Naver Labs Europe on Monday unveiled DUNE, a universal encoder that helps run artificial intelligence (AI) robots with less memory and computing power.
An encoder is a device that converts data collected by robots through sensors such as cameras and LiDAR into a form that AI models can process. Autonomous robots use multiple AI encoders together to perceive their surroundings and handle various tasks. However, in this process, data was processed redundantly, causing excessive increases in memory usage and computing load. DUNE is a universal encoder that integrates multiple encoders into one, capable of supporting various visual AI functions ranging from image understanding to spatial and human recognition.
Naver Labs Europe solved this problem using a "multi-teacher distillation" method, which integrates the core information-processing capabilities learned by each encoder into a single encoder. Multi-teacher distillation is a method that extracts only the essential knowledge from expert "teacher" models specialized in each field, such as image, spatial, and human recognition, and transplants it into a single "student" model. Using a student model created this way, various fields can be handled comprehensively without maintaining several large expert models.
When DUNE was run in an actual experimental environment, encoder memory usage was reduced by about 90% compared with mounting multiple encoders. Encoding processing speed improved up to 12-fold. In addition, the robot's overall memory usage decreased by about 62%, and system processing speed improved up to 4-fold. Since AI functions can be executed with less memory and computing power, it is expected to serve as a foundation for applying high-performance AI to more diverse types of robots.
Furthermore, DUNE is designed to allow new AI functions to be added easily. Each time an AI model is upgraded, performance can be easily enhanced by updating the DUNE installed on existing robots, without introducing new robots equipped with the updated model.
"Globally, lightening the robot brain is emerging as a major topic for commercializing physical AI," said Lee Dong-hwan, leader of Naver Labs' Vision Group. "DUNE will contribute to lowering the barriers to adopting AI robots across daily life and industrial sites."
