Paper Title
Learning Pushing and Grasping via Deep Reinforcement Learning and Grasping Potential Module

Abstract
Robotic arms are often used for grasping objects, but grasping in cluttered scenarios may lead to collisions due to insufficient space. Pushing can help scatter cluttered objects to make space for robotic arm hence increasing the grasp success rate. In this paper, we propose a method that combines deep reinforcement learning with instance segmentation networks to achieve collaborative pushing and grasping. In our work, we use a grasping potential module to estimate the grasp potential of different objects in the environment.The agent prioritizes grasping high-potential objects, allowing it to quickly learn the pushing and grasping policies in complex environments. Additionally, a novel pushing strategy is proposed to prevent ineffective pushing. Experimental results validate that our method achieves a 93% grasp success rate and converges faster than other baseline models. Keywords - Deep Reinforcement Learning, Instance Segmentation, Grasping