In the future, general purpose AI systems will be installed in robots. There will be various means of achieving this, and it is particularly noteworthy that NAVER Labs Europe, a global robotics AI lab, has developed something that may very well be a technological turning point: the DIVINE. DIVINE is a universal AI encoder that simplifies the complex structures of a robot AI system and maximizes computational efficiency.
*An encoder is a module that transforms unprocessed input data such as data from cameras and sensors into a version that AI models can understand and process.
How DIVINE Eliminates Redundancy
DIVINE is a single, lightweight encoder compressed from many AI encoders. In executing tasks such as localization, distance and depth estimation, obstacle perception, and scene understanding, the current generation of robots relies on repetitive processing of input data via a separate encoder for each task. This leads to structural limitations because it bloats the amount of computed data and memory usage. However, with DIVINE, a single data input enables many AI systems and eliminates redundancies in the process.
NAVER LABS Europe adopted the “Multi-Teacher Distillation” method, which allows for learnings and knowledge stored in various specialized AI models to be compressed into a single encoder. Specifically, DIVINE integrates knowledge from specialized AI models, including DUSt3R, which excels in understanding 3D spaces, and Multi-HMR, which recognizes human posture and movement. However, DIVINE is not merely a condensed result of merging multiple AI models; each and every model used and their distinct characteristics and strengths have been seamlessly integrated into DIVINE. Moreover, DIVINE boasts strong scalability in that NAVER LABS Europe’s proprietary AI models as well as external models can be easily added or integrated into it. Through this process, DIVINE can use knowledge and data outside its components, which actually helps DIVINE perform better than its original models in certain tasks.
DIVINE’s scope of usability is not limited to just visual-based capabilities. Because it is a general-purpose framework that can learn from various kinds of foundation models, DIVINE can scale into a multimodal AI that can process language, audio, and more.
Evolving as a More Efficient On-Board AI
Changes in structure lead to improvements in performance and efficiency. According to test results, memory usage in the encoding stage decreased by 90%, and processing speed improved by up to twelve-fold. Looking at the robot system as a whole, total memory usage dropped by 62%, and overall processing speed improved by up to four-fold.
By utilizing DIVINE, more AI tasks can be processed with the limited computing resources available in a robot, and the latest AI models can be integrated seamlessly. Indeed, DIVINE is a key technology that will enhance the performance of future on-board AI.
DIVINE Researcher Commentary
Q: What was a focal point during the development of DIVINE?
A. “Model distillation” (a technique for compressing and transferring the knowledge of a large-scale AI model to a smaller one) is a field that has been researched by many different institutions, but NAVER LABS Europe is the first to integrate tasks that are completely different from another, such as semantic segmentation, 3D space reconstruction, and human perception, into a single encoder. This goes beyond just combining multiple models; during this distillation process, each model absorbs knowledge from one another so that they perform better at certain tasks than if they executed said task individually. This remarkable performance has been published in top global computer vision conferences such as the CVPR 2025 and ECCV 2024.
Q. What other fields besides robotics can DIVINE be applied to?
A. DIVINE is a general-purpose technology that has applications in environments that require many AI processes to be simultaneously executed. From AR and VR, medical videos, manufacturing, agriculture, to environmental monitoring, DIVINE can be used across many different industries to enhance multi-AI system efficiency.
Q. Plans for DIVINE in NAVER’s robots?
A. We plan to apply DIVINE in various robots in the future, including NAVER’s own. NAVER possesses a wide range of robotics platforms, and the organic connection between our AI research and these platforms provides a highly advantageous environment for testing and advancing new AI technologies.