Reproduce “NavOne: One-Step Global Planning for Vision-Language Navigation on Top-Down 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
Dijia Zhan, Jinyi Li, Chenxi Zheng, Shaoyu Huang, Yong Li, Jie Tang, Xuemiao Xu
NavOne introduces a one-step global path planning framework for Vision-Language Navigation (VLN) using pre-built multi-modal top-down maps, reformulating the task to directly predict complete navigation paths. This approach achieves up to an 8x speedup over existing map-based methods and an 80x speedup over egocentric methods in planning, while improving success rates by up to 3 percentage points on the challenging R2R-TopDown Val Unseen split.
Key insights
- Reformulating VLN as one-step global planning on top-down maps significantly boosts computational efficiency and mitigates error accumulation inherent in sequential decision-making.
- The combination of Full Attention Residuals and Spatial-Aware Depth Queries within the ViT encoder enhances generalization by enabling more effective position-dependent feature mixing across different architectural depths.
- Effective fusion of multi-modal map information (RGB, occupancy, semantic) is crucial for robust top-down VLN, with an orthogonality analysis supporting the parameter-free element-wise fusion of pose data.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.