[2O1-GS-8] Robot and real worlds

[2O1-GS-8-05] Efficient Multiple Task Robot Learning Guided by Large Language Models

〇Shota Takashiro¹, Shohei Taniguchi¹, Akihiro Nakano¹, Yusuke Iwasawa¹, Masahiro Suzuki¹, Wataru Kumagai¹, Hitomi Yanaka¹, Yutaka Matsuo¹ (1. Tokyo University)

Keywords:Large Language Model, Reinforcement Learning, Imitation Learning, Robotics Manipulation

Large language models, such as GPT-3 and ChatGPT, show high general performance in various tasks, and are widely applied not only to natural language processing but also to various other domains. In this paper, we utilize large language models for imitation learning of robot control and examine their contribution to improving learning efficiency and sample efficiency. In our experiments, we verify the effectiveness of the proposed method using a benchmark dataset called RLBench. When learning multiple tasks, we utilize prompting to the large language model to generate text that explains the procedure for solving the tasks and use it as a subgoal to improve learning efficiency and generalizability.