Advancing Slip Classification in Robotic Manipulation Through Tactile Data Representation and Model Selection

The accurate slip detection is essential for achieving reliable robotic manipulation. The representation of tactile data and the architecture of the slip classification model significantly influence detection performance. However, the impact of different data representations and model architectures on slip classification remains underexplored. In this article, we systematically analyze the effect of various tactile data representations and data-driven classification architectures on both slip detection and prediction tasks. Our proposed synthetic minority oversampling technique convolutional neural network (SMOTE-CNN) model demonstrates superior performance, outperforming baseline models in slip detection by improving the ${F}1$ -score by 3% and reducing the false negative rate by over 95%—a critical improvement for preventing undetected slip events. To evaluate the practical applicability, we integrate the slip detection models into a slip avoidance controller, which employs model predictive control (MPC) for real-time trajectory optimization. Furthermore, the dataset of robotic manipulation collected in this study will be publicly available to facilitate future research and benchmarking.