2P1-G09 BRLを用いた移動ロボット群の協調箱押しタスクにおける段階的な行動獲得

概要

論文の詳細を見る
This paper investigates incremental learning ability of our proposed reinforcement learning technique, BRL, where a multi-robot system faces a sequence of progressively more complex tasks. BRL has a mechanism for segmenting continuous state and action spaces adaptively, and is proven to be useful for the behavior acquisition in not single-robot systems but also multi-robot systems. Our previous work also shows, by means of the adaptive segmentation, BRL has high robustness against an environmental change. In other words, after the environmental change, BRL robots are expected to accelerate learning by reaching a situation where robots have experienced. Physical experiments of a box pushing task by three mobile robots are conducted. We examine how BRL robots utilize their knowledge acquired in the previous environments.
一般社団法人日本機械学会の論文