Reproduce “Uncertainty-Aware Gaussian Map for Vision-Language Navigation” 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
Jianzhe Gao, Rui Liu, Yuxuan Xu, Tongtong Cao, Yingxue Zhang, Zhanguang Zhang, Sida Peng, Yi Yang and 1 more
An uncertainty-aware framework for Vision-Language Navigation (VLN) is introduced, utilizing a 3D Gaussian Map to integrate geometric, semantic, and appearance uncertainties into a unified 3D Value Map. This approach enhances navigation robustness by explicitly modeling perceptual ambiguity, leading to a 3.94% improvement in Remote Grounding Success (RGS) on the REVERIE benchmark and a 2% increase in Success Rate (SR) on R2R.
Key insights
- Explicitly modeling geometric, semantic, and appearance uncertainties provides valuable reliability cues for perception and action feasibility, significantly improving decision-making in VLN.
- Leveraging the explicit and interpretable structure of 3D Gaussian Splatting allows for granular, grounded, and region-specific uncertainty estimation tied to physical attributes.
- Consolidating distinct uncertainty types into a unified 3D Value Map effectively transforms these estimates into actionable affordances and constraints for more informed and reliable navigation trajectories.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.