NAVER LABS Europe said on the 23rd that it unveiled "DIVINE," a universal encoder that will make the work of autonomous robots in industrial and everyday environments easier.
Autonomous robots have used multiple artificial intelligence (AI) encoders together to perceive their surroundings and handle various tasks. An encoder is a device that converts data collected by sensors such as cameras and LiDAR into a form that an AI model can process.
DIVINE is a universal encoder that integrates the multiple encoders needed at this stage into one. Previously, for each task—such as ▲pose estimation ▲depth calculation ▲spatial understanding ▲person recognition—each AI model used a separate encoder to redundantly process the same input data multiple times, which excessively increased memory usage and computation.
NAVER LABS Europe addressed this by using a "multi-teacher distillation" approach that integrates the core of the information-processing capabilities learned by each specialized encoder into a single encoder. Multi-teacher distillation extracts only the essential knowledge from expert "teacher" models specialized in each domain—such as images, space, and person recognition—and implants it into a single "student" model. Using this student model allows broad handling of diverse domains without maintaining multiple large expert models.
In DIVINE's case, it condenses into one the functions of several encoders that each specialize in 2D image understanding, 3D spatial reconstruction, and person recognition. As a result, robots can perform diverse AI tasks with DIVINE alone, without needing to mount multiple different encoders.
When DIVINE was run in real experimental settings, computational load fell while performance was maximized. Compared with mounting multiple encoders, encoder memory usage was reduced by about 90%, and encoding throughput increased by up to 12 times. Overall robot memory usage decreased by about 62%, and system processing speed improved by up to 4 times.
NAVER LABS Europe said, "Even without large hardware equipped with expensive and heavy compute units, robots can carry out various tasks autonomously by grasping situations on their own if they are fitted with DIVINE, a 'fast and smart brain,'" and added, "Each time an AI model is upgraded, performance can be easily boosted by updating DIVINE installed on existing robots, without introducing new robots with the updated model."