Robotic Grasp Detection via Residual Efficient Channel Attention and Multiscale Feature Learning

Current robotic grasp detection methods frequently exhibit limited accuracy due to inadequate attention to critical object features and inefficient utilization of multiscale object information. To address these limitations, we propose a grasp detection method that systematically integrates attention mechanisms with multiscale feature learning. A key component of our method is the residual efficient channel attention (RECA) module. By utilizing lightweight 1-D convolutional operations for cross-channel interaction and integrating residual connections to enhance feature representation, this module significantly strengthens the network’s capacity to focus on the graspable areas of objects. Furthermore, our method implements a multiscale feature processing architecture that combines selective kernel (SK) convolution featured by dynamic receptive fields with frequency-domain upsampling operations, allowing the model to flexibly adjust the grasping regions in accordance with the object scales. Experimental evaluations on benchmark datasets reveal superior performance, with accuracies reaching 99.2% on the Cornell dataset and 95.9% on the Jacquard dataset. Additional assessments under multiobject scenarios further reveal that our method generates grasp representations with significantly higher grasp quality scores. Physical experiments conducted on an FR5 robotic platform confirm operational reliability, achieving 97% average success rate across 200 grasp trials, demonstrating practical efficacy in real-world applications.