Reproduce “Scalable Task Planning via Large Language Models and Structured World Representations” 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

Rodrigo Pérez-Dattari, Zhaoting Li, Robert Babuška, Jens Kober, Cosimo Della Santina

A framework combines Large Language Models with structured world representations to achieve scalable robotic task planning by pruning irrelevant objects and relationships from the state space. This approach reduced state space from hundreds to dozens of objects, enabling successful planning for complex tasks that were otherwise intractable and demonstrating real-world execution on a 7-DoF manipulator.

Key insights

  • Large Language Models (LLMs) can effectively serve as intelligent agents for pruning irrelevant state space components, leveraging their commonsense knowledge for zero-shot object relevance identification.
  • Integrating LLMs with structured graph states and object taxonomies allows for scalable interaction with complex environments and grounding LLM predictions in concrete world representations.
  • Significant state-space reduction is achievable even for multi-subtask problems, transforming computationally intractable planning problems into manageable ones.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.