NAVER LABS Europe @ECCV 2026

At ECCV 2026 in Malmö, the 10 accepted papers of NAVER LABS Europe span 3D reconstruction and scene generation, visual navigation and representation learning, and human and video understanding. We’re also present with 5 workshop keynotes, an invited spotlight and several live demo sessions! This is part of a 23-paper showing from Team NAVER overall. The detailed agenda of what and where to find us is in the news item.

The work presented is part of our research on AI for Robotics. Robots need to map their environment as they move through it, recognize and reason about the people and objects around them, and do all of this with models that don’t require a data center. That’s the thread running through our papers this year: faster, more robust SLAM for building maps in real time; a navigation policy that runs on a frozen backbone with almost no extra training cost; distillation and merging techniques for keeping perception stacks small without sacrificing capability; and human and object understanding built for the cluttered, dynamic, long-horizon conditions a robot actually encounters vs. curated benchmark images. Here’s a recap of what we’re presenting by theme.

Geometric Foundation Models (for robotics)

BLASt3R: bundle adjustment of any image set with multi-view matching and monocular priors Vincent Leroy, Philippe Weinzaepfel, Lojze Zust, Yohann Cabon, Jerome Revaud
BLASt3R extends the DUSt3R line of work with a bundle-adjustment formulation that fuses multi-view matching with monocular depth priors, reconciling any number of overlapping views into one consistent, metric-scale reconstruction. The approach improves robustness and accuracy over prior DUSt3R-family models while remaining lightweight enough for online, on-device SLAM which is the core capability a mobile robot needs to build a map of its surroundings and localize itself within it as it moves, without offloading computation to the cloud.

Sparse autoregressive modeling for scene generation from multi-view images Thomas Lucas, Maxime Pietrantoni, Wonjune Cho, Bardienus Pieter Duisterhof, Philippe Weinzaepfel, Vincent Leroy, Jerome Revaud
This work reframes 3D scene generation as sequential, token-by-token prediction over a sparse set of scene elements conditioned on multi-view input, rather than dense volumetric synthesis. The autoregressive formulation generates coherent, controllable 3D scenes and is well suited to producing synthetic training environments for robot navigation and manipulation policies. These are scenes a robot could plausibly encounter, generated at a fraction of the cost of capturing and annotating them in the real world.

Vincent Leroy, Jerome Revaud and Boris Chidlovskii from our Geometric Deep Learning team and creators of DUSt3R can also be found at the workshops below plus Christian Wolf from our Spatial AI team at ViLMa.

Also in the field of 3D reconstruction you can catch a live demo of MUSt3R Real-Time VSLAM (CVPR25) and PanSt3R 3D Panoptic Segmentation (ICCV25) in demo session 3.

Visual navigation, representation learning and model efficiency

A scalar per patch from pre-trained ViTs enables fast moving navigation in the real world Steeven Janny, Leonid Antsfeld, Christian Wolf
This paper shows that a single scalar extracted per image patch from a frozen, pre-trained vision transformer is sufficient to drive fast-moving robot navigation in real, physical environments, validated on a physical robot rather than in simulation alone. Only a small downstream layer is trained; the backbone is never fine-tuned, which meaningfully lowers the cost of deploying and updating strong visual navigation policies on a robot fleet.

IDeaL: data-free multi-teacher distillation via improved dead leaves Feyza Yavuz, Mert Bülent Sarıyıldız, Diane Larlus
IDeaL distills a single student encoder from several specialized teacher models without access to their original training data, using an improved dead-leaves image synthesis procedure to generate the training signal instead. This extends the data-free distillation setting to the multi-teacher case, letting a compact encoder inherit capabilities from multiple foundation models at once — a practical route to the kind of small, capable, easy-to-update perception encoder an on-board robotics stack needs, without depending on proprietary training data the robot’s builder may not have access to.

Task alignment: a simple proxy for practical model merging across diverse vision tasks Pau de Jorge, César Roberto de Souza, Björn Michele, Mert Bülent Sarıyıldız, Philippe Weinzaepfel, Florent Perronnin, Diane Larlus, Yannis Kalantidis
This work introduces a lightweight proxy metric that predicts whether two or more task-specific vision models can be merged into a single model without significant performance loss, before any merging is actually attempted. A robot’s perception stack typically runs several task-specific models such as detection, segmentation, depth, human pose side by side. Merging compatible ones into a single backbone is one of the more direct ways to cut the compute and memory budget on-board; this proxy turns that search from costly trial-and-error into something that can be screened for cheaply across many candidate pairs.

Human understanding for HRI

Human mesh modeling for Anny body (arXiv · blog · demo) Romain Brégier, Guénolé Fiche, Matthieu Armando, Laura Bravo-Sánchez, Thomas Lucas, Philippe Weinzaepfel, Grégory Rogez, Fabien Baradel
Anny is a free-to-use, open parametric 3D human body model built to cover the full spectrum of humanity, from infants to elders, with interpretable, easy-to-use shape and pose controls. This paper details the mesh modeling behind Anny, positioning it as an open foundation for the body shapes and poses a robot needs to represent in order to reason about the humans sharing its workspace for motion prediction, safe path planning around people and human-robot collaboration.

MuCHeR: multi-person camera-centric human detection, mesh recovery and tracking (arXiv · blog· demo) Guénolé Fiche, Philippe Weinzaepfel, Romain Brégier, Fabien Baradel
MuCHeR (aka Multi-HMR2) unifies multi-person detection, whole-body mesh recovery and tracking in a single camera-centric framework, extending the Multi-HMR line of work to handle crowded, dynamic scenes where people enter, exit and occlude one another across frames which is the kind of shared, unstructured space a robot operating alongside people, such as in an office or warehouse, actually has to perceive and react to in real time.

Collaborative research papers

Vulnerability of privacy-preserving visual localization against diffusion-based attacks Maxime Pietrantoni, Torsten Sattler, Gabriela Csurka
Privacy-preserving visual localization methods store obfuscated scene representations specifically so that the underlying imagery can’t be reconstructed. It is the approach a fleet of robots or AR devices would use to localize against a shared map of a home or office without exposing what that space actually looks like. This paper shows that diffusion-based generative models can partially reverse that obfuscation, recovering scene detail the representation was designed to hide. This is a result relevant to anyone building or relying on these pipelines for deployed, privacy-sensitive localization. Given the breadth of communities it touches, this paper is presented at three ECCV workshops: PPCV, PPVLM and ILR+G.

Syn4D: a multiview synthetic 4D dataset (project page) Zeren Jiang, Yushi Lan, Yihang Luo, Yufan Deng, Zihang Lai, Edgar Sucar, Christian Rupprecht, Iro Laina, Diane Larlus, Chuanxia Zheng, Andrea Vedaldi
Syn4D is a large-scale synthetic dataset of multi-view sequences with ground-truth dynamic 4D geometry, built to give the community a shared benchmark for dynamic scene reconstruction and generation. Real robot environments are rarely static with people and objects moving through them and ground-truth 4D geometry for moving scenes is scarce and hard to capture at scale, which is exactly the gap this dataset is meant to fill.

Where and what: long-term object tracking in egocentric videos (project page) Jacob Chalk, Saptarshi Sinha, Dima Damen, Yannis Kalantidis, Diane Larlus
This work (aka Whareformer) addresses long-term object tracking in egocentric video, where objects are routinely put down, occluded, handed off or moved out of frame for extended periods. The model jointly reasons over “where” an object is and “what” state it’s in, maintaining identity across these long gaps far better than standard short-horizon trackers. This is directly relevant to any robot or wearable assistant that needs to remember what happened to an object well after it last saw it, rather than just what’s in its current field of view.