Portrait de Jana Pavlasek

Jana Pavlasek

Membre académique associé
Polytechnique Montréal, Département de génie informatique et génie logiciel
Sujets de recherche
Apprentissage profond
Inférence bayésienne
Planification
Robotique
Vision par ordinateur
Site web
Google Scholar
LinkedIn

Étudiants actuels

Haechan Mark Bong
Doctorat - Polytechnique
Superviseur⋅e principal⋅e :
Giovanni Beltrame
Site web
Google Scholar
Yann Bouteiller
Collaborateur·rice de recherche - Polytechnique Montreal
Superviseur⋅e principal⋅e :
Giovanni Beltrame
Site web
Github
Google Scholar
Nicolas Fleury-Rousseau
Doctorat - Polytechnique
Github
Maeva Guerrier
Doctorat - Polytechnique
Superviseur⋅e principal⋅e :
Giovanni Beltrame
Github
Edgar Kappauf
Maîtrise recherche - Polytechnique
Olivier Lessard
Doctorat - Polytechnique
Superviseur⋅e principal⋅e :
Chris Pal
Site web
Google Scholar
Simon Roy
Maîtrise recherche - UdeM
Superviseur⋅e principal⋅e :
Giovanni Beltrame
Site web
Github
Soma Soma
Doctorat - Polytechnique
Superviseur⋅e principal⋅e :
Giovanni Beltrame

Publications

Can Vision Foundation Models Navigate? Zero-Shot Real-World Evaluation and Lessons Learned
Maeva Guerrier
Karthik Soma
Jana Pavlasek
Giovanni Beltrame
Visual Navigation Models (VNMs) promise generalizable, robot navigation by learning from large-scale visual demonstrations. Despite growing … (voir plus)real-world deployment, existing evaluations rely almost exclusively on success rate, whether the robot reaches its goal, which conceals trajectory quality, collision behavior, and robustness to environmental change. We present a real-world evaluation of five state-of-the-art VNMs (GNM, ViNT, NoMaD, NaviBridger, and CrossFormer) across two robot platforms and five environments spanning indoor and outdoor settings. Beyond success rate, we combine path-based metrics with vision-based goal-recognition scores and assess robustness through controlled image perturbations (motion blur, sunflare). Our analysis uncovers three systematic limitations: (a) even architecturally sophisticated diffusion and transformer-based models exhibit frequent collisions, indicating limited geometric understanding; (b) models fail to discriminate between different locations that are perceptually similar, however some semantics differences are present, causing goal prediction errors in repetitive environments; and (c) performance degrades under distribution shift. We will publicly release our evaluation codebase and dataset to facilitate reproducible benchmarking of VNMs.
2026-03-25
arXiv (prépublication)
doi.org
arxiv.org