Naver has integrated visual artificial intelligence (AI) processing models for autonomous robots into a single system. It said the next-generation brain technology lets robots quickly recognise surrounding environments without heavy high-performance equipment.

Naver said on Monday that its subsidiary Naver Labs Europe unveiled a general-purpose encoder, DIVINE, that helps autonomous robots carry out tasks in industrial and everyday environments. An encoder converts data collected by a robot through cameras and sensors such as LiDAR into a form AI models can process.

Existing autonomous robots had inefficiencies because they used separate AI models and encoders for each task, such as localisation, depth calculation, spatial understanding and human recognition, processing the same data repeatedly. Naver Labs Europe said it addressed the problem through a "multi-teacher distillation" method that extracts only the core knowledge of specialised models and transfers it into a single model. It compressed 2D image understanding, 3D space reconstruction and human recognition into DIVINE.

Tests showed a sharp reduction in computational burden and improved performance. Compared with using multiple encoders, encoder memory usage fell by about 90 percent and encoding speed was up to 12 times faster. Total memory use across the robot system fell by about 62 percent and overall system processing speed improved by up to four times.

This will make it easier to equip small robots without large computing hardware with high-performance AI, and the scalability of robot operations is expected to increase.

Lee Dong-hwan (이동환), vision group leader at Naver Labs, said, "Around the world, lightweight robot brains are emerging as a key topic for commercialising physical AI." He added, "DIVINE will contribute to lowering barriers to introducing AI robots across everyday and industrial sites."

Two studies related to DIVINE were accepted by the European Conference on Computer Vision (ECCV) and the Conference on Computer Vision and Pattern Recognition (CVPR), respectively, receiving recognition for the technology.