DSC-GraspNet: A Lightweight Convolutional Neural Network for Robotic Grasp Detection

Grasp detection is an essential task for robots to achieve autonomous operation, it can also make virtual reality-based teleoperation more intelligent and reliable. Existing learning-based grasp detection methods usually fail to strike a balance between high accuracy and low time consumption. Also, the large number of model parameters tends to make these methods expensive to deploy. To solve this problem, a lightweight generative grasp detection network DSC-GraspNet is proposed. Firstly, Depth-separable convolutional blocks with Coordinate Attention (CA) are stacked to obtain a lightweight backbone network for feature extraction. Then multi-level features extracted by the backbone network are fused by the Cross Stage Partial (CSP) block in the up-sampling network. Finally, pixel-level grasp candidates are generated by grasp generating heads. Experimental results shows that an accuracy of 98.3% under image-wise splitting and 97.7% under object-wise splitting can be achieved on the Cornell public dataset. Meanwhile, an accuracy of 94.7% is achieved on the Jacquard dataset using the depth map as inputs. Our method also achieve a grasp success rate of 86.4% in the simulated grasp test. In addition, our network is able to inference an RGB-D image within 14ms, and can be applied to closed-loop grasping scenarios.