摘要
Accurate grasping state detection is critical to the dexterous operation of robots. Robots must use multiple modalities to perceive external information, similar to humans. The direct fusion method of visual and tactile sensing may not provide effective visual–tactile features for the grasping state detection network of the target. To address this issue, we present a novel visual–tactile fusion model (i.e., RFCT) and provide an incremental dimensional tensor product method for detecting grasping states of weak-stiffness targets. We investigate whether convolutional block attention mechanisms (CBAM) can enhance feature representations by selectively attending to salient visual and tactile cues while suppressing less important information and eliminating redundant information for the initial fusion. We conducted 2250 grasping experiments using 15 weak-stiffness targets. We used 12 targets for training and three for testing. When evaluated on untrained targets, our RFCT model achieved a precision of 82.89%, a recall rate of 82.07%, and an F1 score of 81.65%. We compared RFCT model performance with various combinations of Resnet50 + LSTM and C3D models commonly used in grasping state detection. The experimental results show that our RFCT model significantly outperforms these models. Our proposed method provides accurate grasping state detection and has the potential to provide robust support for robot grasping operations in real-world applications.