時系列データによる動作生成のための負例を利用した模倣学習

抄録

Imitation learning is a method to learn motions by imitating data from human manipulations of a robot. However, since learning is performed using only successful data in general imitation learning, failure data is not used, and the cost of data collection is high. Failure data generated by imitation learning is based on human manipulation of the robot to obtain a successful motion, so it is similar to the success data and includes the actions that causes it to fail. It is thought that we can learn motions to avoid failures by performing imitation learning that predicts failure motions in advance and using a loss function for learning that closer to successful data and away from failure in other neural networks. In this study, we constructed an algorithm for imitation learning that increases the success rate of a motion using a negative example, and verified its effectiveness on an actual device.