Grasp Planning Based on Deep Reinforcement Learning: A Brief Survey

Grasping is a fundamental ability for robots to inter-act with the environment. Recent advances show great progress on grasping unstructured household objects by leveraging deep reinforcement learning. This paper presents a brief survey on grasp planning based on deep reinforcement learning. The goal of grasp planning is to determine the actions of the robotic manipulator moving from its initial position to the target one so that the object could be grasped and manipulated. However, objects could be stacked or occluded, making the target object ungraspable by applying a simple grasp action. To tackle this issue, auxiliary actions (such as pushing) should be conducted along with grasping. According to whether auxiliary actions are performed, we categorize the existing grasping planning methods into two classes: methods without auxiliary actions and methods with auxiliary actions. We review them in details and show the advantages and limitations of each type. Furthermore, we show several popular evaluation metrics that are widely used in grasp planning. Finally, difficulties and future research directions are discussed.