Comparisons of data-driven models for detecting slip occurrence and direction based on simulations of tactile sensing

Abstract

Slip detection based on tactile sensing plays a crucial role for a robot to achieving stable grasps by promptly adjusting its grasping state. Detecting the occurrence of slippage has been a focus of previous research, but identifying the direction of slippage can provide richer reference information for correcting grasp issues. Due to the variations of experiment setup, the performances of these data-driven models have not been sufficiently compared. In this study, we compared the performance of three traditional machine learning algorithms—Random Forest (RF), Support Vector Machine (SVM), and k-Nearest Neighbors (KNN)—with four deep learning models—Gated Recurrent Unit (GRU), Convolutional Gated Recurrent Unit (CNN-GRU), Convolutional Long Short-Term Memory (ConvLSTM), and 3D Convolution(Conv3D)—in detecting slippage states and directions. We conducted experiments using a simulated dataset collected in Gazebo. Additionally, we compared the noise resistance capability of each model and their generalization performance when facing new objects. The results show that when facing known grasped objects, all models can effectively detect slippage occurrences and their directions. However, as noise increases, Conv3D exhibits the strongest robustness. When generalized to unknown grasped objects, the deep learning models outperforms the three traditional learning algorithms, with CNN-GRU showing the best performance.