反復自動データ収集を用いた模倣学習による環境設備操作タスクの実現

抄録

When a new robot is introduced into a human-activated environment, there is a need for a function that enables users to teach the robot the tasks and motion skills they want it to perform on the spot. Since imitation learning, a typical skill learning method, has the problem of high cost for teaching repetitive actions, we propose a skill learning method that uses data automatically collected by a robot that iteratively executes actions based on a single action instruction from the user. In this study, we have classified the properties of repetitive and automatic data-collection tasks into two: ease of locating and ease of detecting failures. By learning the motion skills of these properties, a series of environmental equipment manipulation tasks can be robustly executed. The effectiveness of the proposed system was confirmed through experiments on actual robot.