SC-RAG-CoT: An Optimization Method for Robotic Instruction Generation

Large Language Models (LLMs) based robotic instruction generation and control methods provide flexible and efficient solutions for autonomous robot decision-making with natural language interaction and task generalization. However, traditional LLMs struggle to adapt to the dynamic physical environments and address the ambiguity of natural language instructions. Therefore, a mechanism that integrates physical environment information and guides users to clarify specific needs through human-computer interaction is required. To this end, we propose the SC-RAG-CoT framework, which combines spatially constrained retrieval augmented generation (SC-RAG) and Chain of Thought (CoT) interaction for robotic instruction generation. SC-RAG-CoT employs SC-RAG to retrieve contextual information from spatial knowledge bases (e.g., spatial data, object attributes) to mitigate hallucinations caused by environmental dynamics. Additionally, CoT decomposes ambiguous instructions into multi-step reasoning chains. The reliability of instruction generation is improved through explicit verification of intermediate logic, and users are guided to gradually clarify their needs through interactive mechanisms. Simulation and ablation experiments demonstrate that the proposed SC-RAG-CoT strategy reduces average communication rounds by 2.6 and failure rates by $25.3 \%$ compared to traditional LLM-based methods in standardized test scenarios.