Single-shot RGB-D grasping of objects using a multi-finger robot: a grasp rectangle approach with post-processing

Grasping objects with robots is a complex challenge in the field of robotics. This research introduces a fast and dependable method for picking up objects of varying shapes and colours. The primary aim is to develop a flexible approach capable of handling a wide array of different objects. The proposed method is designed to function when objects are placed on a table, and the robot is positioned either above or in front of them. Building upon an existing technique called the grasp rectangle, we employ a trained network to enhance the way we grasp objects. What sets this apart is our ability to expand the network's capabilities to work with robots equipped with multiple fingers. To achieve this, we incorporate a post-processing step into the network. Our experiments validate the effectiveness of our approach. We achieved a successful object grasp rate of 94.4% when viewed from above and an accuracy of 95.6% when grasping from the side. These findings highlight the considerable potential of our method in addressing the challenging problem of robotic grasping, particularly in scenarios involving different object placements and colours.