Reproduce “FILM: Following Instructions in Language with Modular Methods” 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
So Yeon Min, Devendra Singh Chaplot, Pradeep Ravikumar, Yonatan Bisk, Ruslan Salakhutdinov
Researchers at Carnegie Mellon University and Facebook AI Research developed FILM, a modular architecture for Embodied Instruction Following that translates high-level instructions into structured subtasks and uses a semantic map with a dedicated search policy. This approach achieves a 24.46% success rate on the challenging ALFRED "Tests Unseen" benchmark, an 8.17% absolute improvement over prior state-of-the-art methods, by enhancing generalization and reducing reliance on costly expert data.
Key insights
- Explicit modularization with structured representations (semantic maps, subtask sequences) significantly enhances an agent's ability to understand, plan, and generalize compared to end-to-end methods.
- A dedicated semantic search policy effectively addresses the challenge of locating small, unobserved objects by predicting their likely positions, improving exploration efficiency.
- The design inherently reduces the need for expensive expert trajectories and low-level language instructions, making the approach more data-efficient and scalable.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.