Reproduce “Multi-Level Compositional Reasoning for Interactive Instruction Following” 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
Suvaansh Bhambri, Byeonghwi Kim, Jonghyun Choi
MCR-Agent from Yonsei University introduces a three-level hierarchical learning framework for interactive instruction following, achieving 2.03% absolute gain in Path Length Weighted Success Rate on unseen ALFRED environments compared to similar state-of-the-art methods, without relying on external supervision or rule-based planning. This modular approach disentangles navigation from object interaction, leading to improved generalization and faster learning for complex household tasks.
Key insights
- A deeper, three-level hierarchical policy with distinct modules for high-level planning, navigation, and specific interactions improves performance and generalization for complex tasks.
- Explicitly disentangling navigation and interaction policies into specialized modules significantly enhances learning efficiency and task success, particularly by avoiding issues like sample imbalance.
- Identifying the target object for navigation via an Object Encoding Module is crucial for efficient and accurate navigation to interaction points, acting as a direct navigational subgoal.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.