Reproduce “GAMap: Zero-Shot Object Goal Navigation with Multi-Scale Geometric-Affordance Guidance” 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
Shuaihang Yuan, Hao Huang, Yu Hao, Congcong Wen, Anthony Tzes, Yi Fang
GAMap guides robots searching for unseen object categories using language-generated geometric-part and affordance descriptions, scored by pretrained CLIP across multiple image scales. It projects scores into a map and selects frontiers using FMM. On HM3D, GAMap reports 53.1% success rate and 26.0% SPL, versus 50.9% and 23.6% for VLFM value maps with FMM; the comparison does not isolate individual components.
Key insights
- Integrating distinctive geometric parts and functional affordance attributes provides a more robust and human-like understanding of objects for navigation, especially under partial observability.
- A multi-scale perception mechanism effectively captures both global context and fine-grained visual details, enhancing object recognition and localization across different distances and viewing angles.
- Foundation models (LLMs for attribute inference, VLMs for visual-text scoring) can be effectively combined to bridge high-level symbolic reasoning with low-level visual perception for enhanced embodied AI tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.