LLM Text Generation for Service Robot Context

Service robots are designed to perform useful activities for humans or other machines, such as cleaning a house, providing guidance, or cooking. To perform these activities, robots must be increasingly autonomous and capable of executing physical and cognitive tasks. Such skills involve human-robot interaction, where humans give a specific command, and the robot performs the task. In this scenario, voice recognition and detection tools can ensure communication through natural language. However, obtaining comprehensive and contextualized answers to specific questions or topics becomes challenging for the robot, which relies on static and pre-existing information to generate rigid and repetitive responses. This limitation hinders the creation of a more natural and intuitive dialogue between humans and machines. This paper proposes the LLP4ServiceRobot solution, which uses fine-tuned large language models to generate more dynamic texts and responses. The development methodology involves selecting large language models, fine-tuning them, and training them for text generation and question-answering tasks. We validated our model using two performance metrics: the Fl-Score, for the question-answering model; and ROUGE, for the text generation model. Results showed that the question-answering model achieved an F'l-Score of 52%, while the text generation model achieved a ROUGE-l score of 67%. Additionally, an experiment was conducted to evaluate the Question Answering (QA) model. Results showed that the model could correctly answer 60 % of the questions in a test dataset. The proposed model will be used in the service robot BILL, part of the RoboCup@Home league research at ACSO - the Center for Research in Computer Architecture, Intelligent Systems, and Robotics at the State University of Bahia - UNEB.