Year: 2025
Type: article
Abstract: Robotic grasping is crucial in manufacturing, logistics, and service robotics, but existing methods struggle with object occlusion and complex arrangements in cluttered scenes. We propose the Generative Residual Attention Network (GR-AttNet), based on the Generative Residual Convolutional Neural Network (GR-CNN). The spatial attention mechanism enhances its adaptability to clutter, and architectural optimization maintains high accuracy while reducing parameters and boosting efficiency. Experiments show GR-AttNet achieves 98.1% accuracy on the Cornell Grasping Dataset, 94.9% on the Jacquard Dataset, with processing speeds of 148 ms and 163 ms respectively, outperforming several state-of-the-art methods. With only 2.8 million parameters, it surpasses various existing models. Simulations further validate over 50% success rates across different scenario complexities. However, challenges remain in handling small object missed detection and severe occlusion cases. GR-AttNet offers a novel, highly practical solution for robotic grasping tasks.
Source: PLoS ONE
Authors: Shengbang Zhou, Zhongsheng Liu, Chuanqi Li, Dong Chen, Shutian Liu, Yadong Liu
Institution: Nanning Normal University
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,382
OpenAlex ID: W4417001580