Practical learning synergies between pushing and grasping based on DRL

Abstract

This paper focuses on the comparison of performance of the intelligent robot manipulation systems based on different deep reinforcement learning technologies. An ideal strategy for robotic manipulation involves two primary components: non- prehensile actions, such as pushing, and prehensile actions, such as grasping. Both pushing and grasping play pivotal roles in enhancing the efficiency of robotic manipulation. Pushing actions can effectively separate clustered objects, creating room for robotic grippers to grab target item, while grasping can assist in relocating items to facilitate more precise pushing and prevent collisions. Therefore, it is essential to explore synergies between these two fundamental actions. Reviewing literature in related fields reveals that developing synergies between pushing and grasping from the ground up using model-free deep reinforcement learning is feasible. One successful approach involves training two neural networks simultaneously within a standard DQN framework, relying entirely on self-supervised learning through trial and error, with rewards contingent upon successful grasps. Drawing inspiration from this effective methodology, this paper introduces the dueling DQN framework as a potential enhancement, aiming for integrated performance comparison. More complicated reinforcement learning frameworks can help boost the overall performance of self-supervised robotic manipulation systems. Most importantly, these experiments can work as a clear demonstration of the connection between the efficiency of intelligent robotic manipulations and the complexity of deep reinforcement learning models, which will inspire people to think about other more advanced DRL algorithms.