Reproduce “Bootstrapping Object-level Planning with Large 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
David Paulius, Alejandro Agostini, Benedict Quartey, George Konidaris
A method for robot planning called Object-Level Planning (OLP) integrates large language models with hierarchical Task and Motion Planning by having LLMs generate abstract object-level plans, which are then grounded into PDDL subgoals. This approach achieves higher plan completion (80-86%) and success rates (62-77.14%) in simulation compared to direct LLM planning or LLM-generated PDDL baselines.
Key insights
- Large language models are better suited for reasoning about abstract object interactions and state changes rather than generating precise, embodied robot actions or complex PDDL syntax.
- Introducing an object-level planning layer effectively leverages LLM strengths while offloading the complexities of physical embodiment and syntactic correctness to classical planners.
- Structured representations, specifically Functional Object-Oriented Networks (FOONs), provide a robust framework for LLMs to express task-relevant knowledge that can be systematically grounded.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.