Naver's robotics competitiveness has expanded into the realm of "brain lightweighting." Naver Labs Europe, the company's research and development (R&D) subsidiary, unveiled DIVINE, a universal encoder for autonomous mobile robots, on Tuesday. The technology improves the efficiency of the first gateway through which robots perceive their surrounding environment and process data, and is expected to lower the barrier to adopting robots in everyday and industrial settings.
According to Naver Labs Europe, an encoder is a device that converts data collected by robots through cameras and lidar into a form that AI models can process. Autonomous mobile robots mobilize multiple AI encoders to perceive their surroundings, but DIVINE integrates these into a single one. It handles everything from 2D image understanding to 3D recognition encompassing space and people within a single model.
This development is significant because it can minimize unnecessary robot memory usage. A robot AI model, which can be considered the robot's brain, begins its work by perceiving the surrounding physical environment and processing data. Previously, for each task such as position estimation, depth calculation, spatial understanding, and human recognition, each AI model used a separate encoder to redundantly process the same data. This caused the problem of excessive increases in memory usage and computational load.
The effect was confirmed numerically in in-house experiments. Compared to the case of mounting three encoders, encoder memory usage was reduced by about 90%, and encoding processing speed was up to 12 times faster. As the first gateway of encoding became lighter, the robot's overall memory usage decreased by about 62%, and system processing speed improved up to fourfold. As a result, this leads to the lightweighting of the robot AI model as a whole.
The foundation of the technology is "multi-teacher distillation." This method extracts only core knowledge from specialized expert "teacher" models in each field, such as image, spatial, and human recognition, and transplants it into a single "student" model. It can handle multiple domains without maintaining several large expert models. Related papers were accepted at the European Conference on Computer Vision (ECCV) and the Conference on Computer Vision and Pattern Recognition (CVPR), considered the world's three major computer vision conferences, in 2024 and 2025 respectively, recognizing the technology's capabilities.
Naver Labs believes DIVINE's true value will emerge in the commercialization phase of physical AI. Recently, in the physical AI field, the importance of on-device AI, which allows robots to perform computational tasks on their own without external cloud connection, has been growing increasingly. In this situation, DIVINE, which helps lightweight robot brains, is expected to have high utility in on-device environments. Existing robot AI models required massive computation and have mainly run in server environments, but using DIVINE allows AI functions to be executed with less memory and computational load. Lee Dong-hwan, leader of Naver Labs' Vision Group, said, "Robot brain lightweighting is emerging as a major topic worldwide for the commercialization of physical AI," adding, "DIVINE will contribute to lowering the barrier to adopting AI robots across everyday and industrial settings."
The unveiling of DIVINE is seen as further strengthening Naver's full-stack robotics competitiveness. Naver Labs Europe introduced DUSt3R, a vision model that reconstructs space in 3D from just a single photo, in December 2023, followed by the next-generation DUSt3R2 and the 3D body model ANI in November last year. In particular, DUSt3R has led derivative research by global big tech companies including Meta, Google DeepMind, and Nvidia since its open-source release.
