Total Singulation With Modular Reinforcement Learning

Prehensile robotic grasping of a target object in clutter is challenging because, such conditions, the touches other objects, resulting to lack collision free grasp affordances. To address this problem, we propose modular reinforcement learning method which uses continuous actions totally singulate from its surrounding clutter. A high level policy selects between pushing primitives, are learned separately. Prior knowledge effectively incorporated into learning, through action primitives and feature selection, increasing sample efficiency. Experiments demonstrate that proposed considerably outperforms state-of-the-art methods singulation task. Furthermore, although training performed simulation robustly transferred real environment without significant drop success rate. Finally, tasks different environments addressed by easily adding new primitive retraining only policy.