Reproduce “GraphThink: Graph-Enhanced LLM Thinking for Long-Horizon Embodied Task Planning” 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
Chen Li, Sijie Cheng, Yuelin Zhang, Junxi Li, Maozhi Huang, Yang Liu, Wenbing Huang
GraphThink, a dual-graph enhanced LLM planning system, guides large language models for long-horizon embodied tasks using structured task knowledge and dynamic environmental memory. The system achieved a 68.52% success rate on ALFRED 'Tests Unseen', outperformed baselines on custom long-horizon tasks with 90.04% planning success, and demonstrated strong generalization to novel tasks and environments.
Key insights
- Integrating explicit task graph knowledge with LLM planning significantly grounds the planning process, mitigating physical hallucinations and enabling more logically consistent subtask sequences.
- Designing reinforcement learning rewards based on a task graph, which accounts for multiple valid solution paths, enhances an LLM's generalization capabilities for embodied tasks, reducing reliance on exact-match supervision.
- Utilizing a dynamic, task-aware scene graph as environmental memory allows for a closed-loop system that can proactively identify and correct semantic misalignments and execution failures, leading to more robust and adaptive task completion.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.