Interactive Robot-Robot Reinforcement Learning for Object Balancing Task

摘要

Robots with machine learning are expanding their application fields, such as serving robots and guiding robots, but applying machine learning to robots has a high labor cost due to human intervention. This paper proposes interactive robot-robot reinforcement learning technology to minimize human intervention in learning tasks that successfully balance objects between two robots. In this proposed technique, the teacher and student robot guide and learn object balancing based on reinforcement learning. Teacher robot teaches knowledge that has not been learned in advance to student robot, and teaches more effectively by providing various positive and negative text feedback on the training results. In virtual simulations, the training results of the student robot converge to the optimal policy as an evaluation result of the proposed method, and the trained student robot performs appropriate actions according to various table states by the teacher’s random action. Through our experimental results, the application possibility of robot-robot learning is verified being as good as human-robot learning. Our proposed method could be utilized to seamlessly transfer the learned knowledge when it is challenging to apply the learning model to a heterogeneous robot or an agent in an IoT environment.