Robotic Grasp Detection for Medical Test Tubes Based on Convolutional Neural Network

Visual-based robot grasping detection is a significant challenge in the field of robotics. With the widespread application of convolutional neural networks, significant progress has been made in this field in recent years. However, most of these methods only perform well on general objects, while their effectiveness is significantly reduced for small, smooth, and transparent medical test tube containers in medical scenarios. We propose a medical test tube dataset that specifically focuses on medical scenarios to address this issue. The dataset consists of 9 different types of medical test tubes, and over 200 sets of images were captured using a depth camera from various positions of the test tubes. Each set of images contains RGB information as well as depth information. Manual annotation was performed to label the grasping of the medical test tubes in the later stage. Furthermore, we introduce an end-to-end parallel feature fusion network that takes either RGB or depth maps as inputs and generates pixel-level grasp quality. Through experiments, our method achieves accuracies of 97.22%, 88.89%, and 83.33% on RGB images, depth images, and RGB-D images, respectively. These results validate the efficacy of our proposed approach for handling medical test tube containers.