Task-Oriented Robotic Grasping for Intelligent Manufacturing

Grasping is the first step in many robotic manipulation tasks, and it is essential for applications of robots in intelligent manufacturing scenarios. Traditional grasp planning methods rely on force-closure analysis and fail to adapt for new objects quickly. Recent progress in computer vision and deep learning makes automatic grasping planning promising for unseen new objects. However, these grasp planning methods usually generate arbitrary grasp poses, and few has consider the problem of choosing the appropriate grasp pose for subsequent manipulation tasks. In this study, we propose a novel yet practical grasp mapping method Hand-to-Gripper (H2G). In H2G, a task-oriented semantic grasp strategy is introduced to meet the semantic constraints to make an action feasible. In the learning phase, it constructs a hand-object grasp relationship from both the estimation of the object pose and keypoints of the user’s hand. In the application phase, it repeats this user-specified task-oriented task. The proposed approach is evaluated on a UR-10e robot with a RealSense D435i RGB-D camera. The implemented system provides a low-cost solution for teaching robot task-oriented grasps in intelligent manufacturing scenarios.