The paper introduces ROBOFUME, a system that allows robots to learn new tasks autonomously with minimal human intervention. The system pre-trains a multi-task manipulation policy using diverse datasets, and then self-improves online to learn a target task. The system also uses pre-trained vision language models to build a robust reward classifier for providing reward signals during the online fine-tuning process. The researchers demonstrated that their method can improve on a target task within as little as 3 hours of autonomous real-world experience, and also outperforms prior works in simulation experiments.
Publication date: 24 Oct 2023
Project Page: https://robofume.github.io
Paper: https://arxiv.org/pdf/2310.15145
