2A1-A20 他者の動作から発見した運動プリミティブの組み合わせによる模倣動作の生成 : 自他の対応関係を伴った動作の分節化

概要

論文の詳細を見る
It is generally thought that we must give some rules to robot to learn imitational behaviors, such as coordinate transformation. But this study shows that a robot can perform imitational behavior by combination of action primitives. Our robot experiment suggests that with following three requirement robot can generate imitational behavior: (1) The robot has ability to segment behaviors and combine action primitives. To achieve this, we used Multiple Time-scale Recurrent Neural Network (MTRNN) as a controller of robot and statistical machine learning methods to learn own and other's behaviors; (2) The robot has experienced all action primitives needed for generation of imitational behavior, through learning of behaviors which are combination of the primitives; (3) And the self-other correspondence of behaviors are given. We used so-called "bind" learning methods to give robot the correspondence. Furthermore, analysis of the MTRNN shows that slow time scale neural units represent combination of the primitives, not the behavior directly.
一般社団法人日本機械学会の論文
2009-05-25