High-Performance Pixel-Level Grasp Detection Based on Adaptive Grasping and Grasp-Aware Network

Machine vision-based planar grasping detection is challenging due to uncertainty about object shape, pose, size, etc. Previous methods mostly focus on predicting discrete gripper configurations, and may miss some ground-truth grasp postures. In this article, a pixel-level method proposed, which uses deep neural network predict configurations RGB images. First, novel oriented arrow representation model (OAR-model) introduced represent the configuration of parallel-jaw three-fingered gripper, can partly improve applicability different grippers. Then, adaptive attribute proposed adaptively objects, for resolving angle conflicts in training simplifying labeling. Lastly, feature fusion grasp-aware (AFFGA-Net) OAR-models AFFGA-Net improves robustness unstructured scenarios by using hybrid atrous spatial pyramid decoder connected sequence. On public Cornell dataset actual our structure achieves 99.09% 98.0% accuracy, respectively. over 2400 robotic trials, an average success rate 98.77% single-object 93.69% cluttered scenarios. Moreover, completes pipeline within 15 ms.