D2S: 강화학습을 이용한 매니퓰레이터 제어를 위한 보상 빈도 기반의 커리큘럼 학습 개발

Abstract

In this paper, we present D2S(Dense-to-Sparse) curriculum learning to improve learning efficiency when using reinforcement learning to perform robotic manipulation tasks. We use D2S curriculum learning to perform robotic manipulation tasks on PyRep based RLBench. Firstly, we compared the learning results with using sparse reward alone and with using both sparse reward and distance-based dense rewards too. Next, we set up five phases with different frequencies of the reward and designed three experiments by combining the phases so that the robot can learn the task with gradually increasing difficulty. D2S curriculum learning allowed us to succeed the Reaching task, which was not well-trained when using a sparse reward. However, dividing the phases into three or more does not improve learning efficiency. In the future, we plan to optimize the condition where phases of the D2S curriculum learning ends and use it to perform more complex robotic manipulation tasks.