Naver Labs Europe unveils 'Divine,' a universal encoder to lighten robots' AI workload

Naver Labs Europe on Monday unveiled Divine, a universal encoder designed to improve the operational efficiency of autonomous robots in industrial and everyday settings.

An encoder converts data collected by a robot's sensors — such as cameras and LiDAR — into a form that AI models can process. Conventional autonomous robots have relied on separate encoders for each task, including localization, depth estimation, spatial understanding and human detection.

That approach requires multiple AI models to repeatedly process the same input data, driving up memory consumption and computational load. Divine consolidates the functions of multiple encoders into a single module, addressing that bottleneck. Without installing separate encoders for each task, a robot equipped with Divine can handle a range of visual AI functions — including 2D image understanding, 3D spatial reconstruction and human recognition.

Naver Labs Europe applied a technique called "multi-teacher distillation," which transfers the core processing capabilities of specialized expert models — each trained for a distinct domain such as image understanding, spatial mapping or human recognition — into a single unified model. The approach allows the system to handle diverse functions without running multiple large specialist models simultaneously.

Naver Labs Europe also released experimental results. According to the company, Divine reduces encoder memory consumption by roughly 90 percent compared with running multiple separate encoders. Encoding speed improved by up to 12 times, overall system-level memory use fell by about 62 percent, and system processing speed increased by up to four times.

Divine is also suited for onboard environments where AI functions run directly inside the robot. Conventional AI models for robots have often depended on servers or high-performance computing hardware due to their heavy computational demands. Because Divine can handle multiple AI functions with relatively modest memory and processing resources, it broadens the range of autonomous robot platforms to which it can be applied.

In environments where people and robots move together, the ability to quickly perceive and respond to surrounding conditions is critical. Naver Labs Europe said Divine enables robots to simultaneously recognize both spaces and people within constrained computing resources, which it believes will lower the barrier to deploying robots in industrial and everyday settings.

Scalability was also a design consideration. Divine is built so that new AI capabilities can be added through software updates rather than hardware changes, allowing performance to be upgraded without modifying the robot itself.

"Lightweighting the robot brain has emerged as a key challenge worldwide in commercializing physical AI," said Lee Dong-hwan, leader of the Vision Group at Naver Labs. "Divine will help lower the barrier to adopting AI robots across everyday and industrial environments."

Two research papers from Naver Labs Europe on Divine were accepted at the European Conference on Computer Vision (ECCV 2024) and the Conference on Computer Vision and Pattern Recognition (CVPR 2025), respectively.

Meanwhile, Naver Labs has been advancing its physical AI technology on the foundation of spatial data capabilities including digital twins, autonomous robots and high-precision indoor mapping. The company has already built a real-world environment where AI and robots operate together — with autonomous robots handling in-house delivery and logistics at Naver's second headquarters, "1784" — and the release of Divine extends that push to broaden the technology's applicability in real-world settings.