[3O1-OS-16b] OS-16

[3O1-OS-16b-05] Improving Learning Efficiency in Compositional Robot Tasks Using Prior Knowledge of Large Language Model

〇Shota Takashiro¹, Tatsuya Matsushima¹, Yusuke Iwasawa¹, Yutaka Matsuo¹ (1. University of Tokyo)

5月31日（金）09:20~09:40：[4O1-OS-16d-02] の時間で発表

Keywords:RL, LLM, IL

Large language model have shown high general performance in various tasks, and their applications are expanding not only in natural language processing but also in various other fields. Although there are many existing studies that utilize large language model in robot control, most of them are used for action planning in compositional tasks, and fail if an action is selected that is not prepared in advance by the robot. In other words, the a priori knowledge in large-scale language models can be used for policy selection during inference, but it cannot be used during actual policy learning. In this paper, we aim to decompose a task using prior knowledge from a large language model and intensively reinforce learning only the failed steps, so that the robot can acquire a new strategy with minimal interaction with the environment.