Robotics developers have long been trapped in a frustrating trade-off between intelligence and mobility. To give a robot the ability to understand its environment in real time, engineers typically face two suboptimal choices: strap a heavy, power-hungry GPU to the chassis or tether the machine to a high-performance external server. The former kills battery life and adds cumbersome weight, while the latter introduces latency and a critical dependency on stable network connectivity. This physical bottleneck has remained a primary hurdle in moving autonomous robots from controlled lab environments into the unpredictable chaos of the real world.
The Redundancy Crisis in Robot Perception
NAVER LABS Europe recently addressed this hardware dependency with the introduction of DIVINE, a general-purpose encoder designed to unify image understanding, spatial awareness, and person recognition into a single pipeline. In the context of robotics, an encoder serves as the critical first stage of the AI pipeline, converting raw sensor data from cameras or LiDAR into numerical representations that a model can process. Until now, the industry standard has been a fragmented approach. A typical autonomous robot would run separate AI models and encoders for every specific task: one for localization, another for depth estimation, a third for spatial mapping, and yet another for identifying humans.
This fragmented architecture creates a massive amount of redundant computation. The robot processes the same input data multiple times through different encoders, leading to an exponential increase in memory consumption and computational overhead. By treating every perception task as a silo, robots waste precious on-board resources on repetitive calculations, which in turn necessitates the expensive, heavy hardware that limits their operational efficiency.
Distilling Expert Knowledge into a Single Brain
The breakthrough with DIVINE lies in a technique called multi-teacher distillation. Rather than attempting to build a massive, all-encompassing model from scratch, NAVER LABS Europe utilized a teacher-student framework. In this setup, several specialized expert models—the teachers—each possessing deep knowledge in a specific domain like 2D image understanding or 3D spatial reconstruction, are used to train a single, leaner student model. The student model does not inherit the bloated parameter count of the teachers; instead, it learns to mimic the core functional outputs and decision-making logic of those experts.
This shift transforms the robot's cognitive architecture from a collection of separate specialists into a single, versatile polymath. By condensing the intelligence of multiple high-capacity encoders into DIVINE, the robot no longer needs to run parallel processes for different visual tasks. It performs a single encoding pass to extract all the necessary information for localization, depth, and recognition simultaneously. This eliminates the computational overlap that has plagued on-board AI for years, allowing the robot to maintain high-level perception without the need for a server-grade computer on its back.
Quantifying the Efficiency Leap
The empirical results of this architectural shift are stark. In experimental environments, DIVINE reduced encoder memory usage by approximately 90% compared to systems running multiple individual encoders. Because the data transformation path is now streamlined, the encoding processing speed increased by up to 12 times. When looking at the robot's entire system, the total memory footprint dropped by roughly 62%, and the overall system processing speed improved by up to 4 times.
These gains directly translate to a shorter perception-to-action cycle. In environments where humans and robots coexist, a millisecond of latency can be the difference between a smooth interaction and a collision. By slashing the time it takes to recognize a person or calculate a spatial gap, DIVINE provides the foundation for robots to react instantaneously to their surroundings using only their internal computing power. This efficiency makes the concept of Physical AI—AI that interacts with the physical world in real time—practically viable on consumer-grade or industrial-grade embedded hardware.
Beyond the raw speed, DIVINE is designed for modularity. Because it acts as a unified interface for perception, developers can update the AI's capabilities without replacing the robot's physical hardware. When a new perception model is developed, updating the DIVINE encoder allows the existing fleet to inherit new intelligence via software updates, significantly lowering the total cost of ownership for AI robot deployments.
The technical validity of this approach has already gained international recognition, with two related research papers accepted at the 2024 European Conference on Computer Vision (ECCV) and the 2025 Conference on Computer Vision and Pattern Recognition (CVPR). As noted by Donghwan Lee, leader of the Vision Group at NAVER LABS, this technology is intended to lower the barrier to entry for deploying AI robots across both daily life and industrial settings.
This transition toward lean, on-board intelligence marks the end of the era where robot brains were limited by the size of their batteries or the strength of their Wi-Fi signals.
