Robotic Grasping Technology Integrating Large Kernel Convolution and Residual Connections

摘要

To meet real-time grasping demands in complex environments, this paper proposes a lightweight yet high-performance robotic grasping model. The model integrates large kernel convolution and residual connections to generate grasping information for unknown objects from RGB and depth images, enabling real-time generation of stable grasping plans from the images. The proposed model achieved favorable accuracy on both the Cornell and Jacquard standard grasping datasets. Compared to other methods, the proposed model significantly reduces the number of parameters while achieving comparable performance, making it a lightweight model. Additionally, real-world experiments were conducted using a six-axis collaborative robot on a set of previously unseen household objects with diverse and adversarial shapes, achieving a comprehensive grasping success rate of 93.7%. Experimental results demonstrate that the proposed model not only improves grasping accuracy but also has strong potential for practical applications, particularly in resource-constrained robotic systems.