Lili Cai
Robotic manipulation, grasp detection, adaptive fusion
We focus on the research of robotic arm grasp detection algorithms based on visual information. Aiming at the problems in unstructured environments—such as low accuracy of grasp detection algorithms, failure to effectively distinguish between backgrounds and grasped objects, insufficient extraction and utilisation of multi- scale information, inadequate consideration of global information, and weak perception ability for textureless transparent objects—an innovative algorithm based on deep learning is proposed to improve the robot’s perception and manipulation capabilities. To address the challenges of object grasping in complex scenarios, this paper presents an efficient grasp perception network (EGA-Net), which converts the grasp detection problem into a pixel-level semantic segmentation task to achieve end-to-end dense grasp configuration prediction. By introducing the efficient channel attention ResNet (ECA-ResNet), the network establishes a direct relationship between channels and weights, enabling it to focus on features relevant to the grasping task while suppressing irrelevant information. Experiments on the public Cornell dataset and Jacquard dataset demonstrate that EGA-Net can predict reliable grasp configurations, outperforms multiple existing state-of-the-art algorithms in unified evaluation metrics, and exhibits high real-time performance.
Important Links:
Go Back
