Reproduce “SOON: Scenario Oriented Object Navigation with Graph-based Exploration” 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
Fengda Zhu, Xiwen Liang, Yi Zhu, Xiaojun Chang, Xiaodan Liang
The paper introduces Scenario Oriented Object Navigation (SOON), a task where agents locate target objects using rich, target-oriented language instructions from arbitrary starting points in 3D environments, and presents the From Anywhere to Object (FAO) dataset. It proposes a Graph-based Semantic Exploration (GBE) method, which achieves 1.4% SFPL on the FAO test set and improves performance on the R2R dataset by 5.4% SPL over baselines.
Key insights
- The SOON task and FAO dataset highlight a significant gap between current embodied navigation capabilities and human-like exploration and understanding, with human performance vastly exceeding models.
- Rich, hierarchical language descriptions (object attributes, relationships, region, nearby regions) are crucial for unambiguous target identification and effective navigation, outperforming simple object-goal approaches.
- Integrating a dynamic graph-based memory with a novel exploration strategy that learns from both optimal and sub-optimal trajectories significantly improves policy robustness and generalization for long-horizon navigation tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.