Integration of dynamic knowledge and LLM for adaptive human-robot collaborative assembly solution generation

The active adaptability of robots has always been a central challenge and focus in human-robot collaboration research. In complex product assembly scenarios, humans often struggle to provide robots with clear or sufficient natural language instructions. To enhance a robot's adaptive capability, it is essential to incorporate dynamic collaborative context information, such as assembly requirement changes, object positional adjustments, and product status evolution. Traditional approaches that establish static knowledge bases are more suitable for simple and localized context tasks but overlook the fact that in collaborative assembly tasks, historical contextual knowledge accumulates rapidly. This accumulation makes it increasingly challenging to extract effective knowledge from large volumes of historical data and reduces interference from irrelevant contextual prompts. To address this issue, this paper proposes an adaptive method for intelligent generation of Human-Robot Collaborative Assembly (HRCA) programs by fusing dynamic knowledge and Large Language Models (LLMs). The method generates collaborative assembly solutions based on the logic of knowledge modeling, knowledge evolution, and knowledge enhancement. Specifically, a dynamic knowledge evolution mechanism for HRCA is designed, which establishes a memory iteration loop for dynamic contextual requirements and historical scene states. This loop provides LLMs with comprehensive prompt texts that balance current contextual demands with scene states, reducing the interference of irrelevant information and improving the accuracy and consistency of the generated solutions. The proposed method is applied to a complex product's HRCA, and its performance is compared with various baseline methods. The results show that the proposed method significantly enhances the accuracy of multi-step reasoning in HRCA, with accuracy and consistency in the generated solutions for different collaboration modes approaching 90%, thereby validating the effectiveness of the proposed method.