Martin Humenberger: Making new connections – AI for the physical world at NAVER LABS Europe
Martin Humenberger of NAVER LABS Europe explains how the company connects AI research with real-world robotics, from robot-operated buildings to models for vision, optimization, and interaction.
In this talk
- 2:29
Robots in buildings and data centers
At NAVER’s 1784 headquarters, roughly 100 robots deliver coffee, meals, and parcels, coordinated through ARC Brain and connected building infrastructure. Humenberger also describes city-scale 3D maps made from data collected by planes, cars, and people, plus robots and an autonomous shuttle at the Gak Seong data center, including a machine that moves equipment weighing up to 400 kilograms.
- 6:03
A 3D vision model for robotics
NAVER LABS Europe, based in Grenoble, has about 100 researchers from around 27 countries working across vision, action, and interaction. Its CoCo model learns 3D structure from millions of image pairs by reconstructing masked views, then supports tasks including stereo matching, optical flow, camera localization, and image-goal navigation.
- 13:32
Foundation models for optimization
The action research group aims to train models that can address multiple combinatorial optimization problems, including multi-robot routing and task assignment. BQ-NCO learns policies that transfer from small problem instances to larger ones, while GOAL combines a shared model with task-specific adapters to support transfer across different problems.
- 16:58
Giving users control over model outputs
The interaction group’s DISCO toolkit applies user-defined constraints to sequential models. Users specify desired features, build an energy-based distribution reflecting them, then tune an autoregressive language model to approximate that distribution; one example is balancing gender representation in answers about tennis.
Briefs summarise talks and interviews; timestamps open the video at that moment. Claims in a talk are the speaker's own.