To navigate the world, robots need to understand precisely what surrounds them. Two new developments created by a group led by Professor Zhu Hui of the Institute of Intelligent Machines at the Hefei Institutes of Physical Science of the Chinese Academy of Sciences help accomplish this task. Both works have been accepted for presentation at the ICRA 2026 conference (IEEE International Conference on Robotics and Automation), one of the leading events in robotics.
The first development, called LR-SGS, is responsible for high-quality 3D reconstruction, that is, for building a volumetric model of the surrounding space. The method combines lidar data (a laser rangefinder that measures distances to objects using light pulses) and camera images. By using the intensity of the reflected signal, the system captures scene details under different lighting conditions, including at night. In addition, the algorithm identifies key structural features and discards redundant information, which improves processing efficiency.
Experiments on datasets, including Waymo, showed that LR-SGS improves reconstruction accuracy and training speed while maintaining a compact three-dimensional representation. The reconstructed scenes can be edited and used to create virtual driving environments, for example, for simulating and testing autonomous driving.
The second development, MMD-SLAM, improves robot localization and map building. It is based on the geometric features of buildings and other man-made objects: points, lines and planes. By combining these elements, the method builds more consistent three-dimensional point clouds (sets of points in space that describe the shape of objects) and improves positioning accuracy. Tests demonstrated improved mapping and tracking metrics while maintaining real-time processing speed.
According to the researchers, both methods provide new tools for scene reconstruction and may find application in robotics and intelligent systems. At the same time, these are laboratory developments, not finished commercial products. The paper on the reconstruction method was among three finalists for the best paper award in the field of robot perception; more than 5,000 submissions were made to the conference in total.