A CNN-Based Grasp Planning Method for Random Picking of Unknown Objects with a Vacuum Gripper

Abstract

Robotic grasping is still challenging due to limitations in perception and control, especially when the CAD models of objects are unknown. Although some grasp planning approaches using computer vision have been proposed, these methods can be seen as open-loop grasp planning methods and are often not robust enough. In this paper, a novel grasp planning method combining CNN-based quality prediction and closed-loop control (CNNB-CL) is proposed for a vacuum gripper. A large-scale dataset is generated for CNN training, which contains more than 2.3 million synthetic grasps and their grasp qualities evaluated by grasp simulations with 3D models. Unlike other neural networks which predict grasp success by assigning a binary value or grasp quality level by assigning an integer value, the proposed CNN predicts the grasp quality via a linear regression architecture. Additionally, the method adjusts the grasp strategies and detects the optimal grasp based on feedback from a force-torque sensor. Various simulations and physical experiments prove that the CNNB-CL method is robust for random noise disturbance in observation and compatible with different depth cameras and vacuum grippers. The proposed method finds the optimal grasp from 2,000 candidates within 300 ms and achieves a 92.18% average success rate for different vacuum grippers, which outperforms the state-of-the-art methods regarding success rate and robustness.