Grasp detection takes on a critical significance for the robot. However, detecting object positions and corresponding grasp positions in a stacked environment can be quite difficult for a robot. Based on this practical problem, in order to achieve more accurate object position detection and grasp position detection, a new method called MMD (Multi-stage network for multi-object grasp detection algorithm) is proposed in this paper. MMD covers two parts, including the feature extractor and the multi-stage object predictor. The feature extractor refers to a deep convolutional neural network that can generate shared feature layers as well as the initial ROIs (region of interest). A multi-stage refiner serves as the multi-stage object predictor, which continuously regresses the initial ROI to obtain more accurate object detection and grasping detection results. Ablation experiments show that the proposed MMD has better grasp detection performance. The specific performance is that the recognition precision achieves a state-of-the-art 76.71% mAPg on the VMRD dataset. Moreover, test experiments demonstrate the feasibility of our method on the Kinova robot.
Keywords: Kinova robot; VMRD; grasp detection; multi-stage network; multi-task; stack scenarios.