Multi-task learning from demonstration via embedded network

Abstract

With the development of the intelligent home service robot, autonomous-learning applied in the field of robot has aroused considerable attentions of researchers. Home service robot is hoped to help his master to do so many trivial. Nowadays, robot already can learn a new skill autonomously rather than set by programs. However, when meeting multi-task at the same time, there are still so many things to do, the robot need to decide whether going on the current task, or suspending it turning to the next task. This is an arbitrate problem, and for a person who meet the same situation in reality scene he may make his decision subconsciously. Teaching the robot to learn to deal with multi-task is an urgent problem. In order to solve the multi-task issue, an embedded neural network is put forward. The embedded neural network consists of RNN which is equipped with a unique memory module and CNN is embedded in each node of the special RNN network. The single node of the RNN represents a task, and the RNN expresses the transference between tasks. An additional input image works for scene shifting. The new structure of neural network has been proved invaluable in multi-task learning.