Scholarlink

Long-horizon task planning is essential for robotic autonomy, yet LLM-based agents often generate plans that are logically inconsistent or physically infeasible. We propose a systematic framework that combines rule-guided Chain-of-Thought prompting with knowledge graph (KG)-based symbolic validation. This approach enables agents to decompose complex instructions into coherent subgoals, reason about dependencies, and produce executable action plans while the KG enforces object–action relationships and environmental constraints. Experiments in VirtualHome show that our method raises task success rates from 34% to 78%, substantially outperforming existing LLM-based baselines in both plan quality and execution reliability.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

LLM-KGPlan: Long-Horizon Task Planning via Knowledge-Guided Reasoning

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Wei Fang,
  • Dingyu Yang,
  • Niansheng Chen,
  • Guangyu Fan,
  • Lei Rao,
  • Songlin Cheng,
  • Xiaoyong Song,
  • Yingzhou Yu

摘要

Long-horizon task planning is essential for robotic autonomy, yet LLM-based agents often generate plans that are logically inconsistent or physically infeasible. We propose a systematic framework that combines rule-guided Chain-of-Thought prompting with knowledge graph (KG)-based symbolic validation. This approach enables agents to decompose complex instructions into coherent subgoals, reason about dependencies, and produce executable action plans while the KG enforces object–action relationships and environmental constraints. Experiments in VirtualHome show that our method raises task success rates from 34% to 78%, substantially outperforming existing LLM-based baselines in both plan quality and execution reliability.