Reproduce “Graphormer-Guided Task Planning: Beyond Static Rules with LLM Safety Perception” 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
Wanjing Huang, Tongjie Pan, Yalan Ye
Researchers from UCSB and UESTC developed a framework that integrates large language models (LLMs) with Graphormers to provide robots with real-time, structured safety perception in dynamic environments. This approach enables adaptive task replanning in response to emergent risks, achieving high hazard detection and mitigation rates in simulated household tasks within AI2-THOR.
Key insights
- Integrating Graphormers provides LLM-driven robots with robust, real-time structured safety perception that goes beyond static rules.
- A spatio-semantic safety graph, combined with LLM-generated risk annotations, offers an interpretable and adaptive representation of environmental hazards.
- The framework successfully bridges the gap between flexible high-level LLM planning and robust, real-time safety compliance by enabling online hazard detection and adaptive task refinement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.