This academic article explores how to enhance the teachability of large language models (LLMs) used for writing robot codes. These LLMs allow non-experts to direct robot behaviours, modify them based on feedback, or compose them to perform new tasks. However, their capabilities are limited to short-term interactions, as they can forget user feedback over longer interactions. The paper proposes fine-tuning these LLMs to remember their in-context interactions and improve their adaptability to human inputs. The authors introduce a framework called Language Model Predictive Control (LMPC) that fine-tunes the models to improve their teachability across various tasks and robot embodiments, resulting in a significant improvement in non-expert teaching success rates and a reduction in the average number of human corrections.
Publication date: 20 Feb 2024
Project Page: Not provided
Paper: https://arxiv.org/pdf/2402.11450
