Decision making on robot with multi-task using deep reinforcement learning for each task
Yuya Shimoguchi, Kentarou Kurashige · 2019
In recent years, a robot is required to perform multi-task autonomously in human living space. It needs to take actions according to situations. We proposed a method which does decision making on a robot with multi-task according to a situation by using an importance of each task. To respond to changes in importance of task, the robot learned each task independently by using reinforcement learning. An action is selected uniquely using action values and importance of each task in this system. In reinforcement learning, a learning space is constructed by Q-table. When the design of states is detailed, the learning times increase sharply in a case of Q-table. Therefore, a state which a robot recognizes is not detailed in reinforcement learning. There are some researches to adapt to detailed states using deep reinforcement learning. However, it is difficult to make decisions when considering multi-task. In this paper, we propose a method which does decision making on a robot with multi-task using deep reinforcement learning for each task. As an approach, the learning method uses Deep Q-Network for each task, and the decision-making uses importance of each task. This allows the robot to recognize a continuous state and select action according to a situation in multi-task. We carried out an experiment which set three tasks to the robot applied proposal method. From experimental results, we confirmed the usefulness of proposed method.