Reproduce “Embodied Instruction Following in Unknown Environments” 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
Zhenyu Wu, Ziwei Wang, Xiuwei Xu, Hang Yin, Yinan Liang, Angyuan Ma, Jiwen Lu, Haibin Yan
A framework is presented for embodied agents to execute complex human instructions in previously unknown environments by dynamically constructing semantic maps and generating plans based on currently perceived objects. The method achieves a 45.09% success rate in large house-level scenes, demonstrating improved scalability and performance over existing approaches.
Key insights
- Dynamically constructing a semantic representation of an unknown environment allows LLM-based planners to generate feasible actions grounded in perceived reality.
- A hierarchical VLM architecture effectively bridges high-level instruction comprehension with low-level action execution and scene exploration.
- Directing attention to task-relevant visual clues through "dynamic region attention" is crucial for efficient information processing and planning in large, unknown scenes.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.