Reproduce “VLG-Loc: Vision-Language Global Localization from Labeled Footprint Maps” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Mizuho Aoki, Kohei Honda, Yasuhiro Yoshimura, Takeshi Ishita, Ryo Yonetani

VLG-Loc enables robotic global localization using human-readable labeled footprint maps by integrating Vision-Language Models (VLMs) within a Monte Carlo Localization framework. The fused vision-language and scan-based approach achieved 0.52m translational error in a real retail environment, outperforming scan-only methods which recorded 1.97m for MCL and 2.31m for CBGL.

Key insights

  • Vision-Language Models can effectively ground abstract, human-defined map labels in real-world visual observations, enabling localization without explicit geometric or appearance descriptors.
  • The fusion of semantic understanding from VLMs with precise geometric information from LiDAR provides a highly robust localization solution, overcoming the limitations of single-modality approaches.
  • Utilizing low-maintenance, text-based labeled footprint maps offers a practical alternative to complex dense maps, adaptable to dynamic environments with minimal update effort.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.