NAVER LABS takes DUSt3R successor BLASt3R and two other spatial-AI papers to ECCV 2026, one picked as a spotlight

NAVER LABS said in a blog post on Sept 1 that three of its spatial-intelligence papers were accepted at ECCV 2026, the European computer-vision conference. NAVER LABS Europe presented BLASt3R (Bundle Adjustment of Any Image Set with Multi-view Matching and Monocular Priors), a follow-up to its DUSt3R 3D-vision model that reconstructs a scene from the relationships between many images; the lab says it is markedly faster and more accurate than earlier models and folds in SLAM, so a moving robot can estimate its own position and map its surroundings on the device rather than relying on a server.

A second paper, Open-Vocabulary BEV Segmentation with 3D-Aware Geometric Constraints, uses a vision-language model to label objects, including ones never seen in training, and writes them into a real-time bird's-eye-view map for vehicles and robots.

The third, InSpace, rebuilds a room's walls, floor and furniture in 3D from a single 360-degree panorama and was selected as an ECCV spotlight, a slot the lab says goes to about 1 to 2% of submissions.

The lab did not give benchmark numbers in the post. On MEKAWIRE's pages this is the research layer under NAVER LABS' commercial robot stack: three days later it unveiled FLOW, a no-code tool for authoring services on its ARC brain robot platform, which already runs in its 1784 headquarters and is being deployed in Saudi Arabia and Tokyo's Yaesu.