An Attention-Based Approach for Enhanced Robot Grasp Detection in Unstructured Environments

Abstract

Accurate grasp detection in unstructured environments remains a challenging problem in robot manipulation tasks due to irregular layouts, occlusion, complex backgrounds, and different object shapes. In order to better understand the feature relationship between different positions in the image, a new grasp detection network is proposed in this paper. The attention part of this network combines the advantages of Softmax and linear attention and is named SoLiNet. SoLiNet introduces an additional set of agent labels in the traditional attention module and computes them for each window in each agent. By adjusting the attention weight and position bias of input features, the model can better capture global context information and thus obtain better performance when grasping unknown objects. SoLiNet's accuracy was 98.9% on the Cornell dataset and 96.2% on the Jacquard dataset. It has been successfully used in the actual robot arm grasping system.