An Exploratory Guided Push and Grasp

Abstract

Teaching a robot to grasp cluttered and stacked objects is a common and challenging but important research topic that can potentially benefit many real-life applications. The challenge can be addressed by incorporating pushing action into grasping strategy. To make robot capable of pushing and grasping (PG) stacked objects, deep neural network (DNN), typically reinforcement learning, has been reported widely, while the limitations of long training time and low action efficiency are still obvious. In this work, an exploratory guided pushing and grasping approach was proposed using a self-supervised technique to expedite training cycle, and a memory buffer enhancement strategy to guide the training continuously. The network architecture, data collection and learning strategy were investigated and analysed to study their effects. Comprehensive experiments were conducted on both simulation and real environment to validate the proposed solution with high action efficiency, outperforming the state-of-the-art solutions with a large margin.