STEM Approach to Enhance Robot-Human Interaction Through AI Large Language Models and Reinforcement Learning

• Published in: Integrated STEM Education Conference (Online) pp. 1 - 2
• Main Authors: Shibi, Siddhartha, Zaidi, Sohail
• Format: Conference Proceeding
• Language: English
• Published: IEEE 09.03.2024
• Subjects: Human-robot interaction; Humanoid; Humanoid robots; Large language models; LLMs; NLP; Oral communication; Reinforcement learning; RLHF; Robot vision systems
• ISSN: 2473-7623
• DOI: 10.1109/ISEC61299.2024.10665163

Humanoid Robots with their limitless capabilities have revolutionized the world. Their applications range from household assistance to advertising. As these technologies age however, the use of their sensors, motors, cameras, all become outdated; making previous humanoids a thing of the past. This project takes a STEM approach towards enhancing these robots by tackling the most crucial issue that such humanoids face; their adequacy in human-robot interactions. This study explores the promise of integrating LLMs (Large Language Models)-such as Google PaLM2 and ChatGPT-to supplement the capabilities of such robots; as well as bringing CoT (Chain of Thought). The subject of this project is the humanoid robot Pepper, by SoftBank Robotics, a popular robot designed to interact with humans; however, due to its weak natural language processing (NLP) capabilities, it struggles to adequately articulate responses in human-robot conversation. For instance, the robot was capable of easily listing responses to simple questions such as, "What is your name?", or "What are you", yet struggled with providing adequate responses to queries such as, "Who is the president of the United States", or "When is the next World Cup?". By AI/ML LLM integration, such questions were handled by using much improved LLMs in place of the previous built-in responses the robot had. This demonstration has been shown in the video that is uploaded at: https://youtu.be/hF7aRlQmnqs. Our approach targeted the main weak points of the robot; it's ability to provide responses to asked questions, and remembering prior questions/conversation. By intercepting the robot's own NLP Dialog Module, the asked prompt can be connected through a chatAdapter, bringing conversations to a chat database for context as well as LLM of choice. This approach, implemented through use of Android Studio to create an appropriate application for their procedure, addresses the contextual-based reasoning by pulling from the chat database as well as provides adequate responses limited only by the AI/ML model of choice. This project involved integrating ChatGPT /PaLM2 into Pepper's existing system to enable generation of more natural and engaging responses. In addition to the pre-existing development in bringing artificial intelligence into these humanoid robots, further work has been in the process; the aim being to develop a way for the robot to simultaneously extract other situational data from conversation such as facial and tonal expressions, bringing human feedback in order for responses to be further fine-tuned. Aside from the work-in-progress development of integrating RLHF (Reinforcement Learning with Human Feedback), the effectiveness of the aforementioned approach was further evaluated through a user study, comparing it with and without integration. The results indicated that integrating LLM/s into the robot's NLP system significantly improved its ability to generate more coherent responses, leading to more natural human-robot interactions. Overall, this presentation will demonstrate the potential of using LLMs to enhance the NLP capability of human robots like Pepper. It's believed that the proposed approach can pave the way for developing more intelligent human-robot interactions in the future.