ディープニューラルネットワークとリカレントニューラルネットワークによる把持位置を考慮した道具使用モデル

抄録

We propose a tool-use model considering grasping position introducing deep neural network (DNN) and recurrent neural network (RNN). In previous studies, one grasping position was designed for using the tool. Also, labelling or modeling of the tools was needed. In this research, a robot learns how to use tools with several grasping positions without labelling or modeling of tools through the tool-use experience that are considered several grasping positions and shaped tools. To evaluate the model, the robot generates the motion to manipulate an object with unknown grasping positions from only target image and initial state data. The results show that robot was able to recognize function of tools considering grasping-position, even when showing unknown grasping positions.