The Robot’s Understanding of Classification Concepts Based on Large Language Model

Cognition of classification concepts empowers robots to enhance their planning and task execution capabilities. By incorporating classification cognition of the environment and tasks, robots can make more informed decisions and improve their autonomy. In various domains, specific concept classification is crucial, such as disease categorization in medical diagnosis, transaction classification in the financial field, etc. Therefore, this article conducts an in-depth exploration of the cognition and understanding of abstract classification concepts (object color, object type, object size, etc.), and builds a programmed classification structure based on a large language model (PCS-LLM) to improve the robot’s ability to classify the different categories. First, through the programmed definition of interactive objects and executable action libraries in the scene, constraints are added to the prediction of general LLM, thereby improving the executability of predicted actions. Secondly, for abstract classification concepts, typical tasks and decomposed actions performed by robots are programmatically defined as an example of LLM generating new task decomposed actions. Finally, in mixed scenarios, the robot’s understanding and decomposition of any concept combination task is verified based on the UR5 robotic arm. Experimental results show that the robot can achieve corresponding classification operations when faced with tasks involving any combination of concepts.