Host: The Japan Society of Mechanical Engineers
Name : [in Japanese]
Date : June 05, 2019 - June 08, 2019
Various objects were successfully manipulated in our previous research. However, the network had to be trained for each different motion. Therefore, there is a hardware load for getting training data for each motion. Specifically, four-fingered in-hand manipulation is difficult to control because of a high number of joints. This paper suggests a method that reduces the required training data for in-hand manipulation with the concept of pre-training and mutual finger motions. The training data included various sized and shaped objects for making the network more versatile. After pre-training the network, one shot learning was used to do training with a new task; mutual finger motions can be used with 3-fingered pre-training data for 4-fingered manipulation. Importantly, pre-training data from fingers with the same kinematic chain is required. As a result. the importance of morphology specific learning was confirmed. Moreover, untrained sizes and shapes of objects could be manipulated.
