基于大语言模型的反馈纠正下机器人组装任务规划与实现OA

To address the challenges of operational step planning errors and incorrect recognition when applying vision-enabled large language models(LLMs)directly to multi-step automated assembly tasks,a multi-step task planning method integrating multi-modal prompts with feedback correction is proposed.First,an object feature-based chain-of-thought reasoning approach is proposed,enabling LLM to accurately recognize complex objects through single-step inference.A multi-round interactive prompting framework is further introduced to enhance the recognition stability of LLMs for input tasks(assembly manuals).Subsequently,a robotic planning method is designed by incorporating skill prompts for assembly tasks and integrating CS-FOON-based planning verification and feedback mechanisms,achieving accurate and stable multi-step robotic automated assembly task planning.The effectiveness of the method is validated through the construc-tion of multiple assembly task scenarios in the Unity simulation environment and the execution of assembly operations using robotic arms.The results show that,compared to other planning methods based on large language model,this frame-work enables LLM to achieve a 98.7%success rate in tool identification across multiple tasks,with an average improve-ment of 22.0 percentage points,and a 84.0%success rate in task planning,with an average improvement of 45.0 percent-age points.