A Grasping Posture Selection Method Based on Large Language Model

Large Language Models exhibit excellent generalization capabilities, particularly in their ability to semantically analyze and perceive new concepts. This potential can be harnessed to enhance the cognitive abilities of robots in tasks such as grasping decision-making. However, despite the semantic understanding LLMs provide, the process of transitioning from text-based instances generated by LLMs to actual item instance grasping decisions is fraught with challenges. These challenges include multi-instance selection and the absence of a unified framework paradigm that allows for modular and interchangeable decision-making processes related to grasping gestures. In this paper, a layered architecture is proposed that addresses these challenges by incorporating hot-swappable modules. This architecture integrates LLMs, visual perception, and grasping gesture decision-making to enable embodied intelligence to perform grasping tasks effectively. By utilizing the current pretrained foundation models, our approach benefits from the advantages of requiring no fine-tuning and permitting hot-pluggable modules. The feasibility of our architecture has been validated through real photo experiments, demonstrating its practical applicability.