[Edaily Reporter Lee So-Hyun ] NAVER LABS Europe, the research and development division of NAVER (NAVER(035420)), has developed a next-generation universal encoder that reduces the artificial intelligence (AI) processing load on autonomous robots.

NAVER LABS Europe announced on the 23rd that it has unveiled “DIVINE,” a general-purpose encoder designed to assist autonomous robots in performing tasks in industrial and everyday environments.

An encoder is a device that converts data collected by a robot’s sensors—such as cameras and LiDAR—into a format that AI models can process.

Autonomous robots have traditionally used multiple AI encoders in tandem to perceive their surroundings and perform various tasks. DIVINE is a universal encoder that integrates these functions into a single unit. It is designed to handle a wide range of visual AI functions—from image understanding to 3D spatial reconstruction and person recognition—all through a single encoder.

Previously, each task—such as position estimation, depth calculation, spatial understanding, and person recognition—required its own AI model and separate encoder. This led to the problem of increased memory usage and computational load, as the same input data had to be processed multiple times.

Naver Labs Europe solved this problem by utilizing a “multi-teacher distillation” approach. Multi-teacher distillation is a method that extracts core knowledge from expert models specialized in specific domains—such as image, spatial, and person recognition—and transfers it to a single student model. This technology enables a single model to handle various domains without the need for multiple large expert models.

DIVINE consolidates the functions of multiple encoders—each specialized in 2D image understanding, 3D spatial reconstruction, and person recognition—into a single unit. As a result, robots can perform various AI tasks using just DIVINE, without needing to be equipped with multiple separate encoders.

In environments where humans and robots coexist, the ability to quickly perceive the surrounding situation and respond immediately is crucial. DIVINE helps robots rapidly understand their surroundings even with limited computing resources.

Experimental results showed that Divine improved performance while reducing computational load. Compared to when multiple encoders were installed, encoder memory usage was reduced by approximately 90%. Encoding processing speed improved by up to 12 times. The robot’s overall memory usage decreased by about 62%, and system processing speed increased by up to 4 times.

Existing AI models for robots were primarily run in server environments or on high-performance computing equipment due to their massive computational demands. In contrast, DIVINE can execute AI functions with minimal memory and computational resources, thereby enhancing the usability of on-board environments where processing occurs directly within the robot.

Naver Labs Europe expects this to enable the application of high-performance AI to a wider variety of robot types. The company explains that by utilizing DIVINE, it is possible to operate autonomous AI robots that can assess their surroundings and perform various tasks without relying on large, heavy hardware equipped with expensive computing devices.

Another key feature is its design, which allows for the easy addition of new AI capabilities. Whenever an AI model is upgraded, performance can be enhanced by updating the DIVINE installed in existing robots, eliminating the need to introduce new robots.

Lee Dong-hwan, Leader of the Vision Group at Naver Labs, stated, “Globally, the lightweighting of robot brains is emerging as a key topic for the commercialization of physical AI,” adding, “Divine will contribute to lowering the barriers to adopting AI robots across everyday life and industrial settings.”

Two Divine-related research papers from Naver Labs Europe were accepted for presentation at the 2024 European Conference on Computer Vision and the 2025 Conference on Computer Vision and Pattern Recognition, respectively. Based on these findings, Naver Labs plans to continue its research on robot foundation models, which are considered the core of physical AI.