Hefei Institutes of Physical Science develops methods for 3D scene reconstruction and robot mapping
The research team developed a method that combines LiDAR and cameras, and another that uses geometric features. The first targets 3D reconstruction of driving scenes; the second addresses robot localization and 3D mapping. Both studies are scheduled to be presented at ICRA 2026, and the paper on scene reconstruction was selected as a finalist for the Best Paper Award in Robot Perception.
What happened
TechXplore reported on November 11, 2026, that a research team led by Professor Zhu Hui of the Chinese Academy of Sciences’ Hefei Institutes of Physical Science had developed two methods to improve 3D scene reconstruction and robot mapping. LR-SGS, which combines LiDAR and cameras, targets reconstruction of driving scenes, while MMD-SLAM, which uses geometric features, addresses robot localization and 3D mapping.
Why it matters
According to the team, LR-SGS uses LiDAR and camera images to reconstruct driving scenes under different lighting conditions. The fact that separate methods address representation of a robot’s surroundings and estimation of its position and map provides a concrete point to consider when deciding which approach to apply to a given perception task.
Analysis
The key point in this report is that the two methods address different problems, even though both concern “robot perception.” LR-SGS reconstructs the surrounding scene in 3D, while MMD-SLAM targets the robot’s own localization and mapping. The two should therefore be understood not as a single integrated system but as proposals for distinct tasks: representing the environment and estimating position and maps. The report presents both as work by the same team, but does not describe an implementation that links the methods or any performance gains from using them together.
LR-SGS combines LiDAR with camera images and is said to use intensity information to capture details under different lighting conditions. This indicates that the issue is not simply the use of multiple sensors, but which information is used for reconstruction. For MMD-SLAM, the report cites the use of geometric features, but the article text does not specify the features or processing steps. The methods use different approaches to their inputs and processing, but no comparison under the same conditions is provided, so there is no basis for ranking which is more effective in which environments.
The scene-reconstruction paper’s selection as one of three finalists for the Best Paper Award in Robot Perception, from more than 5,000 submissions, is information about its academic recognition. However, being a finalist does not demonstrate safety on physical robots or accuracy in long-term operation. The report also does not establish how well LR-SGS performs in real-world conditions involving moving objects and changes in lighting. Assessing the research’s practical significance requires distinguishing its academic evaluation from results of deployment and operation on robots.
Three points warrant further attention. First, the evaluation conditions used to measure reconstruction and localization accuracy, and comparisons with existing methods. Second, specific validation results for performance with moving objects, changing lighting and complex shapes. Third, operation on physical robots, including computational demands, and the configuration that would be used if the two methods were combined. Until these details are provided, it is appropriate to assess the work as two research methods for different applications.
Sources
Translated from the Japanese original: /news/chuugokukagakuin-gouhi-busshitsukagaku-kenkyuu-in-no-chiimu-27e7cb. The Japanese version is authoritative if they differ.