Reproduce “ViTL: Temporal Logic-Guided Zero-Shot Natural Language Navigation via Vision-Language Models” 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
Kaier Liang, Hengde Dai, Cristian-Ioan Vasile
The ViTL framework from Lehigh University enables zero-shot robot navigation for long-horizon natural language commands containing complex temporal and logical constraints in unknown environments. It achieves a 95.3% task completion rate for multi-target tasks, showing a 15.31% reduction in path length compared to LLM-based planners, and demonstrates competitive performance for single-target ObjectNav with a 54.1% SR and 30.9 SPL.
Key insights
- Translating natural language commands into LTL and then a DFA provides a formal, provably correct mechanism for ensuring temporal and logical constraint satisfaction in complex multi-step navigation, outperforming incremental LLM-based planning.
- Dynamically weighting DFA edges based on environmental discoveries allows for efficient, opportunistic replanning and sub-task reordering, leading to shorter overall path lengths.
- Leveraging VLMs to generate 'directional scores' for distinct frontier candidates offers more granular and effective guidance for zero-shot exploration compared to direction-agnostic scene scores, enhancing single-target navigation performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.