With the rapid advancement of tactile sensing technologies, real-time prediction of robotic grasp stability has become increasingly critical for ensuring safe and reliable operation in automated industrial environments. Achieving accurate real-time predictions of grasp instability is essential for preventing failures during robotic manipulation tasks. Tactile signals serves as a primary source of contact information during grasping and provides detailed insight into the interaction forces and surface dynamics. However, tactile signals often exhibit complex and non-stationary temporal characteristics, posing significant challenges for efficient and robust modeling. To address these challenges while maintaining computational efficiency, we propose GraspLite-Net, a lightweight deep learning-based model that relies solely on tactile time-series data, without requiring visual input. Our proposed GraspLite-Net captures both local and global temporal dependencies while remaining suitable for real-time deployment and comprises three main components: (1) a deformable patch segmentation module that adaptively identifies informative temporal segments; (2) a multi-scale re-parameterized convolutional block for enhanced temporal feature extraction; and (3) an attention-based multiple-instance learning (MIL) pooling module for effective feature aggregation. We conduct extensive experiments on the BiGS dataset and show that GraspLite-Net surpasses baseline methods in both prediction accuracy and inference speed, underscoring its effectiveness and suitability for deployment in real-world robotic grasping systems. We further perform a cross-dataset evaluation on the SnapFitForceProfiles dataset to demonstrate the model’s generalization capability to new robotic scenarios.
Wang et al. (Sun,) studied this question.